Bump mold version to 2.2.0

Fixes https://github.com/rui314/mold/issues/1107

.github/workflows/ci.yml

@@ -26,11 +26,20 @@ jobs:
     - name: build and test
       run: |
         echo "/usr/lib/ccache:/usr/local/opt/ccache/libexec" >> $GITHUB_PATH
+        sudo apt-get install -y clang++-12
         mkdir build
         cd build
-        cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ ${{ matrix.target }} ..
+        cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++-12 ${{ matrix.target }} ..
         cmake --build . -j$(nproc)
         ctest . -j$(nproc)
+    - name: archive test results
+      uses: actions/upload-artifact@v3
+      if: failure()
+      with:
+        name: test-results-clang
+        path: |
+          build/out
+          build/Testing
 
   build-gcc:
     runs-on: ubuntu-20.04
@@ -42,26 +51,45 @@ jobs:
         # Install cross toolchains
         dpkg --add-architecture i386
         ./install-build-deps.sh update
-        apt-get install -y sudo qemu-user gdb zstd dwarfdump {gcc,g++}-10-{i686,aarch64,riscv64,powerpc,powerpc64,powerpc64le,s390x,sparc64,m68k,sh4,alpha}-linux-gnu {gcc,g++}-10-arm-linux-gnueabihf
+        apt-get install -y sudo qemu-user gdb zstd dwarfdump xz-utils {gcc,g++}-10-{i686,aarch64,riscv64,powerpc,powerpc64,powerpc64le,s390x,sparc64,m68k,sh4,alpha}-linux-gnu {gcc,g++}-10-arm-linux-gnueabihf
 
-        for i in i686 aarch64 riscv64 powerpc powerpc64 powerpc64le s390x sparc64 m68k sh4 alpha; do
-          ln -sf /usr/bin/$i-linux-gnu-gcc-10 /usr/bin/$i-linux-gnu-gcc
-          ln -sf /usr/bin/$i-linux-gnu-g++-10 /usr/bin/$i-linux-gnu-g++
+        for i in {i686,aarch64,riscv64,powerpc,powerpc64,powerpc64le,s390x,sparc64,m68k,sh4,alpha}-linux-gnu arm-linux-gnueabihf; do
+          ln -sf /usr/bin/$i-gcc-10 /usr/bin/$i-gcc
+          ln -sf /usr/bin/$i-g++-10 /usr/bin/$i-g++
         done
 
-        ln -sf /usr/bin/arm-linux-gnueabihf-gcc-10 /usr/bin/arm-linux-gnueabihf-gcc
-        ln -sf /usr/bin/arm-linux-gnueabihf-g++-10 /usr/bin/arm-linux-gnueabihf-g++
-
         # Install a RV32 toolchain from third party since it's not available
         # as an Ubuntu package.
         mkdir /usr/local/rv32
-        wget -O- -q https://github.com/riscv-collab/riscv-gnu-toolchain/releases/download/2022.08.26/riscv32-glibc-ubuntu-20.04-nightly-2022.08.26-nightly.tar.gz | tar -C /usr/local/rv32 --strip-components=1 -xzf -
+        wget -O- -q https://github.com/riscv-collab/riscv-gnu-toolchain/releases/download/2023.07.07/riscv32-glibc-ubuntu-20.04-gcc-nightly-2023.07.07-nightly.tar.gz | tar -C /usr/local/rv32 --strip-components=1 -xzf -
 
         ln -sf /usr/local/rv32/sysroot /usr/riscv32-linux-gnu
-        ln -sf /usr/local/rv32/bin/riscv32-unknown-linux-gnu-objdump /usr/bin/riscv32-linux-gnu-objdump
         echo '/usr/local/rv32/bin/riscv32-unknown-linux-gnu-gcc -L/usr/riscv32-linux-gnu "$@"' > /usr/bin/riscv32-linux-gnu-gcc
         echo '/usr/local/rv32/bin/riscv32-unknown-linux-gnu-g++ -L/usr/riscv32-linux-gnu "$@"' > /usr/bin/riscv32-linux-gnu-g++
-        chmod 755 /usr/bin/riscv32-linux-gnu-gcc /usr/bin/riscv32-linux-gnu-g++
+        chmod 755 /usr/bin/riscv32-linux-gnu-{gcc,g++}
+
+        for i in objdump objcopy strip; do
+          ln -sf /usr/local/rv32/bin/riscv32-unknown-linux-gnu-$i /usr/bin/riscv32-linux-gnu-$i
+        done
+
+        # Install a LoongArch toolchain
+        mkdir /usr/local/larch
+        wget -O- -q https://github.com/loongson/build-tools/releases/download/2023.08.08/CLFS-loongarch64-8.1-x86_64-cross-tools-gcc-glibc.tar.xz | tar -C /usr/local/larch --strip-components=1 --xz -xf -
+
+        ln -sf /usr/local/larch/target /usr/loongarch64-linux-gnu
+        cp -r /usr/local/larch/loongarch64-unknown-linux-gnu/lib/* /usr/loongarch64-linux-gnu/lib64/
+
+        for i in objdump objcopy strip; do
+          ln -sf /usr/local/larch/bin/loongarch64-unknown-linux-gnu-$i /usr/bin/loongarch64-linux-gnu-$i
+        done
+
+        echo '/usr/local/larch/bin/loongarch64-unknown-linux-gnu-gcc -L/usr/local/larch/loongarch64-unknown-linux-gnu "$@"' > /usr/bin/loongarch64-linux-gnu-gcc
+        echo '/usr/local/larch/bin/loongarch64-unknown-linux-gnu-g++ -L/usr/local/larch/loongarch64-unknown-linux-gnu "$@"' > /usr/bin/loongarch64-linux-gnu-g++
+        chmod 755 /usr/bin/loongarch64-linux-gnu-{gcc,g++}
+
+        wget -O/usr/local/bin/qemu-loongarch64 -q https://github.com/loongson/build-tools/releases/download/2023.08.08/qemu-loongarch64
+        chmod 755 /usr/local/bin/qemu-loongarch64
+
       shell: bash
     - name: ccache
       uses: hendrikmuhs/ccache-action@v1
@@ -70,14 +98,14 @@ jobs:
         echo "/usr/lib/ccache:/usr/local/opt/ccache/libexec" >> $GITHUB_PATH
         mkdir build
         cd build
-        cmake -DMOLD_ENABLE_QEMU_TESTS=On -DMOLD_ENABLE_QEMU_TESTS_RV32=On ..
+        cmake -DMOLD_ENABLE_QEMU_TESTS=On -DMOLD_ENABLE_QEMU_TESTS_RV32=On -DMOLD_ENABLE_QEMU_TESTS_LOONGARCH=On ..
         cmake --build . -j$(nproc)
         ctest . -j$(nproc)
     - name: archive test results
       uses: actions/upload-artifact@v3
-      if: always()
+      if: failure()
       with:
-        name: test-results
+        name: test-results-gcc
         path: |
           build/out
           build/Testing

.github/workflows/linux-packages.yml

@@ -0,0 +1,15 @@
+name: linux-packages
+on:
+  push:
+    branches: [ main ]
+jobs:
+  build:
+    runs-on: ubuntu-20.04
+    steps:
+    - uses: actions/checkout@v3
+    - name: build
+      run: ./dist.sh x86_64
+    - name: upload
+      uses: actions/upload-artifact@v3
+      with:
+        path: mold-*-linux.tar.gz

CMakeLists.txt

@@ -43,8 +43,8 @@
 # of the mold linker of the same version will have the exactly same set of
 # features and behave exactly the same.
 
-cmake_minimum_required(VERSION 3.13)
-project(mold VERSION 1.8.0)
+cmake_minimum_required(VERSION 3.14)
+project(mold VERSION 2.2.0)
 
 include(CMakeDependentOption)
 include(GNUInstallDirs)
@@ -52,7 +52,7 @@ include(GNUInstallDirs)
 # Build mold itself using mold if -DMOLD_USE_MOLD=ON
 option(MOLD_USE_MOLD "Use mold to build mold" OFF)
 if(MOLD_USE_MOLD)
-  add_link_options(-fuse-ld=mold)
+  add_link_options(-fuse-ld=mold -Wl,--gdb-index)
 endif()
 
 add_executable(mold)
@@ -65,7 +65,17 @@ if(NOT "${CMAKE_CXX_COMPILER_FRONTEND_VARIANT}" STREQUAL "MSVC")
     -fno-unwind-tables
     -fno-asynchronous-unwind-tables
     -Wno-sign-compare
-    -Wno-unused-function)
+    -Wno-unused-function
+    -Wunused-variable
+    -ggnu-pubnames)
+endif()
+
+if(CMAKE_BUILD_TYPE STREQUAL "Debug")
+  target_compile_options(mold PRIVATE -D_GLIBCXX_ASSERTIONS)
+endif()
+
+if(CMAKE_SYSTEM_NAME STREQUAL "OpenBSD")
+  set(OPENBSD ON)
 endif()
 
 # Build mold with -flto if -DMOLD_LTO=ON
@@ -88,7 +98,7 @@ if(MOLD_USE_TSAN)
   target_link_options(mold PRIVATE -fsanitize=thread)
 endif()
 
-# Statically-link libstdc++ and libcrypto if -DMOLD_MOSTLY_STATIC=ON.
+# Statically-link libstdc++ if -DMOLD_MOSTLY_STATIC=ON.
 #
 # This option is intended to be used by `./dist.sh` script to create a
 # mold binary that works on various Linux distros. You probably don't
@@ -96,13 +106,12 @@ endif()
 option(MOLD_MOSTLY_STATIC "Statically link libstdc++ and libcrypto" OFF)
 if(MOLD_MOSTLY_STATIC)
   target_link_options(mold PRIVATE -static-libstdc++)
-  target_link_libraries(mold PRIVATE libcrypto.a)
 endif()
 
 # Find zlib. If libz.so is not found, we compile a bundled one and
 # statically-link it to mold.
 find_package(ZLIB QUIET)
-if(ZLIB_FOUND)
+if(ZLIB_FOUND AND NOT MOLD_MOSTLY_STATIC)
   target_link_libraries(mold PRIVATE ZLIB::ZLIB)
 else()
   add_subdirectory(third-party/zlib EXCLUDE_FROM_ALL)
@@ -111,12 +120,23 @@ else()
   target_link_libraries(mold PRIVATE zlibstatic)
 endif()
 
-# Find zstd compression library. Just like zlib, if libzstd.so is not
-# found, we compile a bundled one and statically-link it to mold.
+# Find BLAKE3 cryptographic hash library. Just like zlib, if libblkae3.so
+# is not found, we compile a bundled one and statically-link it to mold.
+find_package(BLAKE3 QUIET)
+if(BLAKE3_FOUND AND NOT MOLD_MOSTLY_STATIC)
+  target_link_libraries(mold PRIVATE BLAKE3::blake3)
+else()
+  add_subdirectory(third-party/blake3/c EXCLUDE_FROM_ALL)
+  target_link_libraries(mold PRIVATE blake3)
+  target_include_directories(mold PUBLIC third-party/blake3/c)
+endif()
+
+# Find zstd compression library. If libzstd.so is not found, we compile a
+# bundled one and statically-link it to mold.
 include(CheckIncludeFile)
 check_include_file(zstd.h HAVE_ZSTD_H)
 
-if(HAVE_ZSTD_H)
+if(HAVE_ZSTD_H AND NOT MOLD_MOSTLY_STATIC)
   target_link_libraries(mold PRIVATE zstd)
 else()
   add_subdirectory(third-party/zstd/build/cmake EXCLUDE_FROM_ALL)
@@ -142,7 +162,7 @@ check_cxx_source_compiles(
   "static_assert(sizeof(void *) == 4); int main() {}" P32)
 
 cmake_dependent_option(MOLD_USE_MIMALLOC "Use mimalloc" ON
-  "NOT APPLE; NOT ANDROID; NOT P32" OFF)
+  "NOT APPLE; NOT ANDROID; NOT OPENBSD; NOT P32" OFF)
 cmake_dependent_option(
   MOLD_USE_SYSTEM_MIMALLOC "Use system or vendored mimalloc" OFF
   MOLD_USE_MIMALLOC OFF)
@@ -198,12 +218,16 @@ endif()
 
 # We always use Clang to build mold on Windows. MSVC can't compile mold.
 if(WIN32)
-  if(NOT CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+  if(MSVC AND NOT CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
     message(FATAL_ERROR
       "Your compiler is not supported; install Clang from Visual Studio Installer and re-run cmake with '-T clangcl'")
   endif()
 
   target_compile_definitions(mold PRIVATE NOGDI NOMINMAX)
+  if(MINGW)
+    target_compile_definitions(mold PRIVATE _WIN32_WINNT=0xA00)
+    target_link_libraries(mold PRIVATE bcrypt)
+  endif()
 else()
   include(CheckLibraryExists)
   check_library_exists(m pow "" LIBM_FOUND)
@@ -223,11 +247,6 @@ if(NOT APPLE AND NOT WIN32)
   target_sources(mold-wrapper PRIVATE elf/mold-wrapper.c)
 endif()
 
-if(NOT APPLE AND NOT WIN32 AND NOT MOLD_MOSTLY_STATIC)
-  find_package(OpenSSL REQUIRED COMPONENTS Crypto)
-  target_link_libraries(mold PRIVATE OpenSSL::Crypto)
-endif()
-
 # If atomics doesn't work by default, add -latomic.
 # We need the flag on riscv, armv6 and m68k.
 include(CheckCXXSourceCompiles)
@@ -245,7 +264,7 @@ if(NOT HAVE_FULL_ATOMIC_SUPPORT)
 endif()
 
 # Add -pthread
-if(NOT APPLE AND NOT WIN32)
+if(NOT APPLE AND NOT MSVC)
   target_compile_options(mold PRIVATE -pthread)
   target_link_options(mold PRIVATE -pthread)
 endif()
@@ -255,15 +274,15 @@ add_custom_target(git_hash
   COMMAND ${CMAKE_COMMAND}
     -DSOURCE_DIR=${CMAKE_SOURCE_DIR}
     -DOUTPUT_FILE=${CMAKE_BINARY_DIR}/git-hash.cc
-    -P ${CMAKE_SOURCE_DIR}/update-git-hash.cmake
-  DEPENDS update-git-hash.cmake
+    -P ${CMAKE_SOURCE_DIR}/common/update-git-hash.cmake
+  DEPENDS common/update-git-hash.cmake
   BYPRODUCTS git-hash.cc
   VERBATIM)
 
 add_dependencies(mold git_hash)
 
 # Create config.h file
-configure_file(config.h.in config.h)
+configure_file(common/config.h.in config.h)
 include_directories(${CMAKE_CURRENT_BINARY_DIR})
 
 # Almost all functions are template in mold which take a target type
@@ -275,16 +294,19 @@ include_directories(${CMAKE_CURRENT_BINARY_DIR})
 # compiler instances. This is hacky but greatly reduces compile time
 # on a multicore machine.
 list(APPEND MOLD_ELF_TARGETS
-  X86_64 I386 ARM64 ARM32 RV32LE RV32BE RV64LE RV64BE
-  PPC32 PPC64V1 PPC64V2 S390X SPARC64 M68K SH4 ALPHA)
+  X86_64 I386 ARM64 ARM32 RV32LE RV32BE RV64LE RV64BE PPC32 PPC64V1 PPC64V2
+  S390X SPARC64 M68K SH4 ALPHA LOONGARCH32 LOONGARCH64)
 
 list(APPEND MOLD_ELF_TEMPLATE_FILES
+  elf/arch-loongarch.cc
+  elf/arch-riscv.cc
   elf/cmdline.cc
   elf/dwarf.cc
   elf/gc-sections.cc
   elf/icf.cc
   elf/input-files.cc
   elf/input-sections.cc
+  elf/jobs.cc
   elf/linker-script.cc
   elf/lto.cc
   elf/main.cc
@@ -294,6 +316,7 @@ list(APPEND MOLD_ELF_TEMPLATE_FILES
   elf/relocatable.cc
   elf/subprocess.cc
   elf/thunks.cc
+  elf/tls.cc
   )
 
 list(APPEND MOLD_MACHO_TARGETS X86_64 ARM64)
@@ -312,10 +335,12 @@ list(APPEND MOLD_MACHO_TEMPLATE_FILES
 
 function(mold_instantiate_templates SOURCE TARGET)
   set(PATH ${CMAKE_BINARY_DIR}/${SOURCE}.${TARGET}.cc)
-  file(WRITE ${PATH} "#define MOLD_${TARGET} 1
+  if(NOT EXISTS ${PATH})
+    file(WRITE ${PATH} "#define MOLD_${TARGET} 1
 #define MOLD_TARGET ${TARGET}
 #include \"${CMAKE_SOURCE_DIR}/${SOURCE}\"
 ")
+  endif()
   target_sources(mold PRIVATE ${PATH})
 endfunction()
 
@@ -335,8 +360,16 @@ endif()
 
 # Add other non-template source files.
 target_sources(mold PRIVATE
-  compress.cc
-  demangle.cc
+  common/compress.cc
+  common/demangle.cc
+  common/filepath.cc
+  common/glob.cc
+  common/hyperloglog.cc
+  common/main.cc
+  common/multi-glob.cc
+  common/perf.cc
+  common/tar.cc
+  common/uuid.cc
   elf/arch-alpha.cc
   elf/arch-arm32.cc
   elf/arch-arm64.cc
@@ -345,22 +378,13 @@ target_sources(mold PRIVATE
   elf/arch-ppc32.cc
   elf/arch-ppc64v1.cc
   elf/arch-ppc64v2.cc
-  elf/arch-riscv.cc
   elf/arch-s390x.cc
   elf/arch-sh4.cc
   elf/arch-sparc64.cc
   elf/arch-x86-64.cc
   elf/elf.cc
-  filepath.cc
   git-hash.cc
-  glob.cc
-  hyperloglog.cc
-  main.cc
-  multi-glob.cc
-  perf.cc
-  tar.cc
   third-party/rust-demangle/rust-demangle.c
-  uuid.cc
   )
 
 if(MOLD_IS_SOLD)
@@ -421,9 +445,15 @@ endif()
 
 if(NOT CMAKE_SKIP_INSTALL_RULES)
   install(TARGETS mold RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
-  install(FILES LICENSE DESTINATION ${CMAKE_INSTALL_DOCDIR})
   install(FILES docs/mold.1 DESTINATION ${CMAKE_INSTALL_MANDIR}/man1/)
 
+  if(EXISTS "${CMAKE_SOURCE_DIR}/LICENSE")
+    install(FILES LICENSE DESTINATION ${CMAKE_INSTALL_DOCDIR})
+  endif()
+  if(EXISTS "${CMAKE_SOURCE_DIR}/LICENSE.third-party")
+    install(FILES "LICENSE.third-party" DESTINATION "${CMAKE_INSTALL_DOCDIR}")
+  endif()
+
   function(mold_install_relative_symlink OLD NEW)
     install(CODE "
       get_filename_component(PREFIX_ABS \${CMAKE_INSTALL_PREFIX}/ ABSOLUTE)
@@ -436,10 +466,10 @@ if(NOT CMAKE_SKIP_INSTALL_RULES)
       file(CREATE_LINK \${OLD_REL} \$ENV{DESTDIR}\${NEW_ABS} SYMBOLIC)")
   endfunction()
 
-  mold_install_relative_symlink(${CMAKE_INSTALL_BINDIR}/mold
-    ${CMAKE_INSTALL_LIBEXECDIR}/mold/ld)
-  mold_install_relative_symlink(${CMAKE_INSTALL_BINDIR}/mold
-    ${CMAKE_INSTALL_BINDIR}/ld.mold)
+  mold_install_relative_symlink(${CMAKE_INSTALL_BINDIR}/mold${CMAKE_EXECUTABLE_SUFFIX}
+    ${CMAKE_INSTALL_LIBEXECDIR}/mold/ld${CMAKE_EXECUTABLE_SUFFIX})
+  mold_install_relative_symlink(${CMAKE_INSTALL_BINDIR}/mold${CMAKE_EXECUTABLE_SUFFIX}
+    ${CMAKE_INSTALL_BINDIR}/ld.mold${CMAKE_EXECUTABLE_SUFFIX})
   mold_install_relative_symlink(${CMAKE_INSTALL_MANDIR}/man1/mold.1
     ${CMAKE_INSTALL_MANDIR}/man1/ld.mold.1)
 

CMakeSettings.json

@@ -1,15 +0,0 @@
-{
-  "configurations": [
-    {
-      "name": "x64-Debug",
-      "generator": "Ninja",
-      "configurationType": "Debug",
-      "inheritEnvironments": [ "clang_cl_x64" ],
-      "buildRoot": "${projectDir}\\out\\build\\${name}",
-      "installRoot": "${projectDir}\\out\\install\\${name}",
-      "cmakeCommandArgs": "",
-      "buildCommandArgs": "",
-      "ctestCommandArgs": ""
-    }
-  ]
-}

CONTRIBUTING.md

@@ -1,44 +0,0 @@
-# Contributing to mold
-
-We request all patches to mold to be released under the dual license
-of the GNU AGPLv3 and the MIT license. Please read the following and
-sign-off your patch as an indication of agreement.
-
-> ## Developer's Certificate of Origin
->
-> By making a contribution to this project, I certify that:
->
->  1. The contribution was created in whole or in part by me and I
->     have the right to submit it under the Dual License of the GNU
->     AGPLv3 and the MIT license; or
->
->  2. The contribution is based upon previous work that, to the best
->     of my knowledge, is covered under the the Dual License and I
->     have the right under the Dual License to submit that work with
->     modifications, whether created in whole or in part by me, under
->     the Dual License (unless I am permitted to submit under
->     different licenses), as indicated in the file; or
->
->  3. The contribution was provided directly to me by some other
->     person who certified (1), (2) or (3) and I have not modified it.
->
->  4. I understand and agree that this project and the contribution
->     are public and that a record of the contribution (including all
->     personal information I submit with it, including my sign-off) is
->     maintained indefinitely and may be redistributed consistent with
->     this project or the open source license(s) involved.
-
-If you agree, you just add a line saying:
-
-```
-Signed-off-by: Random J Developer <[email protected]>
-```
-
-using your real name (sorry, no pseudonyms or anonymous contributions.)
-This will be done for you automatically if you use ``git commit -s``.
-Reverts should also include "Signed-off-by". ``git revert -s`` does that
-for you.
-
-We assume that any contribution that's 10 lines or less does not meet
-the threshold of originality and therefore copyright does not apply.
-You don't need to add a "Signed-off-by" line to such small patches.

LICENSE.third-party

@@ -0,0 +1,348 @@
+Mold includes a number of subcomponents with separate copyright notices
+and license terms. Your use of the source code for the these subcomponents
+is subject to the terms and conditions of the following licenses.
+
+
+=== For the third-party/mimalloc component:
+
+MIT License
+
+Copyright (c) 2018-2021 Microsoft Corporation, Daan Leijen
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+
+=== For the third-party/rust-demangle component:
+
+Permission is hereby granted, free of charge, to any
+person obtaining a copy of this software and associated
+documentation files (the "Software"), to deal in the
+Software without restriction, including without
+limitation the rights to use, copy, modify, merge,
+publish, distribute, sublicense, and/or sell copies of
+the Software, and to permit persons to whom the Software
+is furnished to do so, subject to the following
+conditions:
+
+The above copyright notice and this permission notice
+shall be included in all copies or substantial portions
+of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF
+ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED
+TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A
+PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT
+SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
+IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+DEALINGS IN THE SOFTWARE.
+
+
+=== For the third-party/tbb component:
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+
+=== For the third-party/xxhash component:
+
+xxHash Library
+Copyright (c) 2012-2020 Yann Collet
+All rights reserved.
+
+BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this
+  list of conditions and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright notice, this
+  list of conditions and the following disclaimer in the documentation and/or
+  other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+=== For the third-party/zlib component:
+
+ (C) 1995-2022 Jean-loup Gailly and Mark Adler
+
+  This software is provided 'as-is', without any express or implied
+  warranty.  In no event will the authors be held liable for any damages
+  arising from the use of this software.
+
+  Permission is granted to anyone to use this software for any purpose,
+  including commercial applications, and to alter it and redistribute it
+  freely, subject to the following restrictions:
+
+  1. The origin of this software must not be misrepresented; you must not
+     claim that you wrote the original software. If you use this software
+     in a product, an acknowledgment in the product documentation would be
+     appreciated but is not required.
+  2. Altered source versions must be plainly marked as such, and must not be
+     misrepresented as being the original software.
+  3. This notice may not be removed or altered from any source distribution.
+
+  Jean-loup Gailly        Mark Adler
+  [email protected]          [email protected]
+
+
+=== For the third-party/zstd component:
+
+BSD License
+
+For Zstandard software
+
+Copyright (c) 2016-present, Facebook, Inc. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+ * Neither the name Facebook nor the names of its contributors may be used to
+   endorse or promote products derived from this software without specific
+   prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

README.md

@@ -1,22 +1,19 @@
 # mold: A Modern Linker
 
-[![CI](https://github.com/rui314/mold/actions/workflows/ci.yml/badge.svg)](https://github.com/rui314/mold/actions/workflows/ci.yml)
-[![build result](https://build.opensuse.org/projects/home:marxin:mold/packages/mold/badge.svg?type=default)](https://build.opensuse.org/package/show/home:marxin:mold/mold)
-
-<i>This is a repo of a free, AGPL-licensed version of the linker.
-If you are looking for a commercial, non-AGPL version of the same linker,
+<i>This repository contains a free version of the mold linker.
+If you are looking for a commercial version that supports macOS
 please visit the
-[repo of the sold linker](https://github.com/bluewhalesystems/sold).</i>
+[repository of the sold linker](https://github.com/bluewhalesystems/sold).</i>
 
-mold is a faster drop-in replacement for existing Unix linkers.
-It is several times faster than the LLVM lld linker, the second-fastest
-open-source linker which I originally created a few years ago.
-mold is designed to increase developer productivity by reducing
-build time, especially in rapid debug-edit-rebuild cycles.
+mold is a faster drop-in replacement for existing Unix linkers. It is several
+times quicker than the LLVM lld linker, the second-fastest open-source linker,
+which I initially developed a few years ago. mold aims to enhance developer
+productivity by minimizing build time, particularly in rapid
+debug-edit-rebuild cycles.
 
-Here is a performance comparison of GNU gold, LLVM lld, and mold for
-linking final debuginfo-enabled executables of major large programs
-on a simulated 8-core 16-threads machine.
+Here is a performance comparison of GNU gold, LLVM lld, and mold when linking
+final debuginfo-enabled executables for major large programs on a simulated
+8-core, 16-thread machine.
 
 ![Link speed comparison](docs/comparison.png)
 
@@ -26,48 +23,47 @@ on a simulated 8-core 16-threads machine.
 | Clang 13 (3.18 GiB)           | 64.12s   | 5.82s    | 2.90s
 | Firefox 89 libxul (1.64 GiB)  | 32.95s   | 6.80s    | 1.42s
 
-mold is so fast that it is only 2x _slower_ than `cp` on the same
-machine. Feel free to [file a bug](https://github.com/rui314/mold/issues)
-if you find mold is not faster than other linkers.
+mold is so fast that it is only 2x _slower_ than the `cp` command on the same
+machine. If you find that mold is not faster than other linkers, please feel
+free to [file a bug report](https://github.com/rui314/mold/issues).
 
 mold supports x86-64, i386, ARM64, ARM32, 64-bit/32-bit little/big-endian
-RISC-V, 64-bit big-endian PowerPC ELFv1, 64-bit little-endian PowerPC ELFv2,
-s390x, SPARC64 and m68k.
+RISC-V, 32-bit PowerPC, 64-bit big-endian PowerPC ELFv1, 64-bit little-endian
+PowerPC ELFv2, s390x, 64-bit/32-bit LoongArch, SPARC64, m68k, SH-4, and DEC
+Alpha.
 
 mold/macOS is commercial software. For mold/macOS, please visit
 https://github.com/bluewhalesystems/sold.
 
-## Why does the speed of linking matter?
+## Why does linking speed matter?
 
-If you are using a compiled language such as C, C++ or Rust, a build
-consists of two phases. In the first phase, a compiler compiles
-source files into object files (`.o` files). In the second phase,
-a linker takes all object files to combine them into a single executable
-or a shared library file.
+If you are using a compiled language such as C, C++, or Rust, a build consists
+of two phases. In the first phase, a compiler compiles source files into
+object files (`.o` files). In the second phase, a linker takes all object
+files and combines them into a single executable or shared library file.
 
-The second phase takes a long time if your build output is large.
-mold can make it faster, saving your time and keeping you from being
-distracted while waiting for a long build to finish. The difference is
-most noticeable when you are in rapid debug-edit-rebuild cycles.
+The second phase can be time-consuming if your build output is large. mold can
+speed up this process, saving you time and preventing distractions while
+waiting for a lengthy build to finish. The difference is most noticeable
+during rapid debug-edit-rebuild cycles.
 
-## Install
+## Installation
 
-Binary packages for the following systems are currently available.
+Binary packages for the following systems are currently available:
 
 [![Packaging status](https://repology.org/badge/vertical-allrepos/mold.svg)](https://repology.org/project/mold/versions)
 
-## How to build
+## How to Build
 
-mold is written in C++20, so if you build mold yourself, you need a
-recent version of a C++ compiler and a C++ standard library. GCC 10.2
-or Clang 12.0.0 (or later) as well as libstdc++ 10 or libc++ 7 (or
-later) are recommended.
+mold is written in C++20, so if you build mold yourself, you will need a
+recent version of a C++ compiler and a C++ standard library. We recommend GCC
+10.2 or Clang 12.0.0 (or later) and libstdc++ 10 or libc++ 7 (or later).
 
-### Install dependencies
+### Install Dependencies
 
 To install build dependencies, run `./install-build-deps.sh` in this
-directory. It recognizes your Linux distribution and tries to install
-necessary packages. You may want to run it as root.
+directory. It will detect your Linux distribution and attempt to install the
+necessary packages. You may need to run it as root.
 
 ### Compile mold
 
@@ -75,59 +71,58 @@ necessary packages. You may want to run it as root.
 git clone https://github.com/rui314/mold.git
 mkdir mold/build
 cd mold/build
-git checkout v1.8.0
+git checkout v2.2.0
 ../install-build-deps.sh
 cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=c++ ..
 cmake --build . -j $(nproc)
-sudo cmake --install .
+sudo cmake --build . --target install
 ```
 
-You may need to pass a C++20 compiler command name to `cmake`.
-In the above case, `c++` is passed. If it doesn't work for you,
-try a specific version of a compiler such as `g++-10` or `clang++-12`.
+You might need to pass a C++20 compiler command name to `cmake`. In the
+example above, `c++` is passed. If that doesn't work for you, try a specific
+version of a compiler, such as `g++-10` or `clang++-12`.
 
-By default, `mold` is installed to `/usr/local/bin`. You can change
-that by passing `-DCMAKE_INSTALL_PREFIX=<directory>`. For other cmake
-options, see the comments in `CMakeLists.txt`.
+By default, `mold` is installed to `/usr/local/bin`. You can change the
+installation location by passing `-DCMAKE_INSTALL_PREFIX=<directory>`.
+For other cmake options, see the comments in `CMakeLists.txt`.
 
-If you don't use a recent enough Linux distribution, or if for any reason
-`cmake` in the above commands doesn't work for you, you can use Docker to
-build it in a Docker environment. To do so, just run `./dist.sh` in this
-directory instead of `cmake`. The shell script pulls a Docker image,
-builds mold and auxiliary files inside it, and packs them into a
-single tar file `mold-$version-$arch-linux.tar.gz`.  You can extract
-the tar file anywhere and use `mold` executable in it.
+If you are not using a recent enough Linux distribution, or if `cmake` does
+not work for you for any reason, you can use Docker to build mold in a Docker
+environment. To do so, run `./dist.sh` in this directory instead of using
+`cmake`. The shell script will pull a Docker image, build mold and auxiliary
+files inside it, and package them into a single tar file named
+`mold-$version-$arch-linux.tar.gz`. You can extract the tar file anywhere and
+use the mold executable within it.
 
 ## How to use
 
 <details><summary>A classic way to use mold</summary>
 
-On Unix, the linker command (which is usually `/usr/bin/ld`) is
-invoked indirectly by the compiler driver (which is usually `cc`,
-`gcc` or `clang`), which is typically in turn indirectly invoked by
-`make` or some other build system command.
+On Unix, the linker command (usually `/usr/bin/ld`) is indirectly invoked by
+the compiler driver (typically `cc`, `gcc`, or `clang`), which is in turn
+indirectly invoked by `make` or another build system command.
 
-If you can specify an additional command line option to your compiler
-driver by modifying build system's config files, add one of the
-following flags to use `mold` instead of `/usr/bin/ld`:
+If you can specify an additional command line option for your compiler driver
+by modifying the build system's config files, add one of the following flags
+to use mold instead of `/usr/bin/ld`:
 
-- Clang: pass `-fuse-ld=mold`
+- For Clang: pass `-fuse-ld=mold`
 
-- GCC 12.1.0 or later: pass `-fuse-ld=mold`
+- For GCC 12.1.0 or later: pass `-fuse-ld=mold`
 
-- GCC before 12.1.0: `-fuse-ld` does not accept `mold` as a valid
-  argument, so you need to use `-B` option instead. `-B` is an option
-  to tell GCC where to look for external commands such as `ld`.
+- For GCC before 12.1.0: the `-fuse-ld` option does not accept `mold` as a
+  valid argument, so you need to use the `-B` option instead. The `-B` option
+  tells GCC where to look for external commands like `ld`.
 
-  If you have installed mold with `make install`, there should be a
-  directory named `/usr/libexec/mold` (or `/usr/local/libexec/mold`,
-  depending on your `$PREFIX`), and `ld` command should be there. The
-  `ld` is actually a symlink to `mold`. So, all you need is to pass
-  `-B/usr/libexec/mold` (or `-B/usr/local/libexec/mold`) to GCC.
+  If you have installed mold with `make install`, there should be a directory
+  named `/usr/libexec/mold` (or `/usr/local/libexec/mold`, depending on your
+  `$PREFIX`), and the `ld` command should be there. The `ld` is actually a
+  symlink to `mold`. So, all you need is to pass `-B/usr/libexec/mold` (or
+  `-B/usr/local/libexec/mold`) to GCC.
 
 If you haven't installed `ld.mold` to any `$PATH`, you can still pass
-`-fuse-ld=/absolute/path/to/mold` to clang to use mold. GCC does not
-take an absolute path as an argument for `-fuse-ld` though.
+`-fuse-ld=/absolute/path/to/mold` to clang to use mold. However, GCC does not
+accept an absolute path as an argument for `-fuse-ld`.
 
 </details>
 
@@ -141,32 +136,19 @@ linker = "clang"
 rustflags = ["-C", "link-arg=-fuse-ld=/path/to/mold"]
 ```
 
-where `/path/to/mold` is an absolute path to `mold` exectuable. In the
-above example, we use `clang` as a linker driver as it can always take
-the `-fuse-ld` option. If your GCC is recent enough to recognize the
-option, you may be able to remove the `linker = "clang"` line.
+where `/path/to/mold` is an absolute path to the mold executable. In the
+example above, we use `clang` as a linker driver since it always accepts the
+`-fuse-ld` option. If your GCC is recent enough to recognize the option, you
+may be able to remove the `linker = "clang"` line.
 
 ```toml
 [target.x86_64-unknown-linux-gnu]
 rustflags = ["-C", "link-arg=-fuse-ld=/path/to/mold"]
 ```
 
-If you want to use mold for all projects, put the above snippet to
+If you want to use mold for all projects, add the above snippet to
 `~/.cargo/config.toml`.
 
-If you are using macOS, you can modify `config.toml` in a similar manner.
-Here is an example with `mold` installed via [Homebrew](https://brew.sh).
-
-```toml
-[target.x86_64-apple-darwin]
-linker = "clang"
-rustflags = ["-C", "link-arg=-fuse-ld=mold"]
-
-[target.aarch64-apple-darwin]
-linker = "clang"
-rustflags = ["-C", "link-arg=-fuse-ld=mold"]
-```
-
 </details>
 
 <details><summary>If you are using Nim</summary>
@@ -178,67 +160,46 @@ when findExe("mold").len > 0 and defined(linux):
   switch("passL", "-fuse-ld=mold")
 ```
 
-where `mold` must be included in the PATH environment variable. In this example
-The above example uses `gcc` as the linker driver.
-Use the `fuse-ld` option. If your GCC is recent enough to recognize this option.
+where `mold` must be included in the `PATH` environment variable. In this
+example, `gcc` is used as the linker driver. Use the `-fuse-ld` option if your
+GCC is recent enough to recognize this option.
 
-If you want to use mold for all projects, put the above snippet to `~/.config/config.nims`.
-
-If you are using macOS, you can modify config.nims in a similar manner. Here is an example with mold installed via Homebrew.
-
-```nim
-when findExe("ld64.mold").len > 0 and defined(macosx):
-  switch("passL", "-fuse-ld=ld64.mold")
-```
+If you want to use mold for all projects, add the above snippet to
+`~/.config/config.nims`.
 
 </details>
 
 <details><summary>mold -run</summary>
 
-It is sometimes very hard to pass an appropriate command line option
-to `cc` to specify an alternative linker.  To deal with the situation,
-mold has a feature to intercept all invocations of `ld`, `ld.lld` or
-`ld.gold` and redirect it to itself. To use the feature, run `make`
-(or another build command) as a subcommand of mold as follows:
+It is sometimes very hard to pass an appropriate command line option to `cc`
+to specify an alternative linker. To address this situation, mold has a
+feature to intercept all invocations of `ld`, `ld.lld`, or `ld.gold` and
+redirect them to itself. To use this feature, run `make` (or another build
+command) as a subcommand of mold as follows:
 
 ```shell
 mold -run make <make-options-if-any>
 ```
 
-Internally, mold invokes a given command with `LD_PRELOAD` environment
-variable set to its companion shared object file. The shared object
-file intercepts all function calls to `exec(3)`-family functions to
-replace `argv[0]` with `mold` if it is `ld`, `ld.gold` or `ld.lld`.
-
-</details>
-
-<details><summary>On macOS</summary>
-
-mold/macOS is available as an alpha version. It can be used to build not
-only macOS apps but also iOS apps because their binary formats are the same.
-
-The command name of mold/macOS is `ld64.mold`. If you build mold on macOS,
-it still produces `mold` and `ld.mold`, but these executables are useful
-only for cross compilation (i.e. building Linux apps on macOS.)
-
-If you find any issue with mold/macOS, please file it to
-<a href=https://github.com/rui314/mold/issues>our GitHub Issues</a>.
+Internally, mold invokes the given command with the `LD_PRELOAD` environment
+variable set to its companion shared object file. The shared object file
+intercepts all function calls to `exec(3)`-family functions to replace
+`argv[0]` with `mold` if it is `ld`, `ld.gold`, or `ld.lld`.
 
 </details>
 
 <details><summary>GitHub Actions</summary>
 
-You can use our <a href=https://github.com/rui314/setup-mold>setup-mold</a>
-GitHub Action to speed up GitHub-hosted continuous build. GitHub Actions
-runs on a two-core machine, but mold is still significantly faster than
-the default GNU linker there especially when a program being linked is
-large.
+You can use our [setup-mold](https://github.com/rui314/setup-mold) GitHub
+Action to speed up GitHub-hosted continuous builds. Although GitHub Actions
+run on a two-core machine, mold is still significantly faster than the default
+GNU linker, especially when linking large programs.
 
 </details>
 
 <details><summary>Verify that you are using mold</summary>
 
-mold leaves its identification string in `.comment` section in an output
+mold leaves its identification string in the `.comment` section of an output
 file. You can print it out to verify that you are actually using mold.
 
 ```shell
@@ -249,70 +210,53 @@ String dump of section '.comment':
   [    2b]  mold 9a1679b47d9b22012ec7dfbda97c8983956716f7
 ```
 
-If `mold` is in `.comment`, the file is created by mold.
+If `mold` is present in the `.comment` section, the file was created by mold.
 
 </details>
 
 <details><summary>Online manual</summary>
 
-Since mold is a drop-in replacement, you should be able to use it
-without reading its manual. But just in case you need it, it's available
-online at <a href=https://rui314.github.io/mold.html>here</a>.
-You can also read the same manual by `man mold`.
+Since mold is a drop-in replacement, you should be able to use it without
+reading its manual. However, if you need it, [mold's man page](docs/mold.md)
+is available. You can read the same manual by running `man mold`.
 
 </details>
 
 ## Why is mold so fast?
 
-One reason is because it simply uses faster algorithms and efficient
-data structures than other linkers do. The other reason is that the
-new linker is highly parallelized.
+One reason is that it utilizes faster algorithms and more efficient data
+structures compared to other linkers. Another reason is that mold is highly
+parallelized.
 
-Here is a side-by-side comparison of per-core CPU usage of lld (left)
-and mold (right). They are linking the same program, Chromium
-executable.
+Here is a side-by-side comparison of per-core CPU usage for lld (left) and
+mold (right), linking the same program, a Chromium executable.
 
 ![CPU usage comparison in htop animation](docs/htop.gif)
 
-As you can see, mold uses all available cores throughout its execution
-and finishes quickly. On the other hand, lld failed to use available
-cores most of the time. In this demo, the maximum parallelism is
-artificially capped to 16 so that the bars fit in the GIF.
+As you can see, mold uses all available cores throughout its execution and
+finishes quickly. In contrast, lld fails to utilize available cores most of
+the time. In this demo, the maximum parallelism is artificially capped at 16,
+so that the bars fit in the GIF.
 
-For details, please read [design notes](docs/design.md).
-
-## License
-
-mold is available under AGPL. Note that that does not mean that you
-have to license your program under AGPL if you use mold to link your
-program. An output of the mold linker is a derived work of the object
-files and libraries you pass to the linker but not a derived work of
-the mold linker itself.
-
-Besides that, you can also buy a commercial, non-AGPL license with
-technical support from our company, Blue Whale Systems PTE LTD. If you
-are a big company, please consider obtaining it before making hundreds
-or thousands of developers of your company to depend on mold. mold is
-mostly a single-person open-source project, and just like other
-open-source projects, we are not legally obligated to keep maintaining
-it. A legally-binding commercial license contract addresses the
-concern. By purchasing a license, you are guaranteed that mold will be
-maintained for you. Please [contact us](mailto:[email protected])
-for a commercial license inquiry.
+For details, please see the [design notes](docs/design.md).
 
 ## Sponsors
 
 We accept donations via [GitHub Sponsors](https://github.com/sponsors/rui314)
-and [OpenCollective](https://opencollective.com/mold-linker).
-We thank you to everybody who sponsors our project. In particular,
-we'd like to acknowledge the following people and organizations who
-have sponsored $128/month or more:
+and [OpenCollective](https://opencollective.com/mold-linker). We thank
+everyone who sponsors our project. In particular, we'd like to acknowledge the
+following people and organizations who have sponsored $128/month or more:
 
 ### Corporate sponsors
 
-<a href="https://mercury.com/"><img src="docs/mercury-logo.png" align=center height=120 width=400 alt=Mercury></a>
+<a href="https://mercury.com"><img src="docs/mercury-logo.png" align=center height=120 width=400 alt=Mercury></a>
+
+<a href="https://cybozu-global.com"><img src="docs/cyboze-logo.png" align=center height=120 width=133 alt=Cybozu></a>
+
+<a href="https://www.emergetools.com"><img src="docs/emerge-tools-logo.png" align=center height=120 width=240 alt="Emerge Tools"></a><br>
 
 - [Uber](https://uber.com)
+- [G-Research](https://www.gresearch.co.uk)
 - [Signal Slot Inc.](https://github.com/signal-slot)
 
 ### Individual sponsors

common/Dockerfile

@@ -11,13 +11,7 @@ RUN apt-get update && \
   apt-get update && \
   apt-get install -y --no-install-recommends build-essential wget libstdc++-11-dev zlib1g-dev gcc-10 g++-10 python3 && \
   \
-  mkdir /openssl && cd /openssl && \
-  wget -O- -q https://www.openssl.org/source/openssl-3.0.7.tar.gz | tar --strip-components=1 -xzf - && \
-  ./Configure --prefix=/usr/local --libdir=lib && \
-  make -j$(nproc) && \
-  make -j$(nproc) install && \
-  ldconfig && \
   mkdir /cmake && cd /cmake && \
   wget -O- -q https://github.com/Kitware/CMake/releases/download/v3.24.2/cmake-3.24.2.tar.gz | tar --strip-components=1 -xzf - && \
   ./bootstrap --parallel=$(nproc) && make -j$(nproc) && make -j$(nproc) install && \
-  rm -rf /var/lib/apt/lists/* /cmake /openssl
+  rm -rf /var/lib/apt/lists/* /cmake

common/archive-file.h

@@ -25,7 +25,7 @@
 
 #pragma once
 
-#include "mold.h"
+#include "common.h"
 #include "filetype.h"
 
 namespace mold {
@@ -76,12 +76,12 @@ struct ArHdr {
   }
 };
 
-template <typename C>
-std::vector<MappedFile<C> *>
-read_thin_archive_members(C &ctx, MappedFile<C> *mf) {
+template <typename Context, typename MappedFile>
+std::vector<MappedFile *>
+read_thin_archive_members(Context &ctx, MappedFile *mf) {
   u8 *begin = mf->data;
   u8 *data = begin + 8;
-  std::vector<MappedFile<C> *> vec;
+  std::vector<MappedFile *> vec;
   std::string_view strtab;
 
   while (data < begin + mf->size) {
@@ -117,19 +117,18 @@ read_thin_archive_members(C &ctx, MappedFile<C> *mf) {
 
     std::string path = name.starts_with('/') ?
       name : (filepath(mf->name).parent_path() / name).string();
-    vec.push_back(MappedFile<C>::must_open(ctx, path));
+    vec.push_back(MappedFile::must_open(ctx, path));
     vec.back()->thin_parent = mf;
     data = body;
   }
   return vec;
 }
 
-template <typename C>
-std::vector<MappedFile<C> *>
-read_fat_archive_members(C &ctx, MappedFile<C> *mf) {
+template <typename Context, typename MappedFile>
+std::vector<MappedFile *> read_fat_archive_members(Context &ctx, MappedFile *mf) {
   u8 *begin = mf->data;
   u8 *data = begin + 8;
-  std::vector<MappedFile<C> *> vec;
+  std::vector<MappedFile *> vec;
   std::string_view strtab;
 
   while (begin + mf->size - data >= 2) {
@@ -163,10 +162,9 @@ read_fat_archive_members(C &ctx, MappedFile<C> *mf) {
   return vec;
 }
 
-template <typename C>
-std::vector<MappedFile<C> *>
-read_archive_members(C &ctx, MappedFile<C> *mf) {
-  switch (get_file_type(mf, !ctx.arg.plugin.empty())) {
+template <typename Context, typename MappedFile>
+std::vector<MappedFile *> read_archive_members(Context &ctx, MappedFile *mf) {
+  switch (get_file_type(ctx, mf)) {
   case FileType::AR:
     return read_fat_archive_members(ctx, mf);
   case FileType::THIN_AR:

common/cmdline.h

@@ -1,14 +1,14 @@
 #pragma once
 
-#include "mold.h"
+#include "common.h"
 
 namespace mold {
 
-template <typename C>
+template <typename Context>
 std::vector<std::string_view>
-read_response_file(C &ctx, std::string_view path) {
+read_response_file(Context &ctx, std::string_view path) {
   std::vector<std::string_view> vec;
-  MappedFile<C> *mf = MappedFile<C>::must_open(ctx, std::string(path));
+  MappedFile<Context> *mf = MappedFile<Context>::must_open(ctx, std::string(path));
   u8 *data = mf->data;
 
   auto read_quoted = [&](i64 i, char quote) {
@@ -63,8 +63,8 @@ read_response_file(C &ctx, std::string_view path) {
 }
 
 // Replace "@path/to/some/text/file" with its file contents.
-template <typename C>
-std::vector<std::string_view> expand_response_files(C &ctx, char **argv) {
+template <typename Context>
+std::vector<std::string_view> expand_response_files(Context &ctx, char **argv) {
   std::vector<std::string_view> vec;
   for (i64 i = 0; argv[i]; i++) {
     if (argv[i][0] == '@')

common/common.h

@@ -8,6 +8,7 @@
 #include <bitset>
 #include <cassert>
 #include <cstdio>
+#include <cstdlib>
 #include <cstring>
 #include <fcntl.h>
 #include <filesystem>
@@ -32,7 +33,7 @@
 #endif
 
 #define XXH_INLINE_ALL 1
-#include "third-party/xxhash/xxhash.h"
+#include "../third-party/xxhash/xxhash.h"
 
 #ifdef NDEBUG
 # define unreachable() __builtin_unreachable()
@@ -40,6 +41,23 @@
 # define unreachable() assert(0 && "unreachable")
 #endif
 
+// __builtin_assume() is supported only by clang, and [[assume]] is
+// available only in C++23, so we use this macro when giving a hint to
+// the compiler's optimizer what's true.
+#define ASSUME(x) do { if (!(x)) __builtin_unreachable(); } while (0)
+
+// This is an assert() that is enabled even in the release build.
+#define ASSERT(x)                                       \
+  do {                                                  \
+    if (!(x)) {                                         \
+      std::cerr << "Assertion failed: (" << #x          \
+                << "), function " << __FUNCTION__       \
+                << ", file " << __FILE__                \
+                << ", line " << __LINE__ << ".\n";      \
+      std::abort();                                     \
+    }                                                   \
+  } while (0)
+
 inline uint64_t hash_string(std::string_view str) {
   return XXH3_64bits(str.data(), str.size());
 }
@@ -60,7 +78,7 @@ namespace mold {
 using namespace std::literals::string_literals;
 using namespace std::literals::string_view_literals;
 
-template <typename C> class OutputFile;
+template <typename Context> class OutputFile;
 
 inline char *output_tmpfile;
 inline thread_local bool opt_demangle;
@@ -86,41 +104,44 @@ static u64 combine_hash(u64 a, u64 b) {
 // Error output
 //
 
-template <typename C>
+template <typename Context>
 class SyncOut {
 public:
-  SyncOut(C &ctx, std::ostream &out = std::cout) : out(out) {
+  SyncOut(Context &ctx, std::ostream *out = &std::cout) : out(out) {
     opt_demangle = ctx.arg.demangle;
   }
 
   ~SyncOut() {
-    std::scoped_lock lock(mu);
-    out << ss.str() << "\n";
+    if (out) {
+      std::scoped_lock lock(mu);
+      *out << ss.str() << "\n";
+    }
   }
 
   template <class T> SyncOut &operator<<(T &&val) {
-    ss << std::forward<T>(val);
+    if (out)
+      ss << std::forward<T>(val);
     return *this;
   }
 
   static inline std::mutex mu;
 
 private:
-  std::ostream &out;
+  std::ostream *out;
   std::stringstream ss;
 };
 
-template <typename C>
-static std::string add_color(C &ctx, std::string msg) {
+template <typename Context>
+static std::string add_color(Context &ctx, std::string msg) {
   if (ctx.arg.color_diagnostics)
     return "mold: \033[0;1;31m" + msg + ":\033[0m ";
   return "mold: " + msg + ": ";
 }
 
-template <typename C>
+template <typename Context>
 class Fatal {
 public:
-  Fatal(C &ctx) : out(ctx, std::cerr) {
+  Fatal(Context &ctx) : out(ctx, &std::cerr) {
     out << add_color(ctx, "fatal");
   }
 
@@ -136,13 +157,13 @@ public:
   }
 
 private:
-  SyncOut<C> out;
+  SyncOut<Context> out;
 };
 
-template <typename C>
+template <typename Context>
 class Error {
 public:
-  Error(C &ctx) : out(ctx, std::cerr) {
+  Error(Context &ctx) : out(ctx, &std::cerr) {
     if (ctx.arg.noinhibit_exec) {
       out << add_color(ctx, "warning");
     } else {
@@ -157,13 +178,14 @@ public:
   }
 
 private:
-  SyncOut<C> out;
+  SyncOut<Context> out;
 };
 
-template <typename C>
+template <typename Context>
 class Warn {
 public:
-  Warn(C &ctx) : out(ctx, std::cerr) {
+  Warn(Context &ctx)
+    : out(ctx, ctx.arg.suppress_warnings ? nullptr : &std::cerr) {
     if (ctx.arg.fatal_warnings) {
       out << add_color(ctx, "error");
       ctx.has_error = true;
@@ -178,7 +200,64 @@ public:
   }
 
 private:
-  SyncOut<C> out;
+  SyncOut<Context> out;
+};
+
+//
+// Atomics
+//
+
+// This is the same as std::atomic except that the default memory
+// order is relaxed instead of sequential consistency.
+template <typename T>
+struct Atomic : std::atomic<T> {
+  static constexpr std::memory_order relaxed = std::memory_order_relaxed;
+
+  using std::atomic<T>::atomic;
+
+  Atomic(const Atomic<T> &other) { store(other.load()); }
+
+  Atomic<T> &operator=(const Atomic<T> &other) {
+    store(other.load());
+    return *this;
+  }
+
+  void operator=(T val) { store(val); }
+  operator T() const { return load(); }
+
+  void store(T val) { std::atomic<T>::store(val, relaxed); }
+  T load() const { return std::atomic<T>::load(relaxed); }
+  T exchange(T val) { return std::atomic<T>::exchange(val, relaxed); }
+  T operator|=(T val) { return std::atomic<T>::fetch_or(val, relaxed); }
+  T operator++() { return std::atomic<T>::fetch_add(1, relaxed) + 1; }
+  T operator--() { return std::atomic<T>::fetch_sub(1, relaxed) - 1; }
+  T operator++(int) { return std::atomic<T>::fetch_add(1, relaxed); }
+  T operator--(int) { return std::atomic<T>::fetch_sub(1, relaxed); }
+
+  bool test_and_set() {
+    // A relaxed load + branch (assuming miss) takes only around 20 cycles,
+    // while an atomic RMW can easily take hundreds on x86. We note that it's
+    // common that another thread beat us in marking, so doing an optimistic
+    // early test tends to improve performance in the ~20% ballpark.
+    return load() || exchange(true);
+  }
+};
+
+//
+// Bit vector
+//
+
+class BitVector {
+public:
+  BitVector() = default;
+  BitVector(u32 size) : vec((size + 7) / 8) {}
+
+  void resize(u32 size) { vec.resize((size + 7) / 8); }
+  bool get(u32 idx) const { return vec[idx / 8] & (1 << (idx % 8)); }
+  void set(u32 idx) { vec[idx / 8] |= 1 << (idx % 8); }
+
+private:
+  std::vector<u8> vec;
 };
 
 //
@@ -220,7 +299,7 @@ inline u64 bits(u64 val, u64 hi, u64 lo) {
 
 inline i64 sign_extend(u64 val, i64 size) {
   return (i64)(val << (63 - size)) >> (63 - size);
-};
+}
 
 template <typename T, typename Compare = std::less<T>>
 void update_minimum(std::atomic<T> &atomic, u64 new_val, Compare cmp = {}) {
@@ -238,17 +317,6 @@ void update_maximum(std::atomic<T> &atomic, u64 new_val, Compare cmp = {}) {
                                        std::memory_order_relaxed));
 }
 
-// An optimized "mark" operation for parallel mark-and-sweep algorithms.
-// Returns true if `visited` was false and updated to true.
-inline bool fast_mark(std::atomic<bool> &visited) {
-  // A relaxed load + branch (assuming miss) takes only around 20 cycles,
-  // while an atomic RMW can easily take hundreds on x86. We note that it's
-  // common that another thread beat us in marking, so doing an optimistic
-  // early test tends to improve performance in the ~20% ballpark.
-  return !visited.load(std::memory_order_relaxed) &&
-         !visited.exchange(true, std::memory_order_relaxed);
-}
-
 template <typename T, typename U>
 inline void append(std::vector<T> &vec1, std::vector<U> vec2) {
   vec1.insert(vec1.end(), vec2.begin(), vec2.end());
@@ -256,7 +324,12 @@ inline void append(std::vector<T> &vec1, std::vector<U> vec2) {
 
 template <typename T>
 inline std::vector<T> flatten(std::vector<std::vector<T>> &vec) {
+  i64 size = 0;
+  for (std::vector<T> &v : vec)
+    size += v.size();
+
   std::vector<T> ret;
+  ret.reserve(size);
   for (std::vector<T> &v : vec)
     append(ret, v);
   return ret;
@@ -322,30 +395,36 @@ inline i64 write_uleb(u8 *buf, u64 val) {
   return i;
 }
 
-inline u64 read_uleb(u8 *&buf) {
+inline u64 read_uleb(u8 **buf) {
   u64 val = 0;
   u8 shift = 0;
   u8 byte;
   do {
-    byte = *buf++;
+    byte = *(*buf)++;
     val |= (byte & 0x7f) << shift;
     shift += 7;
   } while (byte & 0x80);
   return val;
 }
 
-inline u64 read_uleb(u8 const*&buf) {
-  return read_uleb(const_cast<u8 *&>(buf));
+inline u64 read_uleb(u8 *buf) {
+  u8 *tmp = buf;
+  return read_uleb(&tmp);
 }
 
-inline u64 read_uleb(std::string_view &str) {
-  u8 *start = (u8 *)&str[0];
+inline u64 read_uleb(std::string_view *str) {
+  u8 *start = (u8 *)str->data();
   u8 *ptr = start;
-  u64 val = read_uleb(ptr);
-  str = str.substr(ptr - start);
+  u64 val = read_uleb(&ptr);
+  *str = str->substr(ptr - start);
   return val;
 }
 
+inline u64 read_uleb(std::string_view str) {
+  std::string_view tmp = str;
+  return read_uleb(&tmp);
+}
+
 inline i64 uleb_size(u64 val) {
 #if __GNUC__
 #pragma GCC unroll 8
@@ -356,8 +435,16 @@ inline i64 uleb_size(u64 val) {
   return 9;
 }
 
-template <typename C>
-std::string_view save_string(C &ctx, const std::string &str) {
+inline void overwrite_uleb(u8 *loc, u64 val) {
+  while (*loc & 0b1000'0000) {
+    *loc++ = 0b1000'0000 | (val & 0b0111'1111);
+    val >>= 7;
+  }
+  *loc = val & 0b0111'1111;
+}
+
+template <typename Context>
+std::string_view save_string(Context &ctx, const std::string &str) {
   u8 *buf = new u8[str.size() + 1];
   memcpy(buf, str.data(), str.size());
   buf[str.size()] = '\0';
@@ -392,53 +479,72 @@ public:
   }
 
   ~ConcurrentMap() {
-    if (keys) {
-      free((void *)keys);
-      free((void *)key_sizes);
-      free((void *)values);
-    }
+    free(entries);
   }
 
+  // In order to avoid unnecessary cache-line false sharing, we want
+  // to make this object to be aligned to a reasonably large
+  // power-of-two address.
+  struct alignas(32) Entry {
+    std::atomic<const char *> key;
+    T value;
+    u32 keylen;
+  };
+
   void resize(i64 nbuckets) {
-    this->~ConcurrentMap();
+    this->nbuckets = std::max<i64>(MIN_NBUCKETS, bit_ceil(nbuckets));
 
-    nbuckets = std::max<i64>(MIN_NBUCKETS, bit_ceil(nbuckets));
+    i64 sz = sizeof(Entry) * this->nbuckets;
+    free(entries);
 
-    this->nbuckets = nbuckets;
-    keys = (std::atomic<const char *> *)calloc(nbuckets, sizeof(keys[0]));
-    key_sizes = (u32 *)calloc(nbuckets, sizeof(key_sizes[0]));
-    values = (T *)calloc(nbuckets, sizeof(values[0]));
+#if _WIN32
+    // Even though std::aligned_alloc is defined in C++17, MSVC doesn't
+    // seem to provide that function.
+    entries = (Entry *)_aligned_malloc(sz, alignof(Entry));
+#else
+    entries = (Entry *)std::aligned_alloc(alignof(Entry), sz);
+#endif
+
+    memset(entries, 0, sz);
   }
 
   std::pair<T *, bool> insert(std::string_view key, u64 hash, const T &val) {
-    if (!keys)
-      return {nullptr, false};
-
     assert(has_single_bit(nbuckets));
+
     i64 idx = hash & (nbuckets - 1);
     i64 retry = 0;
 
     while (retry < MAX_RETRY) {
-      const char *ptr = keys[idx].load(std::memory_order_acquire);
-      if (ptr == marker) {
-        pause();
+      Entry &ent = entries[idx];
+      const char *ptr = nullptr;
+      bool claimed = ent.key.compare_exchange_weak(ptr, (char *)-1,
+                                                   std::memory_order_acquire);
+
+      // If we successfully claimed the ownership of an unused slot,
+      // copy values to it.
+      if (claimed) {
+        new (&ent.value) T(val);
+        ent.keylen = key.size();
+        ent.key.store(key.data(), std::memory_order_release);
+        return {&ent.value, true};
+      }
+
+      // Loop on a spurious failure.
+      if (ptr == nullptr)
         continue;
+
+      // If someone is copying values to the slot, do busy wait.
+      while (ptr == (char *)-1) {
+        pause();
+        ptr = ent.key.load(std::memory_order_acquire);
       }
 
-      if (ptr == nullptr) {
-        if (!keys[idx].compare_exchange_weak(ptr, marker,
-                                             std::memory_order_acq_rel))
-          continue;
-        new (values + idx) T(val);
-        key_sizes[idx] = key.size();
-        keys[idx].store(key.data(), std::memory_order_release);
-        return {values + idx, true};
-      }
-
-      if (key.size() == key_sizes[idx] &&
-          memcmp(ptr, key.data(), key_sizes[idx]) == 0)
-        return {values + idx, false};
+      // If the same key is already present, this is the slot we are
+      // looking for.
+      if (key == std::string_view(ptr, ent.keylen))
+        return {&ent.value, false};
 
+      // Otherwise, move on to the next slot.
       u64 mask = nbuckets / NUM_SHARDS - 1;
       idx = (idx & ~mask) | ((idx + 1) & mask);
       retry++;
@@ -448,18 +554,21 @@ public:
     return {nullptr, false};
   }
 
-  bool has_key(i64 idx) {
-    return keys[idx].load(std::memory_order_relaxed);
+  const char *get_key(i64 idx) {
+    return entries[idx].key.load(std::memory_order_relaxed);
+  }
+
+  i64 get_idx(T *value) const {
+    uintptr_t addr = (uintptr_t)value - (uintptr_t)value % sizeof(Entry);
+    return (Entry *)addr - entries;
   }
 
   static constexpr i64 MIN_NBUCKETS = 2048;
   static constexpr i64 NUM_SHARDS = 16;
   static constexpr i64 MAX_RETRY = 128;
 
+  Entry *entries = nullptr;
   i64 nbuckets = 0;
-  std::atomic<const char *> *keys = nullptr;
-  u32 *key_sizes = nullptr;
-  T *values = nullptr;
 
 private:
   static void pause() {
@@ -469,21 +578,19 @@ private:
     asm volatile("yield");
 #endif
   }
-
-  static constexpr const char *marker = "marker";
 };
 
 //
 // output-file.h
 //
 
-template <typename C>
+template <typename Context>
 class OutputFile {
 public:
-  static std::unique_ptr<OutputFile<C>>
-  open(C &ctx, std::string path, i64 filesize, i64 perm);
+  static std::unique_ptr<OutputFile<Context>>
+  open(Context &ctx, std::string path, i64 filesize, i64 perm);
 
-  virtual void close(C &ctx) = 0;
+  virtual void close(Context &ctx) = 0;
   virtual ~OutputFile() = default;
 
   u8 *buf = nullptr;
@@ -646,13 +753,13 @@ public:
   }
 
   Counter &operator++(int) {
-    if (enabled)
+    if (enabled) [[unlikely]]
       values.local()++;
     return *this;
   }
 
   Counter &operator+=(int delta) {
-    if (enabled)
+    if (enabled) [[unlikely]]
       values.local() += delta;
     return *this;
   }
@@ -689,10 +796,10 @@ struct TimerRecord {
 void
 print_timer_records(tbb::concurrent_vector<std::unique_ptr<TimerRecord>> &);
 
-template <typename C>
+template <typename Context>
 class Timer {
 public:
-  Timer(C &ctx, std::string name, Timer *parent = nullptr) {
+  Timer(Context &ctx, std::string name, Timer *parent = nullptr) {
     record = new TimerRecord(name, parent ? parent->record : nullptr);
     ctx.timer_records.push_back(std::unique_ptr<TimerRecord>(record));
   }
@@ -743,16 +850,16 @@ private:
 
 // MappedFile represents an mmap'ed input file.
 // mold uses mmap-IO only.
-template <typename C>
+template <typename Context>
 class MappedFile {
 public:
-  static MappedFile *open(C &ctx, std::string path);
-  static MappedFile *must_open(C &ctx, std::string path);
+  static MappedFile *open(Context &ctx, std::string path);
+  static MappedFile *must_open(Context &ctx, std::string path);
 
   ~MappedFile() { unmap(); }
   void unmap();
 
-  MappedFile *slice(C &ctx, std::string name, u64 start, u64 size);
+  MappedFile *slice(Context &ctx, std::string name, u64 start, u64 size);
 
   std::string_view get_contents() {
     return std::string_view((char *)data, size);
@@ -782,7 +889,6 @@ public:
   std::string name;
   u8 *data = nullptr;
   i64 size = 0;
-  i64 mtime = 0;
   bool given_fullpath = true;
   MappedFile *parent = nullptr;
   MappedFile *thin_parent = nullptr;
@@ -792,8 +898,8 @@ public:
 #endif
 };
 
-template <typename C>
-MappedFile<C> *MappedFile<C>::open(C &ctx, std::string path) {
+template <typename Context>
+MappedFile<Context> *MappedFile<Context>::open(Context &ctx, std::string path) {
   if (path.starts_with('/') && !ctx.arg.chroot.empty())
     path = ctx.arg.chroot + "/" + path_clean(path);
 
@@ -804,8 +910,11 @@ MappedFile<C> *MappedFile<C>::open(C &ctx, std::string path) {
     fd = ::open(path.c_str(), O_RDONLY);
 #endif
 
-  if (fd == -1)
+  if (fd == -1) {
+    if (errno != ENOENT)
+      Fatal(ctx) << "opening " << path << " failed: " << errno_string();
     return nullptr;
+  }
 
   struct stat st;
   if (fstat(fd, &st) == -1)
@@ -817,14 +926,6 @@ MappedFile<C> *MappedFile<C>::open(C &ctx, std::string path) {
   mf->name = path;
   mf->size = st.st_size;
 
-#ifdef _WIN32
-    mf->mtime = st.st_mtime;
-#elif defined(__APPLE__)
-    mf->mtime = (u64)st.st_mtimespec.tv_sec * 1000000000 + st.st_mtimespec.tv_nsec;
-#else
-    mf->mtime = (u64)st.st_mtim.tv_sec * 1000000000 + st.st_mtim.tv_nsec;
-#endif
-
   if (st.st_size > 0) {
 #ifdef _WIN32
     HANDLE handle = CreateFileMapping((HANDLE)_get_osfhandle(fd),
@@ -848,17 +949,18 @@ MappedFile<C> *MappedFile<C>::open(C &ctx, std::string path) {
   return mf;
 }
 
-template <typename C>
-MappedFile<C> *MappedFile<C>::must_open(C &ctx, std::string path) {
+template <typename Context>
+MappedFile<Context> *
+MappedFile<Context>::must_open(Context &ctx, std::string path) {
   if (MappedFile *mf = MappedFile::open(ctx, path))
     return mf;
   Fatal(ctx) << "cannot open " << path << ": " << errno_string();
 }
 
-template <typename C>
-MappedFile<C> *
-MappedFile<C>::slice(C &ctx, std::string name, u64 start, u64 size) {
-  MappedFile *mf = new MappedFile<C>;
+template <typename Context>
+MappedFile<Context> *
+MappedFile<Context>::slice(Context &ctx, std::string name, u64 start, u64 size) {
+  MappedFile *mf = new MappedFile;
   mf->name = name;
   mf->data = data + start;
   mf->size = size;
@@ -868,8 +970,8 @@ MappedFile<C>::slice(C &ctx, std::string name, u64 start, u64 size) {
   return mf;
 }
 
-template <typename C>
-void MappedFile<C>::unmap() {
+template <typename Context>
+void MappedFile<Context>::unmap() {
   if (size == 0 || parent || !data)
     return;
 

common/compress.cc

@@ -14,7 +14,7 @@
 // a little bit. However, if a shard size is large enough, that loss
 // is negligible in practice.
 
-#include "mold.h"
+#include "common.h"
 
 #include <tbb/parallel_for_each.h>
 #include <zlib.h>

common/demangle.cc

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 
 #include <cstdlib>
 
@@ -6,7 +6,7 @@
 #include <cxxabi.h>
 #endif
 
-#include "third-party/rust-demangle/rust-demangle.h"
+#include "../third-party/rust-demangle/rust-demangle.h"
 
 namespace mold {
 

common/filepath.cc

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 
 #include <filesystem>
 #include <sys/stat.h>
@@ -7,8 +7,10 @@ namespace mold {
 
 std::string get_realpath(std::string_view path) {
   std::error_code ec;
-  std::string ret = std::filesystem::canonical(path, ec).string();
-  return ec ? std::string(path) : ret;
+  std::filesystem::path link = std::filesystem::read_symlink(path, ec);
+  if (ec)
+    return std::string(path);
+  return (filepath(path) / ".." / link).lexically_normal().string();
 }
 
 // Removes redundant '/..' or '/.' from a given path.

common/filetype.h

@@ -1,7 +1,7 @@
 #pragma once
 
-#include "mold.h"
-#include "elf/elf.h"
+#include "common.h"
+#include "../elf/elf.h"
 
 namespace mold {
 
@@ -23,15 +23,15 @@ enum class FileType {
   LLVM_BITCODE,
 };
 
-template <typename C>
-bool is_text_file(MappedFile<C> *mf) {
+template <typename MappedFile>
+bool is_text_file(MappedFile *mf) {
   u8 *data = mf->data;
   return mf->size >= 4 && isprint(data[0]) && isprint(data[1]) &&
          isprint(data[2]) && isprint(data[3]);
 }
 
-template <typename E, typename C>
-inline bool is_gcc_lto_obj(MappedFile<C> *mf, bool opt_plugin) {
+template <typename E, typename Context, typename MappedFile>
+inline bool is_gcc_lto_obj(Context &ctx, MappedFile *mf) {
   using namespace mold::elf;
 
   const char *data = mf->get_contents().data();
@@ -45,21 +45,23 @@ inline bool is_gcc_lto_obj(MappedFile<C> *mf, bool opt_plugin) {
     ? sh_begin->sh_link : ehdr.e_shstrndx;
 
   for (ElfShdr<E> &sec : shdrs) {
-    // GCC LTO object contains only sections symbols followed by a common
-    // symbol whose name is `__gnu_lto_slim` (or `__gnu_lto_v1` for older
-    // GCC releases).
-    //
-    // However, FAT LTO objects don't have any of the above mentioned symbols
-    // and can identify LTO by `.gnu.lto_.symtab.` section, similarly
-    // to what lto-plugin does. However, if LTO linker plug-in is not available,
-    // use the emitted assembly instead.
-    std::string_view name = data + shdrs[shstrtab_idx].sh_offset + sec.sh_name;
-    if (opt_plugin && name.starts_with (".gnu.lto_.symtab."))
+    // GCC FAT LTO objects contain both regular ELF sections and GCC-
+    // specific LTO sections, so that they can be linked as LTO objects if
+    // the LTO linker plugin is available and falls back as regular
+    // objects otherwise. GCC FAT LTO object can be identified by the
+    // presence of `.gcc.lto_.symtab` section.
+    if (!ctx.arg.plugin.empty()) {
+      std::string_view name = data + shdrs[shstrtab_idx].sh_offset + sec.sh_name;
+      if (name.starts_with(".gnu.lto_.symtab."))
       return true;
+    }
 
     if (sec.sh_type != SHT_SYMTAB)
       continue;
 
+    // GCC non-FAT LTO object contains only sections symbols followed by
+    // a common symbol whose name is `__gnu_lto_slim` (or `__gnu_lto_v1`
+    // for older GCC releases).
     std::span<ElfSym<E>> elf_syms{(ElfSym<E> *)(data + sec.sh_offset),
                                   (size_t)sec.sh_size / sizeof(ElfSym<E>)};
 
@@ -83,8 +85,8 @@ inline bool is_gcc_lto_obj(MappedFile<C> *mf, bool opt_plugin) {
   return false;
 }
 
-template <typename C>
-FileType get_file_type(MappedFile<C> *mf, bool opt_plugin) {
+template <typename Context, typename MappedFile>
+FileType get_file_type(Context &ctx, MappedFile *mf) {
   using namespace elf;
 
   std::string_view data = mf->get_contents();
@@ -100,10 +102,10 @@ FileType get_file_type(MappedFile<C> *mf, bool opt_plugin) {
 
       if (ehdr.e_type == ET_REL) {
         if (ehdr.e_ident[EI_CLASS] == ELFCLASS32) {
-          if (is_gcc_lto_obj<I386>(mf, opt_plugin))
+          if (is_gcc_lto_obj<I386>(ctx, mf))
             return FileType::GCC_LTO_OBJ;
         } else {
-          if (is_gcc_lto_obj<X86_64>(mf, opt_plugin))
+          if (is_gcc_lto_obj<X86_64>(ctx, mf))
             return FileType::GCC_LTO_OBJ;
         }
         return FileType::ELF_OBJ;
@@ -116,10 +118,10 @@ FileType get_file_type(MappedFile<C> *mf, bool opt_plugin) {
 
       if (ehdr.e_type == ET_REL) {
         if (ehdr.e_ident[EI_CLASS] == ELFCLASS32) {
-          if (is_gcc_lto_obj<M68K>(mf, opt_plugin))
+          if (is_gcc_lto_obj<M68K>(ctx, mf))
             return FileType::GCC_LTO_OBJ;
         } else {
-          if (is_gcc_lto_obj<SPARC64>(mf, opt_plugin))
+          if (is_gcc_lto_obj<SPARC64>(ctx, mf))
             return FileType::GCC_LTO_OBJ;
         }
         return FileType::ELF_OBJ;

common/glob.cc

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 
 #include <cstring>
 

common/hyperloglog.cc

@@ -4,7 +4,7 @@
 // For more info, read
 // https://engineering.fb.com/2018/12/13/data-infrastructure/hyperloglog
 
-#include "mold.h"
+#include "common.h"
 
 #include <cmath>
 
@@ -13,7 +13,7 @@ namespace mold {
 i64 HyperLogLog::get_cardinality() const {
   double z = 0;
   for (i64 val : buckets)
-    z += pow(2, -val);
+    z += std::ldexp(1.0, -val);
   return ALPHA * NBUCKETS * NBUCKETS / z;
 }
 

common/main.cc

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 #include "config.h"
 
 #include <cstring>
@@ -110,14 +110,26 @@ static void sighandler(int signo, siginfo_t *info, void *ucontext) {
   static std::mutex mu;
   std::scoped_lock lock{mu};
 
-  if ((signo == SIGSEGV || signo == SIGBUS) &&
-      output_buffer_start <= info->si_addr &&
-      info->si_addr < output_buffer_end) {
-    const char msg[] = "mold: failed to write to an output file. Disk full?\n";
+  switch (signo) {
+  case SIGSEGV:
+  case SIGBUS:
+    if (output_buffer_start <= info->si_addr &&
+        info->si_addr < output_buffer_end) {
+      const char msg[] = "mold: failed to write to an output file. Disk full?\n";
+      (void)!write(STDERR_FILENO, msg, sizeof(msg) - 1);
+    }
+    break;
+  case SIGABRT: {
+    const char msg[] =
+      "mold: aborted\n"
+      "mold: If mold failed due to a spurious failure of pthread_create, "
+      "it's likely because of https://github.com/oneapi-src/oneTBB/pull/824. "
+      "You should ensure that you are using 2021.9.0 or newer version of libtbb.\n";
     (void)!write(STDERR_FILENO, msg, sizeof(msg) - 1);
+    break;
+  }
   }
 
-  cleanup();
   _exit(1);
 }
 
@@ -127,6 +139,7 @@ void install_signal_handler() {
   sigemptyset(&action.sa_mask);
   action.sa_flags = SA_SIGINFO;
 
+  sigaction(SIGABRT, &action, NULL);
   sigaction(SIGINT, &action, NULL);
   sigaction(SIGTERM, &action, NULL);
   sigaction(SIGBUS, &action, NULL);
@@ -148,7 +161,7 @@ int main(int argc, char **argv) {
 
 #if MOLD_IS_SOLD
   std::string cmd = mold::filepath(argv[0]).filename().string();
-  if (cmd == "ld64" || cmd == "ld64.mold")
+  if (cmd == "ld64" || cmd.starts_with("ld64."))
     return mold::macho::main(argc, argv);
 #endif
 

common/multi-glob.cc

@@ -17,7 +17,7 @@
 // slow, but complex patterns are rare in practice, so it should be
 // OK.
 
-#include "mold.h"
+#include "common.h"
 
 #include <queue>
 #include <regex>

common/output-file-unix.h

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 
 #include <fcntl.h>
 #include <filesystem>
@@ -14,9 +14,9 @@ inline u32 get_umask() {
   return orig_umask;
 }
 
-template <typename C>
+template <typename Context>
 static std::pair<i64, char *>
-open_or_create_file(C &ctx, std::string path, i64 filesize, i64 perm) {
+open_or_create_file(Context &ctx, std::string path, i64 filesize, i64 perm) {
   std::string tmpl = filepath(path).parent_path() / ".mold-XXXXXX";
   char *path2 = (char *)save_string(ctx, tmpl).data();
 
@@ -47,11 +47,11 @@ open_or_create_file(C &ctx, std::string path, i64 filesize, i64 perm) {
   return {fd, path2};
 }
 
-template <typename C>
-class MemoryMappedOutputFile : public OutputFile<C> {
+template <typename Context>
+class MemoryMappedOutputFile : public OutputFile<Context> {
 public:
-  MemoryMappedOutputFile(C &ctx, std::string path, i64 filesize, i64 perm)
-    : OutputFile<C>(path, filesize, true) {
+  MemoryMappedOutputFile(Context &ctx, std::string path, i64 filesize, i64 perm)
+    : OutputFile<Context>(path, filesize, true) {
     i64 fd;
     std::tie(fd, output_tmpfile) = open_or_create_file(ctx, path, filesize, perm);
 
@@ -70,7 +70,7 @@ public:
       ::close(fd2);
   }
 
-  void close(C &ctx) override {
+  void close(Context &ctx) override {
     Timer t(ctx, "close_file");
 
     if (!this->is_unmapped)
@@ -92,18 +92,18 @@ private:
   int fd2 = -1;
 };
 
-template <typename C>
-class MallocOutputFile : public OutputFile<C> {
+template <typename Context>
+class MallocOutputFile : public OutputFile<Context> {
 public:
-  MallocOutputFile(C &ctx, std::string path, i64 filesize, i64 perm)
-    : OutputFile<C>(path, filesize, false), perm(perm) {
+  MallocOutputFile(Context &ctx, std::string path, i64 filesize, i64 perm)
+    : OutputFile<Context>(path, filesize, false), perm(perm) {
     this->buf = (u8 *)mmap(NULL, filesize, PROT_READ | PROT_WRITE,
                            MAP_SHARED | MAP_ANONYMOUS, -1, 0);
     if (this->buf == MAP_FAILED)
       Fatal(ctx) << "mmap failed: " << errno_string();
   }
 
-  void close(C &ctx) override {
+  void close(Context &ctx) override {
     Timer t(ctx, "close_file");
 
     if (this->path == "-") {
@@ -125,9 +125,9 @@ private:
   i64 perm;
 };
 
-template <typename C>
-std::unique_ptr<OutputFile<C>>
-OutputFile<C>::open(C &ctx, std::string path, i64 filesize, i64 perm) {
+template <typename Context>
+std::unique_ptr<OutputFile<Context>>
+OutputFile<Context>::open(Context &ctx, std::string path, i64 filesize, i64 perm) {
   Timer t(ctx, "open_file");
 
   if (path.starts_with('/') && !ctx.arg.chroot.empty())
@@ -142,11 +142,11 @@ OutputFile<C>::open(C &ctx, std::string path, i64 filesize, i64 perm) {
       is_special = true;
   }
 
-  OutputFile<C> *file;
+  OutputFile<Context> *file;
   if (is_special)
-    file = new MallocOutputFile<C>(ctx, path, filesize, perm);
+    file = new MallocOutputFile(ctx, path, filesize, perm);
   else
-    file = new MemoryMappedOutputFile<C>(ctx, path, filesize, perm);
+    file = new MemoryMappedOutputFile(ctx, path, filesize, perm);
 
 #ifdef MADV_HUGEPAGE
   // Enable transparent huge page for an output memory-mapped file.
@@ -163,7 +163,7 @@ OutputFile<C>::open(C &ctx, std::string path, i64 filesize, i64 perm) {
 
   if (ctx.arg.filler != -1)
     memset(file->buf, ctx.arg.filler, filesize);
-  return std::unique_ptr<OutputFile<C>>(file);
+  return std::unique_ptr<OutputFile>(file);
 }
 
 } // namespace mold

common/output-file-win32.h

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 
 #include <fcntl.h>
 #include <filesystem>
@@ -6,17 +6,17 @@
 
 namespace mold {
 
-template <typename C>
-class MallocOutputFile : public OutputFile<C> {
+template <typename Context>
+class MallocOutputFile : public OutputFile<Context> {
 public:
-  MallocOutputFile(C &ctx, std::string path, i64 filesize, i64 perm)
-    : OutputFile<C>(path, filesize, false), perm(perm) {
+  MallocOutputFile(Context &ctx, std::string path, i64 filesize, i64 perm)
+    : OutputFile<Context>(path, filesize, false), perm(perm) {
     this->buf = (u8 *)malloc(filesize);
     if (!this->buf)
       Fatal(ctx) << "malloc failed";
   }
 
-  void close(C &ctx) override {
+  void close(Context &ctx) override {
     Timer t(ctx, "close_file");
 
     if (this->path == "-") {
@@ -39,19 +39,19 @@ private:
   i64 perm;
 };
 
-template <typename C>
-std::unique_ptr<OutputFile<C>>
-OutputFile<C>::open(C &ctx, std::string path, i64 filesize, i64 perm) {
+template <typename Context>
+std::unique_ptr<OutputFile<Context>>
+OutputFile<Context>::open(Context &ctx, std::string path, i64 filesize, i64 perm) {
   Timer t(ctx, "open_file");
 
   if (path.starts_with('/') && !ctx.arg.chroot.empty())
     path = ctx.arg.chroot + "/" + path_clean(path);
 
-  OutputFile<C> *file = new MallocOutputFile<C>(ctx, path, filesize, perm);
+  OutputFile<Context> *file = new MallocOutputFile(ctx, path, filesize, perm);
 
   if (ctx.arg.filler != -1)
     memset(file->buf, ctx.arg.filler, filesize);
-  return std::unique_ptr<OutputFile<C>>(file);
+  return std::unique_ptr<OutputFile<Context>>(file);
 }
 
 } // namespace mold

common/perf.cc

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 
 #include <functional>
 #include <iomanip>
@@ -31,7 +31,7 @@ static i64 now_nsec() {
 #else
   struct timespec t;
   clock_gettime(CLOCK_MONOTONIC, &t);
-  return (i64)t.tv_sec * 1000000000 + t.tv_nsec;
+  return (i64)t.tv_sec * 1'000'000'000 + t.tv_nsec;
 #endif
 }
 
@@ -46,7 +46,7 @@ static std::pair<i64, i64> get_usage() {
   return {to_nsec(user), to_nsec(kernel)};
 #else
   auto to_nsec = [](struct timeval t) -> i64 {
-    return (i64)t.tv_sec * 1000000000 + t.tv_usec * 1000;
+    return (i64)t.tv_sec * 1'000'000'000 + t.tv_usec * 1'000;
   };
 
   struct rusage ru;
@@ -79,9 +79,9 @@ void TimerRecord::stop() {
 
 static void print_rec(TimerRecord &rec, i64 indent) {
   printf(" % 8.3f % 8.3f % 8.3f  %s%s\n",
-         ((double)rec.user / 1000000000),
-         ((double)rec.sys / 1000000000),
-         (((double)rec.end - rec.start) / 1000000000),
+         ((double)rec.user / 1'000'000'000),
+         ((double)rec.sys / 1'000'000'000),
+         (((double)rec.end - rec.start) / 1'000'000'000),
          std::string(indent * 2, ' ').c_str(),
          rec.name.c_str());
 

common/tar.cc

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 
 namespace mold {
 
@@ -24,7 +24,10 @@ struct UstarHeader {
     int sum = 0;
     for (i64 i = 0; i < sizeof(*this); i++)
       sum += ((u8 *)this)[i];
-    assert(sum < 01000000);
+
+    // We need to convince the compiler that sum isn't too big to silence
+    // -Werror=format-truncation.
+    ASSUME(sum < 01'000'000);
     snprintf(checksum, sizeof(checksum), "%06o", sum);
   }
 
@@ -47,6 +50,8 @@ struct UstarHeader {
   char pad[12];
 };
 
+static_assert(sizeof(UstarHeader) == 512);
+
 static std::string encode_path(std::string basedir, std::string path) {
   path = path_clean(basedir + "/" + path);
 
@@ -78,6 +83,7 @@ void TarWriter::append(std::string path, std::string_view data) {
 
   std::string attr = encode_path(basedir, path);
   snprintf(pax.size, sizeof(pax.size), "%011zo", attr.size());
+  pax.name[0] = '/';
   pax.typeflag[0] = 'x';
   pax.finalize();
   fwrite(&pax, sizeof(pax), 1, out);

common/uuid.cc

@@ -1,4 +1,4 @@
-#include "mold.h"
+#include "common.h"
 
 #include <random>
 

dist.sh

@@ -7,20 +7,27 @@
 # libstdc++ and libcrypto but dynamically-linked to libc, libm, libz
 # and librt, as they almost always exist on any Linux systems.
 
-if [ $# -ne 1 ]; then
+case $# in
+0)
+  arch=$(uname -m)
+  [[ $arch = arm* ]] && arch=arm
+  ;;
+1)
+  arch="$1"
+  ;;
+*)
   echo "Usage: $0 [ x86_64 | aarch64 | arm | ppc64le | s390x ]"
   exit 1
-fi
+esac
 
-arch=$1
-echo $arch | grep -Eq '^(x86_64|aarch64|arm|ppc64le|s390x)$' || \
+echo "$arch" | grep -Eq '^(x86_64|aarch64|arm|ppc64le|s390x)$' || \
   { echo "Error: no docker image for $arch"; exit 1; }
 
 version=$(sed -n 's/^project(mold VERSION \(.*\))/\1/p' $(dirname $0)/CMakeLists.txt)
 dest=mold-$version-$arch-linux
 set -e -x
 
-docker run --platform linux/$arch -it --rm -v "$(pwd):/mold" \
+docker run --platform linux/$arch -i --rm -v "$(pwd):/mold" \
   -e "OWNER=$(id -u):$(id -g)" rui314/mold-builder:latest \
   bash -c "mkdir /tmp/build &&
 cd /tmp/build &&

docs/mold.md

@@ -0,0 +1,802 @@
+mold(1) -- a modern linker
+==========================
+
+## SYNOPSIS
+
+`mold` [_option_...] _file_...
+
+## DESCRIPTION
+`mold` is a faster drop-in replacement for the default GNU ld(1).
+
+### How to use
+See <https://github.com/rui314/mold#how-to-use>.
+
+### Compatibility
+
+**Mold** is designed to be a drop-in replacement for the GNU linkers for
+linking user-land programs. If your user-land program cannot be built due to
+missing command-line options, please file a bug at
+<https://github.com/rui314/mold/issues>.
+
+Mold supports a very limited set of linker script features, which is just
+sufficient to read `/usr/lib/x86_64-linux-gnu/libc.so` on Linux systems (on
+Linux, that file is contrary to its name not a shared library but an ASCII
+linker script that loads a real `libc.so` file.)
+
+Beyond that, we have no plan to support any additional linker script features.
+The linker script is an ad-hoc, over-designed, complex language which we
+believe needs to be replaced by a simpler mechanism. We have a plan to add a
+replacement for the linker script to `mold` instead.
+
+### Archive symbol resolution
+
+Traditionally, Unix linkers are sensitive to the order in which input files
+appear on the command line. They process input files from the first (leftmost)
+file to the last (rightmost) file one-by-one. While reading input files, they
+maintain sets of defined and undefined symbols. When visiting an archive file
+(`.a` files), they pull out object files to resolve as many undefined symbols
+as possible and move on to the next input file. Object files that weren't
+pulled out will never have a chance for a second look.
+
+Due to this behavior, you usually have to add archive files at the end of a
+command line, so that when a linker reaches archive files, it knows what
+symbols remain as undefined.
+
+If you put archive files at the beginning of a command line, a linker doesn't
+have any undefined symbols, and thus no object files will be pulled out from
+archives. You can change the processing order by using the `--start-group` and
+`--end-group` options, though they make a linker slower.
+
+`mold`, as well as the LLVM lld(1) linker, takes a different approach. They
+remember which symbols can be resolved from archive files instead of
+forgetting them after processing each archive. Therefore, `mold` and lld(1)
+can "go back" in a command line to pull out object files from archives if they
+are needed to resolve remaining undefined symbols. They are not sensitive to
+the input file order.
+
+`--start-group` and `--end-group` are still accepted by `mold` and lld(1)
+for compatibility with traditional linkers, but they are silently ignored.
+
+### Dynamic symbol resolution
+
+Some Unix linker features are difficult to understand without comprehending
+the semantics of dynamic symbol resolution. Therefore, even though it's not
+specific to `mold`, we'll explain it here.
+
+We use "ELF module" or just "module" as a collective term to refer to an
+executable or a shared library file in the ELF format.
+
+An ELF module may have lists of imported symbols and exported symbols, as well
+as a list of shared library names from which imported symbols should be
+imported. The point is that imported symbols are not bound to any specific
+shared library until runtime.
+
+Here is how the Unix dynamic linker resolves dynamic symbols. Upon the start
+of an ELF program, the dynamic linker constructs a list of ELF modules which,
+as a whole, consist of a complete program. The executable file is always at
+the beginning of the list followed by its dependent shared libraries. An
+imported symbol is searched from the beginning of the list to the end. If two
+or more modules define the same symbol, the one that appears first in the list
+takes precedence over the others.
+
+This Unix semantics are contrary to systems such as Windows that have a
+two-level namespace for dynamic symbols. On Windows, for example, dynamic
+symbols are represented as a tuple of (`symbol-name`, `shared-library-name`),
+so that each dynamic symbol is guaranteed to be resolved from some specific
+library.
+
+Typically, an ELF module that exports a symbol also imports the same symbol.
+Such a symbol is usually resolved to itself, but that's not the case if a
+module that appears before it in the symbol search list provides another
+definition of the same symbol.
+
+Let's take `malloc` as an example. Assume that you define your version of
+`malloc` in your main executable file. Then, all `malloc` calls from any
+module are resolved to your function instead of the one in libc, because the
+executable is always at the beginning of the dynamic symbol search list. Note
+that even `malloc` calls within libc are resolved to your definition since
+libc exports and imports `malloc`. Therefore, by defining `malloc` yourself,
+you can overwrite a library function, and the `malloc` in libc becomes dead
+code.
+
+These Unix semantics are tricky and sometimes considered harmful. For example,
+assume that you accidentally define `atoi` as a global function in your
+executable that behaves completely differently from the one in the C standard.
+Then, all `atoi` function calls from any modules (even function calls within
+libc) are redirected to your function instead of the one in libc, which will
+very likely cause a problem.
+
+That is a somewhat surprising consequence for an accidental name conflict. On
+the other hand, this semantic is sometimes useful because it allows users to
+override library functions without rebuilding modules containing them.
+
+Whether good or bad, you should keep these semantics in mind to understand
+Unix linkers' behaviors.
+
+### Build reproducibility
+
+`mold`'s output is deterministic. That is, if you pass the same object files
+and the same command-line options to the same version of `mold`, it is
+guaranteed that `mold` produces the bit-by-bit identical output. The linker's
+internal randomness, such as the timing of thread scheduling or iteration
+orders of hash tables, doesn't affect the output.
+
+`mold` does not have any host-specific default settings. This is contrary to the
+GNU linkers, for which some configurable values, such as system-dependent
+library search paths, are hard-coded. `mold` depends only on its command-line
+arguments.
+
+## MOLD-SPECIFIC OPTIONS
+
+* `--chroot`=_dir_:
+  Set _dir_ as the root directory.
+
+* `--color-diagnostics`=[ _auto_ | _always_ | _never_ ]:
+  Show diagnostic messages in color using ANSI escape sequences. `auto` means
+  that `mold` prints out messages in color only if the standard output is
+  connected to a TTY. Default is `auto`.
+
+* `--color-diagnostics`:
+  Synonym for `--color-diagnostics=auto`.
+
+* `--no-color-diagnostics`:
+  Synonym for `--color-diagnostics=never`.
+
+* `--fork`, `--no-fork`:
+  Spawn a child process and let it do the actual linking. When linking a large
+  program, the OS kernel can take a few hundred milliseconds to terminate a
+  `mold` process. `--fork` hides that latency. By default, it does fork.
+
+* `--perf`:
+  Print performance statistics.
+
+* `--print-dependencies`:
+  Print out dependency information for input files.
+
+  Each line of the output for this option shows which file depends on which
+  file to use a specific symbol. This option is useful for debugging why some
+  object file in a static archive got linked or why some shared library is
+  kept in an output file's dependency list even with `--as-needed`.
+
+* `--repro`:
+  Archive input files, as well as a text file containing command line options,
+  in a tar file so that you can run `mold` with the exact same inputs again.
+  This is useful for reporting a bug with a reproducer. The output filename is
+  `path/to/output.tar`, where `path/to/output` is an output filename specified
+  by `-o`.
+
+* `--reverse-sections`:
+  Reverse the order of input sections before assigning them the offsets in the
+  output file.
+
+  This option is useful for finding bugs that depend on the initialization
+  order of global objects. In C++, constructors of global objects in a single
+  source file are guaranteed to be executed in the source order, but there's
+  no such guarantee across compilation units. Usually, constructors are
+  executed in the order given to the linker, but depending on it is a mistake.
+
+  By reversing the order of input sections using `--reverse-sections`, you can
+  easily test that your program works in the reversed initialization order.
+
+* `--run` _command_ _arg_...:
+  Run _command_ with `mold` `/usr/bin/ld`. Specifically, `mold` runs a given
+  command with the `LD_PRELOAD` environment set to intercept exec(3) family
+  functions and replaces `argv[0]` with itself if it is `ld`, `ld.gold`, or
+  `ld.lld`.
+
+* `--shuffle-sections`, `--shuffle-sections`=_number_:
+  Randomize the output by shuffling the order of input sections before
+  assigning them the offsets in the output file. If a _number_ is given, it's
+  used as a seed for the random number generator, so that the linker produces
+  the same output for the same seed. If no seed is given, a random number is
+  used as a seed.
+
+  This option is useful for benchmarking. Modern CPUs are sensitive to a
+  program's memory layout. A seemingly benign change in program layout, such
+  as a small size increase of a function in the middle of a program, can
+  affect the program's performance. Therefore, even if you write new code and
+  get a good benchmark result, it is hard to say whether the new code improves
+  the program's performance; it is possible that the new memory layout happens
+  to perform better.
+
+  By running a benchmark multiple times with randomized memory layouts using
+  `--shuffle-sections`, you can isolate your program's real performance number
+  from the randomness caused by memory layout changes.
+
+* `--stats`:
+  Print input statistics.
+
+* `--thread-count`=_count_:
+  Use _count_ number of threads.
+
+* `--threads`, `--no-threads`:
+  Use multiple threads. By default, `mold` uses as many threads as the number of
+  cores or 32, whichever is smaller. The reason it is capped at 32 is because
+  `mold` doesn't scale well beyond that point. To use only one thread, pass
+  `--no-threads` or `--thread-count=1`.
+
+* `--quick-exit`, `--no-quick-exit`:
+  Use or do not use `quick_exit` to exit.
+
+## GNU-COMPATIBLE OPTIONS
+
+* `--help`:
+  Report usage information to stdout and exit.
+
+* `-v`, `--version`:
+  Report version information to stdout.
+
+* `-V`:
+  Report version and target information to stdout.
+
+* `-E`, `--export-dynamic`, `--no-export-dynamic`:
+  When creating an executable, using the `-E` option causes all global symbols
+  to be put into the dynamic symbol table, so that the symbols are visible
+  from other ELF modules at runtime.
+
+  By default, or if `--no-export-dynamic` is given, only symbols that are
+  referenced by DSOs at link-time are exported from an executable.
+
+* `-F` _libname_, `--filter`=_libname_:
+  Set the `DT_FILTER` dynamic section field to _libname_.
+
+* `-I` _file_, `--dynamic-linker`=_file_, `--no-dynamic-linker`:
+  Set the dynamic linker path to _file_. If no `-I` option is given, or if
+  `--no-dynamic-linker` is given, no dynamic linker path is set to an output
+  file. This is contrary to the GNU linkers which set a default dynamic linker
+  path in that case. This difference doesn't usually make any difference
+  because the compiler driver always passes `-I` to the linker.
+
+* `-L` _dir_, `--library-path`=_dir_:
+  Add _dir_ to the list of library search paths from which `mold` searches
+  libraries for the `-l` option.
+
+  Unlike the GNU linkers, `mold` does not have default search paths. This
+  difference doesn't usually make any difference because the compiler driver
+  always passes all necessary search paths to the linker.
+
+* `-M`, `--print-map`:
+  Write a map file to stdout.
+
+* `-N`, `--omagic`, `--no-omagic`:
+  Force `mold` to emit an output file with an old-fashioned memory layout.
+  First, it makes the first data segment not aligned to a page boundary.
+  Second, text segments are marked as writable if the option is given.
+
+* `-S`, `--strip-debug`:
+  Omit `.debug_*` sections from the output file.
+
+* `-T` _file_, `--script`=_file_:
+  Read linker script from _file_.
+
+* `-X`, `--discard-locals`:
+  Discard temporary local symbols to reduce the sizes of the symbol table and
+  the string table. Temporary local symbols are local symbols starting with
+  `.L`. Compilers usually generate such symbols for unnamed program elements
+  such as string literals or floating-point literals.
+
+* `-e` _symbol_, `--entry`=_symbol_:
+
+  Use _symbol_ as the entry point symbol instead of the default entry
+  point symbol _start.
+
+* `-f` _shlib_, `--auxiliary`=_shlib_:
+  Set the `DT_AUXILIARY` dynamic section field to _shlib_.
+
+* `-h` _libname_, `--soname`=_libname_:
+  Set the `DT_SONAME` dynamic section field to _libname_. This option is used
+  when creating a shared object file. Typically, when you create `libfoo.so`,
+  you want to pass `--soname=foo` to a linker.
+
+* `-l` _libname_:
+  Search for `lib`_libname_`.so` or `lib`_libname_`.a` from library search
+  paths.
+
+* `-m` _target_:
+  Choose a _target_.
+
+* `-o` _file_, `--output`=_file_:
+  Use _file_ as the output file name instead of the default name `a.out`.
+
+* `-r`, `--relocatable`:
+  Instead of generating an executable or a shared object file, combine input
+  object files to generate another object file that can be used as an input to
+  a linker.
+
+* `--relocatable-merge-sections`:
+  By default, `mold` doesn't merge input sections by name when merging input
+  object files into a single output object file for `-r`. For example,
+  `.text.foo` and `.text.bar` aren't merged for `-r` even though they are
+  merged into `.text` according to the default section merging rules.
+
+  This option changes the behavior so that `mold` merges input sections by
+  name by the default section merging rules.
+
+* `-s`, `--strip-all`:
+  Omit `.symtab` section from the output file.
+
+* `-u` _symbol_, `--undefined`=_symbol_:
+  If _symbol_ remains as an undefined symbol after reading all object files,
+  and if there is a static archive that contains an object file defining
+  _symbol_, pull out the object file and link it so that the output file
+  contains a definition of _symbol_.
+
+* `--Bdynamic`:
+  Link against shared libraries.
+
+* `--Bstatic`:
+  Do not link against shared libraries.
+
+* `--Bsymbolic`:
+  When creating a shared library, make global symbols export-only (i.e. do not
+  import the same symbol). As a result, references within a shared library are
+  always resolved locally, negating symbol override at runtime. See "Dynamic
+  symbol resolution" for more information about symbol imports and exports.
+
+* `--Bsymbolic-functions`:
+  This option has the same effect as `--Bsymbolic` but works only for function
+  symbols. Data symbols remain being both imported and exported.
+
+* `--Bno-symbolic`:
+  Cancel `--Bsymbolic` and `--Bsymbolic-functions`.
+
+* `--Map`=_file_:
+  Write map file to _file_.
+
+* `--Tbss`=_address_:
+  Alias for `--section-start=.bss=`_address_.
+
+* `--Tdata`=_address_:
+  Alias for `--section-start=.data=`_address_.
+
+* `--Ttext`=_address_:
+  Alias for `--section-start=.text=`_address_.
+
+* `--allow-multiple-definition`:
+  Normally, the linker reports an error if there are more than one definition
+  of a symbol. This option changes the default behavior so that it doesn't
+  report an error for duplicate definitions and instead use the first
+  definition.
+
+* `--as-needed`, `--no-as-needed`:
+  By default, shared libraries given to the linker are unconditionally added
+  to the list of required libraries in an output file. However, shared
+  libraries after `--as-needed` are added to the list only when at least one
+  symbol is actually used by the output file. In other words, shared libraries
+  after `--as-needed` are not added to the list of needed libraries if they
+  are not needed by a program.
+
+  The `--no-as-needed` option restores the default behavior for subsequent
+  files.
+
+* `--build-id`=[ `md5` | `sha1` | `sha256` | `uuid` | `0x`_hexstring_ | `none` ]:
+  Create a `.note.gnu.build-id` section containing a byte string to uniquely
+  identify an output file. `sha256` compute a 256-bit cryptographic hash of an
+  output file and set it to build-id. `md5` and `sha1` compute the same hash
+  but truncate it to 128 and 160 bits, respectively, before setting it to
+  build-id. `uuid` sets a random 128-bit UUID. `0x`_hexstring_ sets
+  _hexstring_.
+
+* `--build-id`:
+  Synonym for `--build-id=sha256`.
+
+* `--no-build-id`:
+  Synonym for `--build-id=none`.
+
+* `--compress-debug-sections`=[ `zlib` | `zlib-gabi` | `zstd` | `none` ]:
+  Compress DWARF debug info (`.debug_*` sections) using the zlib or zstd
+  compression algorithm. `zlib-gabi` is an alias for `zlib`.
+
+* `--defsym`=_symbol_=_value_:
+  Define _symbol_ as an alias for _value_.
+
+  _value_ is either an integer (in decimal or hexadecimal with `0x` prefix) or
+  a symbol name. If an integer is given as a value, _symbol_ is defined as an
+  absolute symbol with the given value.
+
+* `--default-symver`:
+  Use soname as a symbol version and append that version to all symbols.
+
+* `--demangle`, `--no-demangle`:
+  Demangle C++ and Rust symbols in log messages.
+
+* `--dependency-file`=_file_:
+  Write a dependency file to _file_. The contents of the written file is
+  readable by make(1), which defines only one rule with the linker's output
+  file as a target and all input files as its prerequisites. Users are
+  expected to include the generated dependency file into a Makefile to
+  automate the dependency management. This option is analogous to the
+  compiler's `-MM -MF` options.
+
+* `--dynamic-list`=_file_:
+  Read a list of dynamic symbols from _file_. Same as
+  `--export-dynamic-symbol-list`, except that it implies `--Bsymbolic`. If
+  _file_ does not exist in the current directory, it is searched from library
+  search paths for the sake of compatibility with GNU ld.
+
+* `--eh-frame-hdr`, `--no-eh-frame-hdr`:
+  Create `.eh_frame_hdr` section.
+
+* `--emit-relocs`:
+  The linker usually "consumes" relocation sections. That is, the linker
+  applies relocations to other sections, and relocation sections themselves
+  are discarded.
+
+  The `--emit-relocs` instructs the linker to leave relocation sections in the
+  output file. Some post-link binary analysis or optimization tools such as
+  LLVM Bolt need them.
+
+* `--enable-new-dtags`, `--disable-new-dtags`:
+  By default, `mold` emits `DT_RUNPATH` for `--rpath`. If you pass
+  `--disable-new-dtags`, `mold` emits `DT_RPATH` for `--rpath` instead.
+
+* `--execute-only`:
+  Traditionally, most processors require both executable and readable bits to
+  1 to make the page executable, which allows machine code to be read as data
+  at runtime. This is actually what an attacker often does after gaining a
+  limited control of a process to find pieces of machine code they can use to
+  gain the full control of the process. As a mitigation, some recent
+  processors allows "execute-only" pages. If a page is execute-only, you can
+  call a function there as long as you know its address but can't read it as
+  data.
+
+  This option marks text segments execute-only. This option currently works
+  only on some ARM64 processors.
+
+* `--exclude-libs`=_libraries_ ...:
+  Mark all symbols in the given _libraries_ hidden.
+
+* `--export-dynamic-symbol`=_symbol_:
+  Put symbols matching _symbol_ in the dynamic symbol table. _symbol_ may be a
+  glob pattern in the same syntax as for the `--export-dynamic-symbol-list` or
+  `--version-script` options.
+
+* `--export-dynamic-symbol-list`=_file_:
+  Read a list of dynamic symbols from _file_.
+
+* `--fatal-warnings`, `--no-fatal-warnings`:
+  Treat warnings as errors.
+
+* `--fini`=_symbol_:
+  Call _symbol_ at unload-time.
+
+* `--gc-sections`, `--no-gc-sections`:
+  Remove unreferenced sections.
+
+* `--gdb-index`:
+  Create a `.gdb_index` section to speed up GNU debugger. To use this, you
+  need to compile source files with the `-ggnu-pubnames` compiler flag.
+
+* `--hash-style`=[ `sysv` | `gnu` | `both` | `none` ]:
+  Set hash style.
+
+* `--icf`=[ `safe` | `all` | `none` ], `--no-icf`:
+  It is not uncommon for a program to contain many identical functions that
+  differ only in name. For example, a C++ template `std::vector` is very
+  likely to be instantiated to the identical code for `std::vector<int>` and
+  `std::vector<unsigned>` because the container cares only about the size of
+  the parameter type. Identical Code Folding (ICF) is a size optimization to
+  identify and merge such identical functions.
+
+  If `--icf=all` is given, `mold` tries to merge all identical functions. This
+  reduces the size of the output most, but it is not a "safe" optimization. It
+  is guaranteed in C and C++ that two pointers pointing two different
+  functions will never be equal, but `--icf=all` breaks that assumption as two
+  identical functions have the same address after merging. So a care must be
+  taken when you use this flag that your program does not depend on the
+  function pointer uniqueness.
+
+  `--icf=safe` is a flag to merge functions only when it is safe to do so.
+  That is, if a program does not take an address of a function, it is safe to
+  merge that function with other function, as you cannot compare a function
+  pointer with something else without taking an address of a function.
+
+  `--icf=safe` needs to be used with a compiler that supports `.llvm_addrsig`
+  section which contains the information as to what symbols are address-taken.
+  LLVM/Clang supports that section by default. Since GCC does not support it
+  yet, you cannot use `--icf=safe` with GCC (it doesn't do any harm but can't
+  optimize at all.)
+
+  `--icf=none` and `--no-icf` disables ICF.
+
+* `--ignore-data-address-equality`:
+  Make ICF to merge not only functions but also data. This option should be
+  used in combination with `--icf=all`.
+
+* `--image-base`=_addr_:
+  Set the base address to _addr_.
+
+* `--init`=_symbol_:
+  Call _symbol_ at load-time.
+
+* `--no-undefined`:
+  Report undefined symbols (even with `--shared`).
+
+* `--noinhibit-exec`:
+  Create an output file even if errors occur.
+
+* `--pack-dyn-relocs`=[ `relr` | `none` ]:
+  If `relr` is specified, all `R_*_RELATIVE` relocations are put into
+  `.relr.dyn` section instead of `.rel.dyn` or `.rela.dyn` section. Since
+  `.relr.dyn` section uses a space-efficient encoding scheme, specifying this
+  flag can reduce the size of the output. This is typically most effective for
+  position-independent executable.
+
+  Note that a runtime loader has to support `.relr.dyn` to run executables or
+  shared libraries linked with `--pack-dyn-relocs=relr`. As of 2022, only
+  ChromeOS, Android and Fuchsia support it.
+
+* `--package-metadata`=_string_:
+  Embed _string_ to a `.note.package` section. This option is intended to be
+  used by a package management command such as rpm(8) to embed metadata
+  regarding a package to each executable file.
+
+* `--pie`, `--pic-executable`, `--no-pie`, `--no-pic-executable`:
+  Create a position-independent executable.
+
+* `--print-gc-sections`, `--no-print-gc-sections`:
+  Print removed unreferenced sections.
+
+* `--print-icf-sections`, `--no-print-icf-sections`:
+  Print folded identical sections.
+
+* `--push-state`, `--pop-state`:
+  `--push-state` saves the current values of `--as-needed`, `--whole-archive`,
+  `--static`, and `--start-lib`. The saved values can be restored by
+  pop-state.
+
+  `--push-state` and `--pop-state` pairs can nest.
+
+  These options are useful when you want to construct linker command line
+  options programmatically. For example, if you want to link `libfoo.so` by
+  as-needed basis but don't want to change the global state of `--as-needed`,
+  you can append `--push-state --as-needed -lfoo --pop-state` to the linker
+  command line options.
+
+* `--relax, --no-relax`:
+  Rewrite machine instructions with more efficient ones for some relocations.
+  The feature is enabled by default.
+
+* `--require-defined`=_symbol_:
+  Like `--undefined`, except the new symbol must be defined by the end of the
+  link.
+
+* `--retain-symbols-file`=_file_:
+  Keep only symbols listed in _file_. _file_ is a text file containing a
+  symbol name on each line. `mold` discards all local symbols as well as
+  global symbol that are not in _file_. Note that this option removes symbols
+  only from `.symtab` section and does not affect `.dynsym` section, which is
+  used for dynamic linking.
+
+* `--rpath`=_dir_:
+  Add _dir_ to runtime search paths.
+
+* `--section-start`=_section_=_address_:
+  Set _address_ to section. _address_ is a hexadecimal number that may start
+  with an optional `0x`.
+
+* `--shared`, `--Bshareable`:
+  Create a share library.
+
+* `--spare-dynamic-tags`=_number_:
+  Reserve the given _number_ of tags in `.dynamic` section.
+
+* `--start-lib`, `--end-lib`:
+  Handle object files between `--start-lib` and `--end-lib` as if they were in
+  an archive file. That means object files between them are linked only when
+  they are needed to resolve undefined symbols. The options are useful if you
+  want to link object files only when they are needed but want to avoid the
+  overhead of running ar(3).
+
+* `--static`:
+  Do not link against shared libraries.
+
+* `--sysroot`=_dir_:
+  Set target system root directory to _dir_.
+
+* `--trace`:
+  Print name of each input file.
+
+* `--undefined-version`, `--no-undefined-version`:
+  By default, `mold` warns on a symbol specified by a version script or by
+  `--export-dynamic-symbol` if it is not defined. You can silence the warning
+  by `--undefined-version`.
+
+* `--unique`=_pattern_:
+  Don't merge input sections that match the given glob pattern _pattern_.
+
+* `--unresolved-symbols`=[ `report-all` | `ignore-all` | `ignore-in-object-files` | `ignore-in-shared-libs` ]:
+  How to handle undefined symbols.
+
+* `--version-script`=_file_:
+  Read version script from _file_. If _file_ does not exist in the current
+  directory, it is searched from library search paths for the sake of
+  compatibility with GNU ld.
+
+* `--warn-common`, `--no-warn-common`:
+  Warn about common symbols.
+
+* `--warn-once`:
+  Only warn once for each undefined symbol instead of warn for each relocation
+  referring an undefined symbol.
+
+* `--warn-unresolved-symbols`, `--error-unresolved-symbols`:
+  Normally, the linker reports an error for unresolved symbols.
+  `--warn-unresolved-symbols` option turns it into a warning.
+  `--error-unresolved-symbols` option restores the default behavior.
+
+* `--whole-archive`, `--no-whole-archive`:
+  When archive files (`.a` files) are given to the linker, only object files
+  that are needed to resolve undefined symbols are extracted from them and
+  linked to an output file. `--whole-archive` changes that behavior for
+  subsequent archives so that the linker extracts all object files and links
+  them to an output. For example, if you are creating a shared object file and
+  you want to include all archive members to the output, you should pass
+  `--whole-archive`. `--no-whole-archive` restores the default behavior for
+  subsequent archives.
+
+* `--wrap`=_symbol_:
+  Make _symbol_ be resolved to `__wrap_`_symbol_. The original symbol can be
+  resolved as `__real_`_symbol_. This option is typically used for wrapping an
+  existing function.
+
+* `-z cet-report`=[ `warning` | `error` | `none` ]:
+  Intel Control-flow Enforcement Technology (CET) is a new x86 feature
+  available since Tiger Lake which is released in 2020. It defines new
+  instructions to harden security to protect programs from control hijacking
+  attacks. You can tell the compiler to use the feature by specifying the
+  `-fcf-protection` flag.
+
+  `-z cet-report` flag is used to make sure that all object files were
+  compiled with a correct `-fcf-protection` flag. If `warning` or `error` are
+  given, `mold` prints out a warning or an error message if an object file was
+  not compiled with the compiler flag.
+
+  `mold` looks for `GNU_PROPERTY_X86_FEATURE_1_IBT` bit and
+  `GNU_PROPERTY_X86_FEATURE_1_SHSTK` bit in `.note.gnu.property` section to
+  determine whether or not an object file was compiled with `-fcf-protection`.
+
+* `-z now`, `-z lazy`:
+  By default, functions referring to other ELF modules are resolved by the
+  dynamic linker when they are called for the first time. `-z now` marks an
+  executable or a shared library file so that all dynamic symbols are resolved
+  when a file is loaded to memory. `-z lazy` restores the default behavior.
+
+* `-z origin`:
+  Mark object requiring immediate `$ORIGIN` processing at runtime.
+
+* `-z ibt`:
+  Turn on `GNU_PROPERTY_X86_FEATURE_1_IBT` bit in `.note.gnu.property` section
+  to indicate that the output uses IBT-enabled PLT. This option implies `-z
+  ibtplt`.
+
+* `-z ibtplt`:
+  Generate Intel Branch Tracking (IBT)-enabled PLT which is the default on
+  x86-64. This is the default.
+
+* `-z execstack`, `-z noexecstack`:
+  By default, the pages for the stack area (i.e. the pages where local
+  variables reside) are not executable for security reasons. `-z execstack`
+  makes it executable. `-z noexecstack` restores the default behavior.
+
+* `-z keep-text-section-prefix`, `-z nokeep-text-section-prefix`:
+  Keep `.text.hot`, `.text.unknown`, `.text.unlikely`, `.text.startup`, and
+  `.text.exit` as separate sections in the final binary instead of merging
+  them as `.text`.
+
+* `-z relro`, `-z norelro`:
+  Some sections such as `.dynamic` have to be writable only during an
+  executable or a shared library file is being loaded to memory. Once the
+  dynamic linker finishes its job, such sections won't be mutated by anyone.
+  As a security mitigation, it is preferred to make such segments read-only
+  during program execution.
+
+  `-z relro` puts such sections into a special segment called `relro`. The
+  dynamic linker makes a relro segment read-only after it finishes its job.
+
+  By default, `mold` generates a relro segment. `-z norelro` disables the
+  feature.
+
+* `-z sectionheader`, `-z nosectionheader`:
+  `-z nosectionheader` tell the linker to omit the section header.
+  By default, the linker does not omit the section header.
+
+* `-z separate-loadable-segments`, `-z separate-code`, `-z noseparate-code`:
+  If one memory page contains multiple segments, the page protection bits are
+  set in such a way that the needed attributes (writable or executable) are
+  satisfied for all segments. This usually happens at a boundary of two
+  segments with two different attributes.
+
+  `separate-loadable-segments` adds paddings between segments with different
+  attributes so that they do not share the same page. This is the default.
+
+  `separate-code` adds paddings only between executable and non-executable
+  segments.
+
+  `noseparate-code` does not add any paddings between segments.
+
+* `-z defs`, `-z nodefs`:
+  Report undefined symbols (even with `--shared`).
+
+* `-z shstk`:
+  Enforce shadow stack by turning `GNU_PROPERTY_X86_FEATURE_1_SHSTK` bit in
+  `.note.gnu.property` output section. Shadow stack is part of Intel
+  Control-flow Enforcement Technology (CET), which is available since Tiger
+  Lake (2020).
+
+* `-z text`, `-z notext`, `-z textoff`:
+  `mold` by default reports an error if dynamic relocations are created in
+  read-only sections. If `-z notext` or `-z textoff` are given, `mold` creates
+  such dynamic relocations without reporting an error. `-z text` restores the
+  default behavior.
+
+* `-z max-page-size`=_number_:
+  Some CPU ISAs support multiple different memory page sizes. This option
+  specifies the maximum page size that an output binary can run on. The
+  default value is 4 KiB for i386, x86-64, and RISC-V, and 64 KiB for ARM64.
+
+* `-z nodefaultlib`:
+  Make the dynamic loader ignore default search paths.
+
+* `-z nodelete`:
+  Mark DSO non-deletable at runtime.
+
+* `-z nodlopen`:
+  Mark DSO not available to dlopen(3). This option makes it possible for the
+  linker to optimize thread-local variable accesses by rewriting instructions
+  for some targets.
+
+* `-z nodump`:
+  Mark DSO not available to dldump(3).
+
+* `-z nocopyreloc`:
+  Do not create copy relocations.
+
+* `-z initfirst`:
+  Mark DSO to be initialized first at runtime.
+
+* `-z interpose`:
+  Mark object to interpose all DSOs but executable.
+
+* `-(`, `-)`, `-EL`, `-O`_number_, `--allow-shlib-undefined`, `--dc`, `--dp`, `--end-group`, `--no-add-needed`, `--no-allow-shlib-undefined`, `--no-copy-dt-needed-entries`, `--no-fatal-warnings`, `--nostdlib`, `--rpath-link=Ar dir`, `--sort-common`, `--sort-section`, `--start-group`, `--warn-constructors`, `--warn-once`, `--fix-cortex-a53-835769`, `--fix-cortex-a53-843419`, `-z combreloc`, `-z common-page-size`, `-z nocombreloc`:
+  Ignored
+
+## ENVIRONMENT VARIABLES
+
+* `MOLD_JOBS`:
+  If this variable is set to `1`, only one process of `mold` runs actively. A
+  mold process invoked while another active mold process is running will wait
+  before doing anything until the active process exits.
+
+  The purpose of this environment variable is to reduce peak memory usage.
+  Since mold is highly parallelized, there's no point in running it
+  simultaneously. If you run N instances of mold in parallel, it would take N
+  times more time and N times more memory. If you run them serially, it would
+  still take N times more to finish, but their peak memory usage is reduced to
+  normal.
+
+  If your build system tends to invoke multiple linker processes
+  simultaneously, you may want to try to set this environment variable to
+  `1` to see if it could improve overall performance.
+
+  Currently, any value other than 1 is silently ignored.
+
+* `MOLD_DEBUG`:
+  If this variable is set to a non-empty string, `mold` embeds its
+  command-line options in the output file's `.comment` section.
+
+* `MOLD_REPRO`:
+  Setting this variable to a non-empty string has the same effect as passing
+  the `--repro` option.
+
+## SEE ALSO
+
+gold(1), ld(1), elf(5), ld.so(8)
+
+## AUTHOR
+
+Rui Ueyama <ruiu@cs.stanford.edu>
+
+## BUGS
+
+Report bugs to <https://github.com/rui314/mold/issues>.

elf/arch-alpha.cc

@@ -62,8 +62,8 @@ template <>
 void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {}
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -103,7 +103,7 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
 
     switch (rel.r_type) {
     case R_ALPHA_REFQUAD:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
     case R_ALPHA_GPREL32:
       *(ul32 *)loc = S + A - GP;
@@ -166,17 +166,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     SectionFragment<E> *frag;
     i64 frag_addend;
     std::tie(frag, frag_addend) = get_fragment(ctx, rel);
@@ -213,16 +208,11 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
       Error(ctx) << sym << ": GNU ifunc symbol is not supported on Alpha";
 
@@ -231,29 +221,30 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
       scan_dyn_absrel(ctx, sym, rel);
       break;
     case R_ALPHA_LITERAL:
-      if (rel.r_addend) {
-        sym.flags.fetch_or(NEEDS_ALPHA_GOT, std::memory_order_relaxed);
+      if (rel.r_addend)
         ctx.extra.got->add_symbol(sym, rel.r_addend);
-      } else {
-        sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
-      }
+      else
+        sym.flags |= NEEDS_GOT;
       break;
     case R_ALPHA_SREL32:
       scan_pcrel(ctx, sym, rel);
       break;
     case R_ALPHA_BRSGP:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_ALPHA_TLSGD:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_TLSGD;
       break;
     case R_ALPHA_TLSLDM:
-      ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+      ctx.needs_tlsld = true;
       break;
     case R_ALPHA_GOTTPREL:
+      sym.flags |= NEEDS_GOTTP;
+      break;
     case R_ALPHA_TPRELHI:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+    case R_ALPHA_TPRELLO:
+      check_tlsle(ctx, sym, rel);
       break;
     case R_ALPHA_GPREL32:
     case R_ALPHA_LITUSE:
@@ -263,7 +254,6 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_ALPHA_GPRELLOW:
     case R_ALPHA_DTPRELHI:
     case R_ALPHA_DTPRELLO:
-    case R_ALPHA_TPRELLO:
       break;
     default:
       Fatal(ctx) << *this << ": unknown relocation: " << rel;
@@ -273,16 +263,15 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
 // An R_ALPHA_LITERAL relocation may request the linker to create a GOT
 // entry for an external symbol with a non-zero addend. This is an unusual
-// request which we cannot find in any other targets.
+// request which is not found in any other targets.
 //
-// Referring an external symbol with a non-zero addend is a bad practice.
-// To see why, assume that we have two R_ALPHA_LITERAL relocs to symbol X
-// with addend 8 and 16. To satisfy that request, we need to create not
-// only one but two GOT entries for symbol X with different addends.
+// Referring an external symbol with a non-zero addend is a bad practice
+// because we need to create as many dynamic relocations as the number of
+// distinctive addends for the same symbol.
 //
 // We don't want to mess up the implementation of the common GOT section
 // for Alpha. So we create another GOT-like section, .alpha_got. Any GOT
-// entry for an R_ALPHA_LITERAL reloc with a non-zero addends is created
+// entry for an R_ALPHA_LITERAL reloc with a non-zero addend is created
 // not in .got but in .alpha_got.
 //
 // Since .alpha_got entries are accessed relative to GP, .alpha_got
@@ -304,9 +293,9 @@ u64 AlphaGotSection::get_addr(Symbol<E> &sym, i64 addend) {
   return this->shdr.sh_addr + (it - entries.begin()) * sizeof(Word<E>);
 }
 
-i64 AlphaGotSection::get_reldyn_size(Context<E> &ctx) {
+i64 AlphaGotSection::get_reldyn_size(Context<E> &ctx) const {
   i64 n = 0;
-  for (Entry &e : entries)
+  for (const Entry &e : entries)
     if (e.sym->is_imported || (ctx.arg.pic && !e.sym->is_absolute()))
       n++;
   return n;
@@ -319,9 +308,8 @@ void AlphaGotSection::finalize() {
 }
 
 void AlphaGotSection::copy_buf(Context<E> &ctx) {
-  ElfRel<E> *dynrel = nullptr;
-  if (ctx.reldyn)
-    dynrel = (ElfRel<E> *)(ctx.buf + ctx.reldyn->shdr.sh_offset + reldyn_offset);
+  ElfRel<E> *dynrel = (ElfRel<E> *)(ctx.buf + ctx.reldyn->shdr.sh_offset +
+                                    reldyn_offset);
 
   for (i64 i = 0; i < entries.size(); i++) {
     Entry &e = entries[i];
@@ -330,7 +318,7 @@ void AlphaGotSection::copy_buf(Context<E> &ctx) {
 
     if (e.sym->is_imported) {
       *buf = ctx.arg.apply_dynamic_relocs ? e.addend : 0;
-      *dynrel++ = ElfRel<E>(P, E::R_ABS, e.sym->get_dynsym_idx(ctx), e.addend);
+      *dynrel++ = ElfRel<E>(P, E::R_DYNAMIC, e.sym->get_dynsym_idx(ctx), e.addend);
     } else {
       *buf = e.sym->get_addr(ctx) + e.addend;
       if (ctx.arg.pic && !e.sym->is_absolute())

elf/arch-arm32.cc

@@ -2,7 +2,7 @@
 // processors support two different instruction encodings: Thumb and
 // ARM (in a narrower sense). Thumb instructions are either 16 bits or
 // 32 bits, while ARM instructions are all 32 bits. Feature-wise,
-// thumb is a subset of ARM, so not all ARM instructions are
+// Thumb is a subset of ARM, so not all ARM instructions are
 // representable in Thumb.
 //
 // ARM processors originally supported only ARM instructions. Thumb
@@ -12,14 +12,16 @@
 // be switched using BX (branch and mode exchange)-family instructions.
 // We need to use that instructions to, for example, call a function
 // encoded in Thumb from a function encoded in ARM. Sometimes, the
-// linker even has to emit an interworking thunk code to switch mode.
+// linker even has to emit interworking thunk code to switch mode.
 //
 // ARM instructions are aligned to 4 byte boundaries. Thumb are to 2
-// byte boundaries.
+// byte boundaries. So the least significant bit of a function address
+// is always 0.
 //
-// You can distinguish Thumb functions from ARM functions by looking
-// at the least significant bit (LSB) of its "address". If LSB is 0,
-// it's ARM; otherwise, Thumb.
+// To distinguish Thumb functions from ARM fucntions, the LSB of a
+// function address is repurposed as a boolean flag. If the LSB is 0,
+// the function referred to by the address is encoded in ARM;
+// otherwise, Thumb.
 //
 // For example, if a symbol `foo` is of type STT_FUNC and has value
 // 0x2001, `foo` is a function using Thumb instructions whose address
@@ -42,8 +44,6 @@ using E = ARM32;
 template <>
 i64 get_addend(u8 *loc, const ElfRel<E> &rel) {
   switch (rel.r_type) {
-  case R_ARM_NONE:
-    return 0;
   case R_ARM_ABS32:
   case R_ARM_REL32:
   case R_ARM_TARGET1:
@@ -76,8 +76,9 @@ i64 get_addend(u8 *loc, const ElfRel<E> &rel) {
   }
   case R_ARM_CALL:
   case R_ARM_JUMP24:
+  case R_ARM_PLT32:
   case R_ARM_TLS_CALL:
-    return sign_extend(*(ul32 *)loc & 0x00ff'ffff, 23) << 2;
+    return sign_extend(*(ul32 *)loc, 23) << 2;
   case R_ARM_MOVW_PREL_NC:
   case R_ARM_MOVW_ABS_NC:
   case R_ARM_MOVT_PREL:
@@ -100,7 +101,7 @@ i64 get_addend(u8 *loc, const ElfRel<E> &rel) {
     return sign_extend(val, 15);
   }
   default:
-    unreachable();
+    return 0;
   }
 }
 
@@ -167,8 +168,8 @@ void write_addend(u8 *loc, i64 val, const ElfRel<E> &rel) {
     write_thm_b_imm(loc, val);
     break;
   case R_ARM_CALL:
-  case R_ARM_PLT32:
   case R_ARM_JUMP24:
+  case R_ARM_PLT32:
     *(ul32 *)loc = (*(ul32 *)loc & 0xff00'0000) | bits(val, 25, 2);
     break;
   case R_ARM_MOVW_PREL_NC:
@@ -194,14 +195,14 @@ void write_addend(u8 *loc, i64 val, const ElfRel<E> &rel) {
 template <>
 void write_plt_header(Context<E> &ctx, u8 *buf) {
   static const ul32 insn[] = {
-    0xe52d'e004, // push {lr}
-    0xe59f'e004, // ldr lr, 2f
+    0xe52d'e004, //    push {lr}
+    0xe59f'e004, //    ldr lr, 2f
     0xe08f'e00e, // 1: add lr, pc, lr
-    0xe5be'f008, // ldr pc, [lr, #8]!
+    0xe5be'f008, //    ldr pc, [lr, #8]!
     0x0000'0000, // 2: .word .got.plt - 1b - 8
-    0xe320'f000, // nop
-    0xe320'f000, // nop
-    0xe320'f000, // nop
+    0xe320'f000, //    nop
+    0xe320'f000, //    nop
+    0xe320'f000, //    nop
   };
 
   memcpy(buf, insn, sizeof(insn));
@@ -210,8 +211,8 @@ void write_plt_header(Context<E> &ctx, u8 *buf) {
 
 static const ul32 plt_entry[] = {
   0xe59f'c004, // 1: ldr ip, 2f
-  0xe08c'c00f, // add ip, ip, pc
-  0xe59c'f000, // ldr pc, [ip]
+  0xe08c'c00f, //    add ip, ip, pc
+  0xe59c'f000, //    ldr pc, [ip]
   0x0000'0000, // 2: .word sym@GOT - 1b
 };
 
@@ -230,8 +231,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 // ARM does not use .eh_frame for exception handling. Instead, it uses
 // .ARM.exidx and .ARM.extab. So this function is empty.
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {}
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {}
 
 // ARM and Thumb branch instructions can jump within ±16 MiB.
 static bool is_jump_reachable(i64 val) {
@@ -247,17 +248,14 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     dynrel = (ElfRel<E> *)(ctx.buf + ctx.reldyn->shdr.sh_offset +
                            file.reldyn_offset + this->reldyn_offset);
 
-  std::span<std::unique_ptr<RangeExtensionThunk<E>>> thunks =
-    output_section->thunks;
-
-  auto get_trampoline_addr = [&](u64 addr) {
-    for (;;) {
-      assert(!thunks.empty());
-      i64 disp = output_section->shdr.sh_addr + thunks[0]->offset - addr;
+  auto get_tls_trampoline_addr = [&, i = 0](u64 addr) mutable {
+    for (; i < output_section->thunks.size(); i++) {
+      i64 disp = output_section->shdr.sh_addr + output_section->thunks[i]->offset -
+                 addr;
       if (is_jump_reachable(disp))
         return disp;
-      thunks = thunks.subspan(1);
     }
+    unreachable();
   };
 
   for (i64 i = 0; i < rels.size(); i++) {
@@ -288,15 +286,12 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     switch (rel.r_type) {
     case R_ARM_ABS32:
     case R_ARM_TARGET1:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
     case R_ARM_REL32:
       *(ul32 *)loc = S + A - P;
       break;
     case R_ARM_THM_CALL: {
-      // THM_CALL relocation refers either BL or BLX instruction.
-      // They are different in only one bit. We need to use BL if
-      // the jump target is Thumb. Otherwise, use BLX.
       if (sym.is_remaining_undef_weak()) {
         // On ARM, calling an weak undefined symbol jumps to the
         // next instruction.
@@ -304,6 +299,9 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
         break;
       }
 
+      // THM_CALL relocation refers either BL or BLX instruction.
+      // They are different in only one bit. We need to use BL if
+      // the jump target is Thumb. Otherwise, use BLX.
       i64 val = S + A - P;
       if (is_jump_reachable(val)) {
         if (T) {
@@ -332,54 +330,66 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_ARM_GOT_BREL:
       *(ul32 *)loc = G + A;
       break;
-    case R_ARM_CALL:
-    case R_ARM_PLT32: {
-      // Just like THM_CALL, ARM_CALL relocation refers either BL or
-      // BLX instruction. We may need to rewrite BL → BLX or BLX → BL.
-      bool is_bl = ((*(ul32 *)loc & 0xff00'0000) == 0xeb00'0000);
-      bool is_blx = ((*(ul32 *)loc & 0xfe00'0000) == 0xfa00'0000);
-      if (!is_bl && !is_blx)
-        Fatal(ctx) << *this << ": R_ARM_CALL refers neither BL nor BLX";
-
+    case R_ARM_CALL: {
       if (sym.is_remaining_undef_weak()) {
-        // On ARM, calling an weak undefined symbol jumps to the
-        // next instruction.
         *(ul32 *)loc = 0xe320'f000; // NOP
         break;
       }
 
+      // Just like THM_CALL, ARM_CALL relocation refers to either BL or
+      // BLX instruction. We may need to rewrite BL → BLX or BLX → BL.
+      bool is_bl = ((*(ul32 *)loc & 0xff00'0000) == 0xeb00'0000);
+      bool is_blx = ((*(ul32 *)loc & 0xfe00'0000) == 0xfa00'0000);
+      if (!is_bl && !is_blx)
+        Fatal(ctx) << *this << ": R_ARM_CALL refers to neither BL nor BLX";
+
       u64 val = S + A - P;
       if (is_jump_reachable(val)) {
-        if (T)
-          *(ul32 *)loc = 0xfa00'0000 | (bit(val, 1) << 24) | bits(val, 25, 2);
-        else
-          *(ul32 *)loc = 0xeb00'0000 | bits(val, 25, 2);
+        if (T) {
+          *(ul32 *)loc = 0xfa00'0000; // BLX
+          *(ul32 *)loc |= (bit(val, 1) << 24) | bits(val, 25, 2);
+        } else {
+          *(ul32 *)loc = 0xeb00'0000; // BL
+          *(ul32 *)loc |= bits(val, 25, 2);
+        }
       } else {
-        *(ul32 *)loc = 0xeb00'0000 | bits(get_arm_thunk_addr() + A - P, 25, 2);
+        *(ul32 *)loc = 0xeb00'0000; // BL
+        *(ul32 *)loc |= bits(get_arm_thunk_addr() + A - P, 25, 2);
       }
       break;
     }
-    case R_ARM_JUMP24:
+    case R_ARM_JUMP24: {
+      if (sym.is_remaining_undef_weak()) {
+        *(ul32 *)loc = 0xe320'f000; // NOP
+        break;
+      }
+
+      // These relocs refers to a B (unconditional branch) instruction.
+      // Unlike BL or BLX, we can't rewrite B to BX in place when the
+      // processor mode switch is required because BX doesn't takes an
+      // immediate; it takes only a register. So if mode switch is
+      // required, we jump to a linker-synthesized thunk which does the
+      // job with a longer code sequence.
+      u64 val = S + A - P;
+      if (!is_jump_reachable(val) || T)
+        val = get_arm_thunk_addr() + A - P;
+      *(ul32 *)loc = (*(ul32 *)loc & 0xff00'0000) | bits(val, 25, 2);
+      break;
+    }
+    case R_ARM_PLT32:
       if (sym.is_remaining_undef_weak()) {
         *(ul32 *)loc = 0xe320'f000; // NOP
       } else {
-        // Unlike BL and BLX, we can't rewrite B to BX because BX doesn't
-        // takes an immediate; it takes only a register. So if mode switch
-        // is required, we jump to a linker-synthesized thunk which constructs
-        // a branch destination in a register and branches to that address.
-        u64 val = S + A - P;
-        if (!is_jump_reachable(val) || T)
-          val = get_arm_thunk_addr() + A - P;
+        u64 val = (T ? get_arm_thunk_addr() : S) + A - P;
         *(ul32 *)loc = (*(ul32 *)loc & 0xff00'0000) | bits(val, 25, 2);
       }
       break;
-    case R_ARM_THM_JUMP11: {
+    case R_ARM_THM_JUMP11:
       assert(T);
-      i64 val = S + A - P;
-      check(val, -(1 << 11), 1 << 11);
-      *(ul16 *)loc = (*(ul16 *)loc & 0xf800) | bits(val, 11, 1);
+      check(S + A - P, -(1 << 11), 1 << 11);
+      *(ul16 *)loc &= 0xf800;
+      *(ul16 *)loc |= bits(S + A - P, 11, 1);
       break;
-    }
     case R_ARM_THM_JUMP19: {
       i64 val = S + A - P;
       check(val, -(1 << 19), 1 << 19);
@@ -398,18 +408,20 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ul16 *)(loc + 2) |= (J2 << 13) | (J1 << 11) | imm11;
       break;
     }
-    case R_ARM_THM_JUMP24:
+    case R_ARM_THM_JUMP24: {
       if (sym.is_remaining_undef_weak()) {
-        *(ul32 *)loc = 0x8000'f3af; // NOP.W
-      } else {
-        // Just like R_ARM_JUMP24, we need to jump to a thunk if we need to
-        // switch processor mode.
-        u64 val = S + A - P;
-        if (!is_jump_reachable(val) || !T)
-          val = get_thumb_thunk_addr() + A - P;
-        write_thm_b_imm(loc, val);
+        *(ul32 *)loc = 0x8000'f3af; // NOP
+        break;
       }
+
+      // Just like R_ARM_JUMP24, we need to jump to a thunk if we need to
+      // switch processor mode.
+      u64 val = S + A - P;
+      if (!is_jump_reachable(val) || !T)
+        val = get_thumb_thunk_addr() + A - P;
+      write_thm_b_imm(loc, val);
       break;
+    }
     case R_ARM_MOVW_PREL_NC:
       write_mov_imm(loc, ((S + A) | T) - P);
       break;
@@ -419,12 +431,11 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_ARM_THM_MOVW_PREL_NC:
       write_thm_mov_imm(loc, ((S + A) | T) - P);
       break;
-    case R_ARM_PREL31: {
-      i64 val = S + A - P;
-      check(val, -(1LL << 30), 1LL << 30);
-      *(ul32 *)loc = (*(ul32 *)loc & 0x8000'0000) | (val & 0x7fff'ffff);
+    case R_ARM_PREL31:
+      check(S + A - P, -(1LL << 30), 1LL << 30);
+      *(ul32 *)loc &= 0x8000'0000;
+      *(ul32 *)loc |= (S + A - P) & 0x7fff'ffff;
       break;
-    }
     case R_ARM_THM_MOVW_ABS_NC:
       write_thm_mov_imm(loc, (S + A) | T);
       break;
@@ -456,12 +467,33 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ul32 *)loc = S + A - ctx.tp_addr;
       break;
     case R_ARM_TLS_GOTDESC:
+      // ARM32 TLSDESC uses the following code sequence to materialize
+      // a TP-relative address in r0.
+      //
+      //       ldr     r0, .L2
+      //  .L1: bl      foo
+      //           R_ARM_TLS_CALL
+      //  .L2: .word   foo + . - .L1
+      //           R_ARM_TLS_GOTDESC
+      //
+      // We may relax the instructions to the following for non-dlopen'd DSO
+      //
+      //       ldr     r0, .L2
+      //  .L1: ldr r0, [pc, r0]
+      //       ...
+      //  .L2: .word   foo(gottpoff) + . - .L1
+      //
+      // or to the following for executable.
+      //
+      //       ldr     r0, .L2
+      //  .L1: nop
+      //       ...
+      //  .L2: .word   foo(tpoff)
       if (sym.has_tlsdesc(ctx)) {
         // A is odd if the corresponding TLS_CALL is Thumb.
-        if (A & 1)
-          *(ul32 *)loc = sym.get_tlsdesc_addr(ctx) - P + A - 6;
-        else
-          *(ul32 *)loc = sym.get_tlsdesc_addr(ctx) - P + A - 4;
+        *(ul32 *)loc = sym.get_tlsdesc_addr(ctx) - P + A - ((A & 1) ? 6 : 4);
+      } else if (sym.has_gottp(ctx)) {
+        *(ul32 *)loc = sym.get_gottp_addr(ctx) - P + A - ((A & 1) ? 5 : 8);
       } else {
         *(ul32 *)loc = S - ctx.tp_addr;
       }
@@ -469,20 +501,25 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_ARM_TLS_CALL:
       if (sym.has_tlsdesc(ctx)) {
         // BL <tls_trampoline>
-        *(ul32 *)loc = 0xeb00'0000 | bits(get_trampoline_addr(P + 8), 25, 2);
+        *(ul32 *)loc = 0xeb00'0000 | bits(get_tls_trampoline_addr(P + 8), 25, 2);
+      } else if (sym.has_gottp(ctx)) {
+        *(ul32 *)loc = 0xe79f'0000; // ldr r0, [pc, r0]
       } else {
-        // BL -> NOP
-        *(ul32 *)loc = 0xe320'f000;
+        *(ul32 *)loc = 0xe320'f000; // nop
       }
       break;
     case R_ARM_THM_TLS_CALL:
       if (sym.has_tlsdesc(ctx)) {
-        u64 val = align_to(get_trampoline_addr(P + 4), 4);
+        u64 val = align_to(get_tls_trampoline_addr(P + 4), 4);
         write_thm_b_imm(loc, val);
         *(ul16 *)(loc + 2) &= ~0x1000; // rewrite BL with BLX
+      } else if (sym.has_gottp(ctx)) {
+        // Since `ldr r0, [pc, r0]` is not representable in Thumb,
+        // we use two instructions instead.
+        *(ul16 *)loc = 0x4478;         // add r0, pc
+        *(ul16 *)(loc + 2) = 0x6800;   // ldr r0, [r0]
       } else {
-        // BL -> NOP.W
-        *(ul32 *)loc = 0x8000'f3af;
+        *(ul32 *)loc = 0x8000'f3af;    // nop.w
       }
       break;
     default:
@@ -497,17 +534,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     SectionFragment<E> *frag;
     i64 frag_addend;
     std::tie(frag, frag_addend) = get_fragment(ctx, rel);
@@ -546,18 +578,13 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
 
     switch (rel.r_type) {
     case R_ARM_ABS32:
@@ -568,16 +595,16 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
       break;
     case R_ARM_THM_CALL:
     case R_ARM_CALL:
-    case R_ARM_PLT32:
     case R_ARM_JUMP24:
+    case R_ARM_PLT32:
     case R_ARM_THM_JUMP24:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_ARM_GOT_PREL:
     case R_ARM_GOT_BREL:
     case R_ARM_TARGET2:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
     case R_ARM_MOVT_PREL:
     case R_ARM_THM_MOVT_PREL:
@@ -585,17 +612,20 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
       scan_pcrel(ctx, sym, rel);
       break;
     case R_ARM_TLS_GD32:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_TLSGD;
       break;
     case R_ARM_TLS_LDM32:
-      ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+      ctx.needs_tlsld = true;
       break;
     case R_ARM_TLS_IE32:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
       break;
-    case R_ARM_TLS_GOTDESC:
-      if (!relax_tlsdesc(ctx, sym))
-        sym.flags.fetch_or(NEEDS_TLSDESC, std::memory_order_relaxed);
+    case R_ARM_TLS_CALL:
+    case R_ARM_THM_TLS_CALL:
+      scan_tlsdesc(ctx, sym);
+      break;
+    case R_ARM_TLS_LE32:
+      check_tlsle(ctx, sym, rel);
       break;
     case R_ARM_REL32:
     case R_ARM_BASE_PREL:
@@ -607,10 +637,8 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_ARM_THM_MOVW_PREL_NC:
     case R_ARM_THM_MOVW_ABS_NC:
     case R_ARM_TLS_LDO32:
-    case R_ARM_TLS_LE32:
-    case R_ARM_TLS_CALL:
-    case R_ARM_THM_TLS_CALL:
     case R_ARM_V4BX:
+    case R_ARM_TLS_GOTDESC:
       break;
     default:
       Error(ctx) << *this << ": unknown relocation: " << rel;
@@ -620,8 +648,6 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
 template <>
 void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
-  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
-
   // TLS trampoline code. ARM32's TLSDESC is designed so that this
   // common piece of code is factored out from object files to reduce
   // output size. Since no one provide, the linker has to synthesize it.
@@ -629,36 +655,44 @@ void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
     0xe08e'0000, // add r0, lr, r0
     0xe590'1004, // ldr r1, [r0, #4]
     0xe12f'ff11, // bx  r1
+    0xe320'f000, // nop
   };
 
   // This is a range extension and mode switch thunk.
   // It has two entry points: +0 for Thumb and +4 for ARM.
   const u8 entry[] = {
     // .thumb
-    0xfc, 0x46,             //    mov  ip, pc
-    0x60, 0x47,             //    bx   ip  # jumps to the following `ldr` insn
+    0x78, 0x47,             //    bx   pc  # jumps to 1f
+    0xc0, 0x46,             //    nop
     // .arm
-    0x04, 0xc0, 0x9f, 0xe5, //    ldr  ip, 2f
-    0x0f, 0xc0, 0x8c, 0xe0, // 1: add  ip, ip, pc
-    0x1c, 0xff, 0x2f, 0xe1, //    bx   ip
-    0x00, 0x00, 0x00, 0x00, // 2: .word sym - 1b
+    0x00, 0xc0, 0x9f, 0xe5, // 1: ldr  ip, 3f
+    0x0f, 0xf0, 0x8c, 0xe0, // 2: add  pc, ip, pc
+    0x00, 0x00, 0x00, 0x00, // 3: .word sym - 2b
   };
 
   static_assert(E::thunk_hdr_size == sizeof(hdr));
   static_assert(E::thunk_size == sizeof(entry));
 
+  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
   memcpy(buf, hdr, sizeof(hdr));
+  buf += sizeof(hdr);
 
-  for (i64 i = 0; i < symbols.size(); i++) {
-    u8 *loc = buf + sizeof(hdr) + i * sizeof(entry);
-    memcpy(loc, entry, sizeof(entry));
+  u64 P = output_section.shdr.sh_addr + offset + sizeof(hdr);
 
-    u64 S = symbols[i]->get_addr(ctx);
-    u64 P = output_section.shdr.sh_addr + offset + sizeof(hdr) + i * sizeof(entry);
-    *(ul32 *)(loc + 16) = S - P - 16;
+  for (Symbol<E> *sym : symbols) {
+    memcpy(buf, entry, sizeof(entry));
+    *(ul32 *)(buf + 12) = sym->get_addr(ctx) - P - 16;
+
+    buf += sizeof(entry);
+    P += sizeof(entry);
   }
 }
 
+template <>
+u64 get_eflags(Context<E> &ctx) {
+  return EF_ARM_EABI_VER5;
+}
+
 // ARM executables use an .ARM.exidx section to look up an exception
 // handling record for the current instruction pointer. The table needs
 // to be sorted by their addresses.
@@ -668,7 +702,13 @@ void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
 // likely that it's due to some historical reason.
 //
 // This function sorts .ARM.exidx records.
-static void sort_exidx(Context<E> &ctx, OutputSection<E> &osec) {
+void fixup_arm_exidx_section(Context<E> &ctx) {
+  Timer t(ctx, "fixup_arm_exidx_section");
+
+  OutputSection<E> *osec = find_section(ctx, SHT_ARM_EXIDX);
+  if (!osec)
+    return;
+
   // .ARM.exidx records consists of a signed 31-bit relative address
   // and a 32-bit value. The relative address indicates the start
   // address of a function that the record covers. The value is one of
@@ -688,11 +728,11 @@ static void sort_exidx(Context<E> &ctx, OutputSection<E> &osec) {
     ul32 val;
   };
 
-  if (osec.shdr.sh_size % sizeof(Entry))
+  if (osec->shdr.sh_size % sizeof(Entry))
     Fatal(ctx) << "invalid .ARM.exidx section size";
 
-  Entry *ent = (Entry *)(ctx.buf + osec.shdr.sh_offset);
-  i64 num_entries = osec.shdr.sh_size / sizeof(Entry);
+  Entry *ent = (Entry *)(ctx.buf + osec->shdr.sh_offset);
+  i64 num_entries = osec->shdr.sh_size / sizeof(Entry);
 
   // Entry's addresses are relative to themselves. In order to sort
   // records by addresses, we first translate them so that the addresses
@@ -719,38 +759,12 @@ static void sort_exidx(Context<E> &ctx, OutputSection<E> &osec) {
     if (is_relative(ent[i].val))
       ent[i].val = 0x7fff'ffff & (ent[i].val - offset);
   });
-}
-
-void fixup_arm_exidx_section(Context<E> &ctx) {
-  Timer t(ctx, "fixup_arm_exidx_section");
-
-  auto find_exidx = [&]() -> OutputSection<E> * {
-    for (Chunk<E> *chunk : ctx.chunks)
-      if (OutputSection<E> *osec = chunk->to_osec())
-        if (osec->shdr.sh_type == SHT_ARM_EXIDX)
-          return osec;
-    return nullptr;
-  };
-
-  OutputSection<E> *exidx = find_exidx();
-  if (!exidx)
-    return;
-
-  // Sort .ARM.exidx contents
-  sort_exidx(ctx, *exidx);
 
   // .ARM.exidx's sh_link should be set to the .text section index.
   // Runtime doesn't care about it, but the binutils's strip command does.
   if (ctx.shdr) {
-    auto find_text = [&]() -> Chunk<E> * {
-      for (Chunk<E> *chunk : ctx.chunks)
-        if (chunk->name == ".text")
-          return chunk;
-      return nullptr;
-    };
-
-    if (Chunk<E> *text = find_text()) {
-      exidx->shdr.sh_link = text->shndx;
+    if (Chunk<E> *text = find_section(ctx, ".text")) {
+      osec->shdr.sh_link = text->shndx;
       ctx.shdr->copy_buf(ctx);
     }
   }

elf/arch-arm64.cc

@@ -102,8 +102,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 }
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -123,6 +123,24 @@ void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
   }
 }
 
+static bool is_adrp(u8 *loc) {
+  // https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/ADRP--Form-PC-relative-address-to-4KB-page-
+  u32 insn = *(ul32 *)loc;
+  return (bits(insn, 31, 24) & 0b1001'1111) == 0b1001'0000;
+}
+
+static bool is_ldr(u8 *loc) {
+  // https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/LDR--immediate---Load-Register--immediate--
+  u32 insn = *(ul32 *)loc;
+  return (bits(insn, 31, 20) & 0b1111'1111'1100) == 0b1111'1001'0100;
+}
+
+static bool is_add(u8 *loc) {
+  // https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/ADD--immediate---Add--immediate--
+  u32 insn = *(ul32 *)loc;
+  return (bits(insn, 31, 20) & 0b1111'1111'1100) == 0b1001'0001'0000;
+}
+
 template <>
 void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
   std::span<const ElfRel<E>> rels = get_rels(ctx);
@@ -155,9 +173,10 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
 
     switch (rel.r_type) {
     case R_AARCH64_ABS64:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
     case R_AARCH64_LDST8_ABS_LO12_NC:
+    case R_AARCH64_ADD_ABS_LO12_NC:
       *(ul32 *)loc |= bits(S + A, 11, 0) << 10;
       break;
     case R_AARCH64_LDST16_ABS_LO12_NC:
@@ -172,57 +191,86 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_AARCH64_LDST128_ABS_LO12_NC:
       *(ul32 *)loc |= bits(S + A, 11, 4) << 10;
       break;
-    case R_AARCH64_ADD_ABS_LO12_NC:
-      *(ul32 *)loc |= bits(S + A, 11, 0) << 10;
+    case R_AARCH64_MOVW_UABS_G0:
+      check(S + A, 0, 1 << 16);
+      *(ul32 *)loc |= bits(S + A, 15, 0) << 5;
       break;
-    case R_AARCH64_MOVW_UABS_G0: {
-      i64 val = S + A;
-      check(val, 0, 1 << 16);
-      *(ul32 *)loc |= bits(val, 15, 0) << 5;
-      break;
-    }
     case R_AARCH64_MOVW_UABS_G0_NC:
       *(ul32 *)loc |= bits(S + A, 15, 0) << 5;
       break;
-    case R_AARCH64_MOVW_UABS_G1: {
-      i64 val = S + A;
-      check(val, 0, 1LL << 32);
-      *(ul32 *)loc |= bits(val, 31, 16) << 5;
+    case R_AARCH64_MOVW_UABS_G1:
+      check(S + A, 0, 1LL << 32);
+      *(ul32 *)loc |= bits(S + A, 31, 16) << 5;
       break;
-    }
     case R_AARCH64_MOVW_UABS_G1_NC:
       *(ul32 *)loc |= bits(S + A, 31, 16) << 5;
       break;
-    case R_AARCH64_MOVW_UABS_G2: {
-      i64 val = S + A;
-      check(val, 0, 1LL << 48);
-      *(ul32 *)loc |= bits(val, 47, 32) << 5;
+    case R_AARCH64_MOVW_UABS_G2:
+      check(S + A, 0, 1LL << 48);
+      *(ul32 *)loc |= bits(S + A, 47, 32) << 5;
       break;
-    }
     case R_AARCH64_MOVW_UABS_G2_NC:
       *(ul32 *)loc |= bits(S + A, 47, 32) << 5;
       break;
     case R_AARCH64_MOVW_UABS_G3:
       *(ul32 *)loc |= bits(S + A, 63, 48) << 5;
       break;
-    case R_AARCH64_ADR_GOT_PAGE: {
-      i64 val = page(G + GOT + A) - page(P);
-      check(val, -(1LL << 32), 1LL << 32);
-      write_adrp(loc, val);
+    case R_AARCH64_ADR_GOT_PAGE:
+      if (sym.has_got(ctx)) {
+        i64 val = page(G + GOT + A) - page(P);
+        check(val, -(1LL << 32), 1LL << 32);
+        write_adrp(loc, val);
+      } else {
+        // Relax GOT-loading ADRP+LDR to an immediate ADRP+ADD
+        i64 val = page(S + A) - page(P);
+        check(val, -(1LL << 32), 1LL << 32);
+        write_adrp(loc, val);
+
+        u32 reg = bits(*(ul32 *)loc, 4, 0);
+        *(ul32 *)(loc + 4) = 0x9100'0000 | (reg << 5) | reg; // ADD
+        *(ul32 *)(loc + 4) |= bits(S + A, 11, 0) << 10;
+        i++;
+      }
       break;
-    }
-    case R_AARCH64_ADR_PREL_PG_HI21: {
+    case R_AARCH64_ADR_PREL_PG_HI21:
+    case R_AARCH64_ADR_PREL_PG_HI21_NC: {
+      // The ARM64 psABI defines that an `ADRP x0, foo` and `ADD x0, x0,
+      // :lo12: foo` instruction pair to materialize a PC-relative address
+      // in a register can be relaxed to `NOP` followed by `ADR x0, foo`
+      // if foo is in PC ± 1 MiB.
+      if (ctx.arg.relax && sym.is_pcrel_linktime_const(ctx) &&
+          i + 1 < rels.size()) {
+        i64 val = S + A - P - 4;
+        const ElfRel<E> &rel2 = rels[i + 1];
+        if (sign_extend(val, 20) == val &&
+            rel2.r_type == R_AARCH64_ADD_ABS_LO12_NC &&
+            rel2.r_sym == rel.r_sym &&
+            rel2.r_offset == rel.r_offset + 4 &&
+            rel2.r_addend == rel.r_addend &&
+            is_adrp(loc) &&
+            is_add(loc + 4)) {
+          u32 reg1 = bits(*(ul32 *)loc, 4, 0);
+          u32 reg2 = bits(*(ul32 *)(loc + 4), 4, 0);
+          if (reg1 == reg2) {
+            *(ul32 *)loc = 0xd503'201f;              // nop
+            *(ul32 *)(loc + 4) = 0x1000'0000 | reg1; // adr
+            write_adr(loc + 4, val);
+            i++;
+            break;
+          }
+        }
+      }
+
       i64 val = page(S + A) - page(P);
-      check(val, -(1LL << 32), 1LL << 32);
+      if (rel.r_type == R_AARCH64_ADR_PREL_PG_HI21)
+        check(val, -(1LL << 32), 1LL << 32);
       write_adrp(loc, val);
       break;
     }
-    case R_AARCH64_ADR_PREL_LO21: {
-      i64 val = S + A - P;
-      check(val, -(1LL << 20), 1LL << 20);
-      write_adr(loc, val);
+    case R_AARCH64_ADR_PREL_LO21:
+      check(S + A - P, -(1LL << 20), 1LL << 20);
+      write_adr(loc, S + A - P);
       break;
-    }
     case R_AARCH64_CALL26:
     case R_AARCH64_JUMP26: {
       if (sym.is_remaining_undef_weak()) {
@@ -238,25 +286,23 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ul32 *)loc |= bits(val, 27, 2);
       break;
     }
+    case R_AARCH64_PLT32:
+      check(S + A - P, -(1LL << 31), 1LL << 31);
+      *(ul32 *)loc = S + A - P;
+      break;
     case R_AARCH64_CONDBR19:
-    case R_AARCH64_LD_PREL_LO19: {
-      i64 val = S + A - P;
-      check(val, -(1LL << 20), 1LL << 20);
-      *(ul32 *)loc |= bits(val, 20, 2) << 5;
+    case R_AARCH64_LD_PREL_LO19:
+      check(S + A - P, -(1LL << 20), 1LL << 20);
+      *(ul32 *)loc |= bits(S + A - P, 20, 2) << 5;
       break;
-    }
-    case R_AARCH64_PREL16: {
-      i64 val = S + A - P;
-      check(val, -(1LL << 15), 1LL << 15);
-      *(ul16 *)loc = val;
+    case R_AARCH64_PREL16:
+      check(S + A - P, -(1LL << 15), 1LL << 15);
+      *(ul16 *)loc = S + A - P;
       break;
-    }
-    case R_AARCH64_PREL32: {
-      i64 val = S + A - P;
-      check(val, -(1LL << 31), 1LL << 32);
-      *(ul32 *)loc = val;
+    case R_AARCH64_PREL32:
+      check(S + A - P, -(1LL << 31), 1LL << 32);
+      *(ul32 *)loc = S + A - P;
       break;
-    }
     case R_AARCH64_PREL64:
       *(ul64 *)loc = S + A - P;
       break;
@@ -309,6 +355,9 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       break;
     }
     case R_AARCH64_TLSLE_ADD_TPREL_LO12:
+      check(S + A - ctx.tp_addr, 0, 1 << 12);
+      *(ul32 *)loc |= bits(S + A - ctx.tp_addr, 11, 0) << 10;
+      break;
     case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC:
       *(ul32 *)loc |= bits(S + A - ctx.tp_addr, 11, 0) << 10;
       break;
@@ -322,38 +371,65 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ul32 *)loc |= bits(sym.get_tlsgd_addr(ctx) + A, 11, 0) << 10;
       break;
     case R_AARCH64_TLSDESC_ADR_PAGE21:
+      // ARM64 TLSDESC uses the following code sequence to materialize
+      // a TP-relative address in x0.
+      //
+      //   adrp    x0, 0
+      //       R_AARCH64_TLSDESC_ADR_PAGE21 foo
+      //   ldr     x1, [x0]
+      //       R_AARCH64_TLSDESC_LD64_LO12  foo
+      //   add     x0, x0, #0
+      //       R_AARCH64_TLSDESC_ADD_LO12   foo
+      //   blr     x1
+      //       R_AARCH64_TLSDESC_CALL       foo
+      //
+      // We may relax the instructions to the following for non-dlopen'd DSO
+      //
+      //   nop
+      //   nop
+      //   adrp    x0, :gottprel:foo
+      //   ldr     x0, [x0, :gottprel_lo12:foo]
+      //
+      // or to the following for executable.
+      //
+      //   nop
+      //   nop
+      //   movz    x0, :tls_offset_hi:foo, lsl #16
+      //   movk    x0, :tls_offset_lo:foo
       if (sym.has_tlsdesc(ctx)) {
         i64 val = page(sym.get_tlsdesc_addr(ctx) + A) - page(P);
         check(val, -(1LL << 32), 1LL << 32);
         write_adrp(loc, val);
       } else {
-        // adrp x0, 0 -> movz x0, #tls_ofset_hi, lsl #16
-        i64 val = (S + A - ctx.tp_addr);
-        check(val, -(1LL << 32), 1LL << 32);
-        *(ul32 *)loc = 0xd2a0'0000 | (bits(val, 32, 16) << 5);
+        *(ul32 *)loc = 0xd503'201f; // nop
       }
       break;
     case R_AARCH64_TLSDESC_LD64_LO12:
-      if (sym.has_tlsdesc(ctx)) {
+      if (sym.has_tlsdesc(ctx))
         *(ul32 *)loc |= bits(sym.get_tlsdesc_addr(ctx) + A, 11, 3) << 10;
-      } else {
-        // ldr x2, [x0] -> movk x0, #tls_ofset_lo
-        u32 offset_lo = (S + A - ctx.tp_addr) & 0xffff;
-        *(ul32 *)loc = 0xf280'0000 | (offset_lo << 5);
-      }
+      else
+        *(ul32 *)loc = 0xd503'201f; // nop
       break;
     case R_AARCH64_TLSDESC_ADD_LO12:
       if (sym.has_tlsdesc(ctx)) {
         *(ul32 *)loc |= bits(sym.get_tlsdesc_addr(ctx) + A, 11, 0) << 10;
+      } else if (sym.has_gottp(ctx)) {
+        *(ul32 *)loc = 0x9000'0000; // adrp x0, 0
+        write_adrp(loc, page(sym.get_gottp_addr(ctx) + A) - page(P));
       } else {
-        // add x0, x0, #0 -> nop
-        *(ul32 *)loc = 0xd503'201f;
+        *(ul32 *)loc = 0xd2a0'0000; // movz x0, 0, lsl #16
+        *(ul32 *)loc |= bits(S + A - ctx.tp_addr, 32, 16) << 5;
       }
       break;
     case R_AARCH64_TLSDESC_CALL:
-      if (!sym.has_tlsdesc(ctx)) {
-        // blr x2 -> nop
-        *(ul32 *)loc = 0xd503'201f;
+      if (sym.has_tlsdesc(ctx)) {
+        // Do nothing
+      } else if (sym.has_gottp(ctx)) {
+        *(ul32 *)loc = 0xf940'0000; // ldr x0, [x0, 0]
+        *(ul32 *)loc |= bits(sym.get_gottp_addr(ctx) + A, 11, 3) << 10;
+      } else {
+        *(ul32 *)loc = 0xf280'0000; // movk x0, 0
+        *(ul32 *)loc |= bits(S + A - ctx.tp_addr, 15, 0) << 5;
       }
       break;
     default:
@@ -368,17 +444,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     auto check = [&](i64 val, i64 lo, i64 hi) {
       if (val < lo || hi <= val)
         Error(ctx) << *this << ": relocation " << rel << " against "
@@ -424,48 +495,77 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
-
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
+    u8 *loc = (u8 *)(contents.data() + rel.r_offset);
 
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
 
     switch (rel.r_type) {
     case R_AARCH64_ABS64:
       scan_dyn_absrel(ctx, sym, rel);
       break;
+    case R_AARCH64_MOVW_UABS_G3:
+      scan_absrel(ctx, sym, rel);
+      break;
     case R_AARCH64_ADR_GOT_PAGE:
+      // An ADR_GOT_PAGE and GOT_LO12_NC relocation pair is used to load a
+      // symbol's address from GOT. If the GOT value is a link-time
+      // constant, we may be able to rewrite the ADRP+LDR instruction pair
+      // with an ADRP+ADD, eliminating a GOT memory load.
+      if (ctx.arg.relax && sym.is_pcrel_linktime_const(ctx) &&
+          i + 1 < rels.size()) {
+        // ADRP+LDR must be consecutive and use the same register to relax.
+        const ElfRel<E> &rel2 = rels[i + 1];
+        if (rel2.r_type == R_AARCH64_LD64_GOT_LO12_NC &&
+            rel2.r_offset == rel.r_offset + 4 &&
+            rel2.r_sym == rel.r_sym &&
+            rel.r_addend == 0 &&
+            rel2.r_addend == 0 &&
+            is_adrp(loc) &&
+            is_ldr(loc + 4)) {
+          u32 rd = bits(*(ul32 *)loc, 4, 0);
+          u32 rn = bits(*(ul32 *)(loc + 4), 9, 5);
+          u32 rt = bits(*(ul32 *)(loc + 4), 4, 0);
+          if (rd == rn && rn == rt) {
+            i++;
+            break;
+          }
+        }
+      }
+      sym.flags |= NEEDS_GOT;
+      break;
     case R_AARCH64_LD64_GOT_LO12_NC:
     case R_AARCH64_LD64_GOTPAGE_LO15:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
     case R_AARCH64_CALL26:
     case R_AARCH64_JUMP26:
+    case R_AARCH64_PLT32:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21:
     case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
       break;
     case R_AARCH64_ADR_PREL_PG_HI21:
+    case R_AARCH64_ADR_PREL_PG_HI21_NC:
       scan_pcrel(ctx, sym, rel);
       break;
     case R_AARCH64_TLSGD_ADR_PAGE21:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_TLSGD;
       break;
-    case R_AARCH64_TLSDESC_ADR_PAGE21:
-    case R_AARCH64_TLSDESC_LD64_LO12:
-    case R_AARCH64_TLSDESC_ADD_LO12:
-      if (!relax_tlsdesc(ctx, sym))
-        sym.flags.fetch_or(NEEDS_TLSDESC, std::memory_order_relaxed);
+    case R_AARCH64_TLSDESC_CALL:
+      scan_tlsdesc(ctx, sym);
+      break;
+    case R_AARCH64_TLSLE_MOVW_TPREL_G2:
+    case R_AARCH64_TLSLE_ADD_TPREL_LO12:
+    case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC:
+      check_tlsle(ctx, sym, rel);
       break;
     case R_AARCH64_ADD_ABS_LO12_NC:
     case R_AARCH64_ADR_PREL_LO21:
@@ -482,20 +582,18 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_AARCH64_MOVW_UABS_G1_NC:
     case R_AARCH64_MOVW_UABS_G2:
     case R_AARCH64_MOVW_UABS_G2_NC:
-    case R_AARCH64_MOVW_UABS_G3:
     case R_AARCH64_PREL16:
     case R_AARCH64_PREL32:
     case R_AARCH64_PREL64:
+    case R_AARCH64_TLSGD_ADD_LO12_NC:
     case R_AARCH64_TLSLE_MOVW_TPREL_G0:
     case R_AARCH64_TLSLE_MOVW_TPREL_G0_NC:
     case R_AARCH64_TLSLE_MOVW_TPREL_G1:
     case R_AARCH64_TLSLE_MOVW_TPREL_G1_NC:
-    case R_AARCH64_TLSLE_MOVW_TPREL_G2:
     case R_AARCH64_TLSLE_ADD_TPREL_HI12:
-    case R_AARCH64_TLSLE_ADD_TPREL_LO12:
-    case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC:
-    case R_AARCH64_TLSGD_ADD_LO12_NC:
-    case R_AARCH64_TLSDESC_CALL:
+    case R_AARCH64_TLSDESC_ADR_PAGE21:
+    case R_AARCH64_TLSDESC_LD64_LO12:
+    case R_AARCH64_TLSDESC_ADD_LO12:
       break;
     default:
       Error(ctx) << *this << ": unknown relocation: " << rel;
@@ -505,24 +603,26 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
 template <>
 void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
-  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
-
-  static const ul32 data[] = {
+  static const ul32 insn[] = {
     0x9000'0010, // adrp x16, 0   # R_AARCH64_ADR_PREL_PG_HI21
     0x9100'0210, // add  x16, x16 # R_AARCH64_ADD_ABS_LO12_NC
     0xd61f'0200, // br   x16
+    0xd503'201f, // nop
   };
 
-  static_assert(E::thunk_size == sizeof(data));
+  static_assert(E::thunk_size == sizeof(insn));
 
-  for (i64 i = 0; i < symbols.size(); i++) {
-    u64 S = symbols[i]->get_addr(ctx);
-    u64 P = output_section.shdr.sh_addr + offset + i * E::thunk_size;
+  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
+  u64 P = output_section.shdr.sh_addr + offset;
 
-    u8 *loc = buf + i * E::thunk_size;
-    memcpy(loc , data, sizeof(data));
-    write_adrp(loc, page(S) - page(P));
-    *(ul32 *)(loc + 4) |= bits(S, 11, 0) << 10;
+  for (Symbol<E> *sym : symbols) {
+    u64 S = sym->get_addr(ctx);
+    memcpy(buf, insn, E::thunk_size);
+    write_adrp(buf, page(S) - page(P));
+    *(ul32 *)(buf + 4) |= bits(S, 11, 0) << 10;
+
+    buf += E::thunk_size;
+    P += E::thunk_size;
   }
 }
 

elf/arch-i386.cc

@@ -31,7 +31,7 @@
 // assume that %ebx is set to .got. For PDE, we need to create position-
 // dependent PLT entries which don't use %ebx.
 //
-// https://github.com/rui314/mold/wiki/i386-psabi.pdf
+// https://github.com/rui314/psabi/blob/main/i386.pdf
 
 #include "mold.h"
 
@@ -42,8 +42,6 @@ using E = I386;
 template <>
 i64 get_addend(u8 *loc, const ElfRel<E> &rel) {
   switch (rel.r_type) {
-  case R_386_NONE:
-    return 0;
   case R_386_8:
   case R_386_PC8:
     return *loc;
@@ -66,8 +64,9 @@ i64 get_addend(u8 *loc, const ElfRel<E> &rel) {
   case R_386_SIZE32:
   case R_386_TLS_GOTDESC:
     return *(ul32 *)loc;
+  default:
+    return 0;
   }
-  unreachable();
 }
 
 template <>
@@ -141,6 +140,7 @@ void write_plt_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
       0xcc,                   // (padding)
     };
     memcpy(buf, insn, sizeof(insn));
+    *(ul32 *)(buf + 5) = sym.get_plt_idx(ctx) * sizeof(ElfRel<E>);
     *(ul32 *)(buf + 11) = sym.get_gotplt_addr(ctx) - ctx.got->shdr.sh_addr;
   } else {
     static const u8 insn[] = {
@@ -150,10 +150,9 @@ void write_plt_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
       0xcc,                   // (padding)
     };
     memcpy(buf, insn, sizeof(insn));
+    *(ul32 *)(buf + 5) = sym.get_plt_idx(ctx) * sizeof(ElfRel<E>);
     *(ul32 *)(buf + 11) = sym.get_gotplt_addr(ctx);
   }
-
-  *(ul32 *)(buf + 5) = sym.get_plt_idx(ctx) * sizeof(ElfRel<E>);
 }
 
 template <>
@@ -178,8 +177,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 }
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -203,6 +202,58 @@ static u32 relax_got32x(u8 *loc) {
   return 0;
 }
 
+// Relax GD to LE
+static void relax_gd_to_le(u8 *loc, ElfRel<E> rel, u64 val) {
+  static const u8 insn[] = {
+    0x65, 0xa1, 0, 0, 0, 0, // mov %gs:0, %eax
+    0x81, 0xc0, 0, 0, 0, 0, // add $tp_offset, %eax
+  };
+
+  switch (rel.r_type) {
+  case R_386_PLT32:
+  case R_386_PC32:
+    memcpy(loc - 3, insn, sizeof(insn));
+    *(ul32 *)(loc + 5) = val;
+    break;
+  case R_386_GOT32:
+  case R_386_GOT32X:
+    memcpy(loc - 2, insn, sizeof(insn));
+    *(ul32 *)(loc + 6) = val;
+    break;
+  default:
+    unreachable();
+  }
+}
+
+// Relax LD to LE
+static void relax_ld_to_le(u8 *loc, ElfRel<E> rel, u64 val) {
+  switch (rel.r_type) {
+  case R_386_PLT32:
+  case R_386_PC32: {
+    static const u8 insn[] = {
+      0x65, 0xa1, 0, 0, 0, 0, // mov %gs:0, %eax
+      0x2d, 0, 0, 0, 0,       // sub $tls_size, %eax
+    };
+    memcpy(loc - 2, insn, sizeof(insn));
+    *(ul32 *)(loc + 5) = val;
+    break;
+  }
+  case R_386_GOT32:
+  case R_386_GOT32X: {
+    static const u8 insn[] = {
+      0x65, 0xa1, 0, 0, 0, 0, // mov %gs:0, %eax
+      0x2d, 0, 0, 0, 0,       // sub $tls_size, %eax
+      0x90,                   // nop
+    };
+    memcpy(loc - 2, insn, sizeof(insn));
+    *(ul32 *)(loc + 5) = val;
+    break;
+  }
+  default:
+    unreachable();
+  }
+}
+
 template <>
 void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
   std::span<const ElfRel<E>> rels = get_rels(ctx);
@@ -234,33 +285,25 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     u64 GOT = ctx.got->shdr.sh_addr;
 
     switch (rel.r_type) {
-    case R_386_8: {
-      i64 val = S + A;
-      check(val, 0, 1 << 8);
-      *loc = val;
+    case R_386_8:
+      check(S + A, 0, 1 << 8);
+      *loc = S + A;
       break;
-    }
-    case R_386_16: {
-      i64 val = S + A;
-      check(val, 0, 1 << 16);
-      *(ul16 *)loc = val;
+    case R_386_16:
+      check(S + A, 0, 1 << 16);
+      *(ul16 *)loc = S + A;
       break;
-    }
     case R_386_32:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
-    case R_386_PC8: {
-      i64 val = S + A - P;
-      check(val, -(1 << 7), 1 << 7);
-      *loc = val;
+    case R_386_PC8:
+      check(S + A - P, -(1 << 7), 1 << 7);
+      *loc = S + A - P;
       break;
-    }
-    case R_386_PC16: {
-      i64 val = S + A - P;
-      check(val, -(1 << 15), 1 << 15);
-      *(ul16 *)loc = val;
+    case R_386_PC16:
+      check(S + A - P, -(1 << 15), 1 << 15);
+      *(ul16 *)loc = S + A - P;
       break;
-    }
     case R_386_PC32:
     case R_386_PLT32:
       *(ul32 *)loc = S + A - P;
@@ -295,72 +338,16 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ul32 *)loc = sym.get_gottp_addr(ctx) + A;
       break;
     case R_386_TLS_GD:
-      if (sym.has_tlsgd(ctx)) {
+      if (sym.has_tlsgd(ctx))
         *(ul32 *)loc = sym.get_tlsgd_addr(ctx) + A - GOT;
-      } else {
-        // Relax GD to LE
-        switch (rels[i + 1].r_type) {
-        case R_386_PLT32:
-        case R_386_PC32: {
-          static const u8 insn[] = {
-            0x65, 0xa1, 0, 0, 0, 0, // mov %gs:0, %eax
-            0x81, 0xe8, 0, 0, 0, 0, // sub $val, %eax
-          };
-          memcpy(loc - 3, insn, sizeof(insn));
-          *(ul32 *)(loc + 5) = ctx.tp_addr - S - A;
-          break;
-        }
-        case R_386_GOT32:
-        case R_386_GOT32X: {
-          static const u8 insn[] = {
-            0x65, 0xa1, 0, 0, 0, 0, // mov %gs:0, %eax
-            0x81, 0xe8, 0, 0, 0, 0, // sub $val, %eax
-          };
-          memcpy(loc - 2, insn, sizeof(insn));
-          *(ul32 *)(loc + 6) = ctx.tp_addr - S - A;
-          break;
-        }
-        default:
-          unreachable();
-        }
-
-        i++;
-      }
+      else
+        relax_gd_to_le(loc, rels[++i], S - ctx.tp_addr);
       break;
     case R_386_TLS_LDM:
-      if (ctx.got->has_tlsld(ctx)) {
+      if (ctx.got->has_tlsld(ctx))
         *(ul32 *)loc = ctx.got->get_tlsld_addr(ctx) + A - GOT;
-      } else {
-        // Relax LD to LE
-        switch (rels[i + 1].r_type) {
-        case R_386_PLT32:
-        case R_386_PC32: {
-          static const u8 insn[] = {
-            0x31, 0xc0,             // xor %eax, %eax
-            0x65, 0x8b, 0x00,       // mov %gs:(%eax), %eax
-            0x81, 0xe8, 0, 0, 0, 0, // sub $tls_size, %eax
-          };
-          memcpy(loc - 2, insn, sizeof(insn));
-          break;
-        }
-        case R_386_GOT32:
-        case R_386_GOT32X: {
-          static const u8 insn[] = {
-            0x31, 0xc0,             // xor %eax, %eax
-            0x65, 0x8b, 0x00,       // mov %gs:(%eax), %eax
-            0x81, 0xe8, 0, 0, 0, 0, // sub $tls_size, %eax
-            0x90,                   // nop
-          };
-          memcpy(loc - 2, insn, sizeof(insn));
-          break;
-        }
-        default:
-          unreachable();
-        }
-
-        *(ul32 *)(loc + 5) = ctx.tp_addr - ctx.tls_begin;
-        i++;
-      }
+        else
+        relax_ld_to_le(loc, rels[++i], ctx.tp_addr - ctx.tls_begin);
       break;
     case R_386_TLS_LDO_32:
       *(ul32 *)loc = S + A - ctx.dtp_addr;
@@ -369,13 +356,32 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ul32 *)loc = sym.esym().st_size + A;
       break;
     case R_386_TLS_GOTDESC:
+      // i386 TLSDESC uses the following code sequence to materialize
+      // a TP-relative address in %eax.
+      //
+      //   lea    0(%ebx), %eax
+      //       R_386_TLS_GOTDESC   foo
+      //   call   *(%eax)
+      //       R_386_TLS_DESC_CALL foo
+      //
+      // We may relax the instructions to the following for non-dlopen'd DSO
+      //
+      //   mov     foo@GOTTPOFF(%ebx), %eax
+      //   nop
+      //
+      // or to the following for executable.
+      //
+      //   mov     $foo@TPOFF, %eax
+      //   nop
       if (sym.has_tlsdesc(ctx)) {
         *(ul32 *)loc = sym.get_tlsdesc_addr(ctx) + A - GOT;
+      } else if (sym.has_gottp(ctx)) {
+        loc[-2] = 0x8b;
+        loc[-1] = 0x83;
+        *(ul32 *)loc = sym.get_gottp_addr(ctx) + A - GOT;
       } else {
-        static const u8 insn[] = {
-          0x8d, 0x05, 0, 0, 0, 0, // lea 0, %eax
-        };
-        memcpy(loc - 2, insn, sizeof(insn));
+        loc[-2] = 0x90;
+        loc[-1] = 0xb8;
         *(ul32 *)loc = S + A - ctx.tp_addr;
       }
       break;
@@ -398,17 +404,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     auto check = [&](i64 val, i64 lo, i64 hi) {
       if (val < lo || hi <= val)
         Error(ctx) << *this << ": relocation " << rel << " against "
@@ -425,36 +426,28 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
     u64 GOT = ctx.got->shdr.sh_addr;
 
     switch (rel.r_type) {
-    case R_386_8: {
-      i64 val = S + A;
-      check(val, 0, 1 << 8);
-      *loc = val;
+    case R_386_8:
+      check(S + A, 0, 1 << 8);
+      *loc = S + A;
       break;
-    }
-    case R_386_16: {
-      i64 val = S + A;
-      check(val, 0, 1 << 16);
-      *(ul16 *)loc = val;
+    case R_386_16:
+      check(S + A, 0, 1 << 16);
+      *(ul16 *)loc = S + A;
       break;
-    }
     case R_386_32:
       if (std::optional<u64> val = get_tombstone(sym, frag))
         *(ul32 *)loc = *val;
       else
         *(ul32 *)loc = S + A;
       break;
-    case R_386_PC8: {
-      i64 val = S + A;
-      check(val, -(1 << 7), 1 << 7);
-      *loc = val;
+    case R_386_PC8:
+      check(S + A, -(1 << 7), 1 << 7);
+      *loc = S + A;
       break;
-    }
-    case R_386_PC16: {
-      i64 val = S + A;
-      check(val, -(1 << 15), 1 << 15);
-      *(ul16 *)loc = val;
+    case R_386_PC16:
+      check(S + A, -(1 << 15), 1 << 15);
+      *(ul16 *)loc = S + A;
       break;
-    }
     case R_386_PC32:
       *(ul32 *)loc = S + A;
       break;
@@ -489,19 +482,24 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = (u8 *)(contents.data() + rel.r_offset);
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
+
+    if (rel.r_type == R_386_TLS_GD || rel.r_type == R_386_TLS_LDM) {
+      if (i + 1 == rels.size())
+        Fatal(ctx) << *this << ": " << rel << " must be followed by PLT or GOT32";
+
+      if (u32 ty = rels[i + 1].r_type;
+          ty != R_386_PLT32 && ty != R_386_PC32 &&
+          ty != R_386_GOT32 && ty != R_386_GOT32X)
+        Fatal(ctx) << *this << ": " << rel << " must be followed by PLT or GOT32";
+    }
 
     switch (rel.r_type) {
     case R_386_8:
@@ -518,55 +516,47 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
       break;
     case R_386_GOT32:
     case R_386_GOTPC:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
-    case R_386_GOT32X: {
-      bool do_relax = ctx.arg.relax && !sym.is_imported &&
-                      sym.is_relative() && relax_got32x(loc - 2);
-      if (!do_relax)
-        sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+    case R_386_GOT32X:
+      // We always want to relax GOT32X even if --no-relax is given
+      // because static PIE doesn't work without it.
+      if (sym.is_pcrel_linktime_const(ctx) && relax_got32x(loc - 2)) {
+        // Do nothing
+      } else {
+        sym.flags |= NEEDS_GOT;
+      }
       break;
-    }
     case R_386_PLT32:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_386_TLS_GOTIE:
-    case R_386_TLS_LE:
     case R_386_TLS_IE:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
       break;
     case R_386_TLS_GD:
-      if (i + 1 == rels.size())
-        Fatal(ctx) << *this << ": TLS_GD reloc must be followed by PLT or GOT32";
-
-      if (u32 ty = rels[i + 1].r_type;
-          ty != R_386_PLT32 && ty != R_386_PC32 &&
-          ty != R_386_GOT32 && ty != R_386_GOT32X)
-        Fatal(ctx) << *this << ": TLS_GD reloc must be followed by PLT or GOT32";
-
-      if (relax_tlsgd(ctx, sym))
+      // We always relax if -static because libc.a doesn't contain
+      // __tls_get_addr().
+      if ((ctx.arg.relax && sym.is_tprel_linktime_const(ctx)) ||
+          ctx.arg.is_static)
         i++;
       else
-        sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+        sym.flags |= NEEDS_TLSGD;
       break;
     case R_386_TLS_LDM:
-      if (i + 1 == rels.size())
-        Fatal(ctx) << *this << ": TLS_LDM reloc must be followed by PLT or GOT32";
-
-      if (u32 ty = rels[i + 1].r_type;
-          ty != R_386_PLT32 && ty != R_386_PC32 &&
-          ty != R_386_GOT32 && ty != R_386_GOT32X)
-        Fatal(ctx) << *this << ": TLS_LDM reloc must be followed by PLT or GOT32";
-
-      if (relax_tlsld(ctx))
+      // We always relax if -static because libc.a doesn't contain
+      // __tls_get_addr().
+      if (ctx.arg.is_static || (ctx.arg.relax && !ctx.arg.shared))
         i++;
       else
-        ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+        ctx.needs_tlsld = true;
       break;
     case R_386_TLS_GOTDESC:
-      if (!relax_tlsdesc(ctx, sym))
-        sym.flags.fetch_or(NEEDS_TLSDESC, std::memory_order_relaxed);
+      scan_tlsdesc(ctx, sym);
+      break;
+    case R_386_TLS_LE:
+      check_tlsle(ctx, sym, rel);
       break;
     case R_386_GOTOFF:
     case R_386_TLS_LDO_32:

elf/arch-loongarch.cc

@@ -0,0 +1,686 @@
+// LoongArch is a new RISC ISA announced in 2021 by Loongson. The ISA
+// feels like a modernized MIPS with a hint of RISC-V flavor, although
+// it's not compatible with either one.
+//
+// While LoongArch is a fresh and clean ISA, its technological advantage
+// over other modern RISC ISAs such as RISC-V doesn't seem to be very
+// significant. It appears that the real selling point of LoongArch is
+// that the ISA is developed and controlled by a Chinese company,
+// reflecting a desire for domestic CPUs. Loongson is actively working on
+// bootstrapping the entire ecosystem for LoongArch, sending patches to
+// Linux, GCC, LLVM, etc.
+//
+// All instructions are 4 bytes long in LoongArch and aligned to 4-byte
+// boundaries. It has 32 general-purpose registers. Among these, $t0 - $t8
+// (aliases for $r12 - $r20) are temporary registers that we can use in
+// our PLT and range extension thunks.
+//
+// The psABI defines a few linker relaxations. We haven't supported them
+// yet.
+//
+// https://loongson.github.io/LoongArch-Documentation/LoongArch-ELF-ABI-EN.html
+
+#if MOLD_LOONGARCH64 || MOLD_LOONGARCH32
+
+#include "mold.h"
+
+namespace mold::elf {
+
+using E = MOLD_TARGET;
+
+static u64 page(u64 val) {
+  return val & 0xffff'ffff'ffff'f000;
+}
+
+static u64 hi20(u64 val, u64 pc) {
+  // A PC-relative address with a 32 bit offset is materialized in a
+  // register with the following instructions:
+  //
+  //   pcalau12i $rN, %pc_hi20(sym)
+  //   addi.d    $rN, $rN, %lo12(sym)
+  //
+  // PCALAU12I materializes bits [63:12] by computing (pc + imm << 12)
+  // and zero-clear [11:0]. ADDI.D sign-extends its 12 bit immediate and
+  // add it to the register. To compensate the sign-extension, PCALAU12I
+  // needs to materialize a 0x1000 larger value than the desired [63:12]
+  // if [11:0] is sign-extended.
+  //
+  // This is similar but different from RISC-V because RISC-V's AUIPC
+  // doesn't zero-clear [11:0].
+  return bits(page(val + 0x800) - page(pc), 31, 12);
+}
+
+static u64 hi64(u64 val, u64 pc) {
+  // A PC-relative 64-bit address is materialized with the following
+  // instructions for the large code model:
+  //
+  //   pcalau12i $rN, %pc_hi20(sym)
+  //   addi.d    $rM, $zero, %lo12(sym)
+  //   lu32i.d   $rM, %pc64_lo20(sym)
+  //   lu52i.d   $rM, $r12, %pc64_hi12(sym)
+  //   add.d     $rN, $rN, $rM
+  //
+  // PCALAU12I computes (pc + imm << 12) to materialize a 64-bit value.
+  // ADDI.D adds a sign-extended 12 bit value to a register. LU32I.D and
+  // LU52I.D simply set bits to [51:31] and to [63:53], respectively.
+  //
+  // Compensating all the sign-extensions is a bit complicated.
+  bool x = val & 0x800;
+  bool y = (page(val + 0x800) - page(pc)) & 0x8000'0000;
+
+  if (x && !y)
+    return val - 0x1'0000'0000;
+  if (!x && y)
+    return val + 0x1'0000'0000;
+  return val;
+}
+
+static u64 higher20(u64 val, u64 pc) {
+  return bits(hi64(val, pc), 51, 32);
+}
+
+static u64 highest12(u64 val, u64 pc) {
+  return bits(hi64(val, pc), 63, 52);
+}
+
+static void write_k12(u8 *loc, u32 val) {
+  // opcode, [11:0], rj, rd
+  *(ul32 *)loc &= 0b1111111111'000000000000'11111'11111;
+  *(ul32 *)loc |= bits(val, 11, 0) << 10;
+}
+
+static void write_k16(u8 *loc, u32 val) {
+  // opcode, [15:0], rj, rd
+  *(ul32 *)loc &= 0b111111'0000000000000000'11111'11111;
+  *(ul32 *)loc |= bits(val, 15, 0) << 10;
+}
+
+static void write_j20(u8 *loc, u32 val) {
+  // opcode, [19:0], rd
+  *(ul32 *)loc &= 0b1111111'00000000000000000000'11111;
+  *(ul32 *)loc |= bits(val, 19, 0) << 5;
+}
+
+static void write_d5k16(u8 *loc, u32 val) {
+  // opcode, [15:0], rj, [20:16]
+  *(ul32 *)loc &= 0b111111'0000000000000000'11111'00000;
+  *(ul32 *)loc |= bits(val, 15, 0) << 10;
+  *(ul32 *)loc |= bits(val, 20, 16);
+}
+
+static void write_d10k16(u8 *loc, u32 val) {
+  // opcode, [15:0], [25:16]
+  *(ul32 *)loc &= 0b111111'0000000000000000'0000000000;
+  *(ul32 *)loc |= bits(val, 15, 0) << 10;
+  *(ul32 *)loc |= bits(val, 25, 16);
+}
+
+template <>
+void write_plt_header<E>(Context<E> &ctx, u8 *buf) {
+  static const ul32 insn_64[] = {
+    0x1a00'000e, // pcalau12i $t2, %pc_hi20(.got.plt)
+    0x0011'bdad, // sub.d     $t1, $t1, $t3
+    0x28c0'01cf, // ld.d      $t3, $t2, %lo12(.got.plt) # _dl_runtime_resolve
+    0x02ff'51ad, // addi.d    $t1, $t1, -44             # .plt entry
+    0x02c0'01cc, // addi.d    $t0, $t2, %lo12(.got.plt) # &.got.plt
+    0x0045'05ad, // srli.d    $t1, $t1, 1               # .plt entry offset
+    0x28c0'218c, // ld.d      $t0, $t0, 8               # link map
+    0x4c00'01e0, // jr        $t3
+  };
+
+  static const ul32 insn_32[] = {
+    0x1a00'000e, // pcalau12i $t2, %pc_hi20(.got.plt)
+    0x0011'3dad, // sub.w     $t1, $t1, $t3
+    0x2880'01cf, // ld.w      $t3, $t2, %lo12(.got.plt) # _dl_runtime_resolve
+    0x02bf'51ad, // addi.w    $t1, $t1, -44             # .plt entry
+    0x0280'01cc, // addi.w    $t0, $t2, %lo12(.got.plt) # &.got.plt
+    0x0044'89ad, // srli.w    $t1, $t1, 2               # .plt entry offset
+    0x2880'118c, // ld.w      $t0, $t0, 4               # link map
+    0x4c00'01e0, // jr        $t3
+  };
+
+  u64 gotplt = ctx.gotplt->shdr.sh_addr;
+  u64 plt = ctx.plt->shdr.sh_addr;
+
+  memcpy(buf, E::is_64 ? insn_64 : insn_32, E::plt_hdr_size);
+  write_j20(buf, hi20(gotplt, plt));
+  write_k12(buf + 8, gotplt);
+  write_k12(buf + 16, gotplt);
+}
+
+static const ul32 plt_entry_64[] = {
+  0x1a00'000f, // pcalau12i $t3, %pc_hi20([email protected])
+  0x28c0'01ef, // ld.d      $t3, $t3, %lo12([email protected])
+  0x4c00'01ed, // jirl      $t1, $t3, 0
+  0x0340'0000, // nop
+};
+
+static const ul32 plt_entry_32[] = {
+  0x1a00'000f, // pcalau12i $t3, %pc_hi20([email protected])
+  0x2880'01ef, // ld.w      $t3, $t3, %lo12([email protected])
+  0x4c00'01ed, // jirl      $t1, $t3, 0
+  0x0340'0000, // nop
+};
+
+template <>
+void write_plt_entry<E>(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
+  u64 gotplt = sym.get_gotplt_addr(ctx);
+  u64 plt = sym.get_plt_addr(ctx);
+
+  memcpy(buf, E::is_64 ? plt_entry_64 : plt_entry_32, E::plt_size);
+  write_j20(buf, hi20(gotplt, plt));
+  write_k12(buf + 4, gotplt);
+}
+
+template <>
+void write_pltgot_entry<E>(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
+  u64 got = sym.get_got_addr(ctx);
+  u64 plt = sym.get_plt_addr(ctx);
+
+  memcpy(buf, E::is_64 ? plt_entry_64 : plt_entry_32, E::plt_size);
+  write_j20(buf, hi20(got, plt));
+  write_k12(buf + 4, got);
+}
+
+template <>
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
+  u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
+
+  switch (rel.r_type) {
+  case R_NONE:
+    break;
+  case R_LARCH_ADD6:
+    *loc = (*loc & 0b1100'0000) | ((*loc + val) & 0b0011'1111);
+    break;
+  case R_LARCH_ADD8:
+    *loc += val;
+    break;
+  case R_LARCH_ADD16:
+    *(ul16 *)loc += val;
+    break;
+  case R_LARCH_ADD32:
+    *(ul32 *)loc += val;
+    break;
+  case R_LARCH_ADD64:
+    *(ul64 *)loc += val;
+    break;
+  case R_LARCH_SUB6:
+    *loc = (*loc & 0b1100'0000) | ((*loc - val) & 0b0011'1111);
+    break;
+  case R_LARCH_SUB8:
+    *loc -= val;
+    break;
+  case R_LARCH_SUB16:
+    *(ul16 *)loc -= val;
+    break;
+  case R_LARCH_SUB32:
+    *(ul32 *)loc -= val;
+    break;
+  case R_LARCH_SUB64:
+    *(ul64 *)loc -= val;
+    break;
+  case R_LARCH_32_PCREL:
+    *(ul32 *)loc = val - this->shdr.sh_addr - offset;
+    break;
+  case R_LARCH_64_PCREL:
+    *(ul64 *)loc = val - this->shdr.sh_addr - offset;
+    break;
+  default:
+    Fatal(ctx) << "unsupported relocation in .eh_frame: " << rel;
+  }
+}
+
+template <>
+void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
+  std::span<const ElfRel<E>> rels = get_rels(ctx);
+
+  ElfRel<E> *dynrel = nullptr;
+  if (ctx.reldyn)
+    dynrel = (ElfRel<E> *)(ctx.buf + ctx.reldyn->shdr.sh_offset +
+                           file.reldyn_offset + this->reldyn_offset);
+
+  for (i64 i = 0; i < rels.size(); i++) {
+    const ElfRel<E> &rel = rels[i];
+
+    if (rel.r_type == R_NONE || rel.r_type == R_LARCH_RELAX ||
+        rel.r_type == R_LARCH_MARK_LA || rel.r_type == R_LARCH_MARK_PCREL ||
+        rel.r_type == R_LARCH_ALIGN)
+      continue;
+
+    Symbol<E> &sym = *file.symbols[rel.r_sym];
+    u8 *loc = base + rel.r_offset;
+
+    auto check = [&](i64 val, i64 lo, i64 hi) {
+      if (val < lo || hi <= val)
+        Error(ctx) << *this << ": relocation " << rel << " against "
+                   << sym << " out of range: " << val << " is not in ["
+                   << lo << ", " << hi << ")";
+    };
+
+    auto check_branch = [&](i64 val, i64 lo, i64 hi) {
+      if (val & 0b11)
+        Error(ctx) << *this << ": misaligned symbol " << sym
+                   << " for relocation " << rel;
+      check(val, lo, hi);
+    };
+
+    // Unlike other psABIs, the LoongArch ABI uses the same relocation
+    // types to refer to GOT entries for thread-local symbols and regular
+    // ones. Therefore, G may refer to a TLSGD or a regular GOT slot
+    // depending on the symbol type.
+    //
+    // Note that as of August 2023, both GCC and Clang treat TLSLD relocs
+    // as if they were TLSGD relocs for LoongArch, which is a clear bug.
+    // We need to handle TLSLD relocs as synonyms for TLSGD relocs for the
+    // sake of bug compatibility.
+    auto get_got_idx = [&] {
+      if (sym.has_tlsgd(ctx))
+        return sym.get_tlsgd_idx(ctx);
+      return sym.get_got_idx(ctx);
+    };
+
+    u64 S = sym.get_addr(ctx);
+    u64 A = rel.r_addend;
+    u64 P = get_addr() + rel.r_offset;
+    u64 G = get_got_idx() * sizeof(Word<E>);
+    u64 GOT = ctx.got->shdr.sh_addr;
+
+    switch (rel.r_type) {
+    case R_LARCH_32:
+      if constexpr (E::is_64)
+        *(ul32 *)loc = S + A;
+      else
+        apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
+      break;
+    case R_LARCH_64:
+      assert(E::is_64);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
+      break;
+    case R_LARCH_B16:
+      check_branch(S + A - P, -(1 << 17), 1 << 17);
+      write_k16(loc, (S + A - P) >> 2);
+      break;
+    case R_LARCH_B21:
+      check_branch(S + A - P, -(1 << 22), 1 << 22);
+      write_d5k16(loc, (S + A - P) >> 2);
+      break;
+    case R_LARCH_B26: {
+      i64 val = S + A - P;
+      if (val < -(1 << 27) || (1 << 27) <= val)
+        val = get_thunk_addr(i) + A - P;
+      write_d10k16(loc, val >> 2);
+      break;
+    }
+    case R_LARCH_ABS_LO12:
+      write_k12(loc, S + A);
+      break;
+    case R_LARCH_ABS_HI20:
+      write_j20(loc, (S + A) >> 12);
+      break;
+    case R_LARCH_ABS64_LO20:
+      write_j20(loc, (S + A) >> 32);
+      break;
+    case R_LARCH_ABS64_HI12:
+      write_k12(loc, (S + A) >> 52);
+      break;
+    case R_LARCH_PCALA_LO12:
+      write_k12(loc, S + A);
+      break;
+    case R_LARCH_PCALA_HI20:
+      check(S + A - P, -(1LL << 31), 1LL << 31);
+      write_j20(loc, hi20(S + A, P));
+      break;
+    case R_LARCH_PCALA64_LO20:
+      write_j20(loc, higher20(S + A, P));
+      break;
+    case R_LARCH_PCALA64_HI12:
+      write_k12(loc, highest12(S + A, P));
+      break;
+    case R_LARCH_GOT_PC_LO12:
+      write_k12(loc, GOT + G + A);
+      break;
+    case R_LARCH_GOT_PC_HI20:
+      check(GOT + G + A - P, -(1LL << 31), 1LL << 31);
+      write_j20(loc, hi20(GOT + G + A, P));
+      break;
+    case R_LARCH_GOT64_PC_LO20:
+      write_j20(loc, higher20(GOT + G + A, P));
+      break;
+    case R_LARCH_GOT64_PC_HI12:
+      write_k12(loc, highest12(GOT + G + A, P));
+      break;
+    case R_LARCH_GOT_LO12:
+      write_k12(loc, GOT + G + A);
+      break;
+    case R_LARCH_GOT_HI20:
+      write_j20(loc, (GOT + G + A) >> 12);
+      break;
+    case R_LARCH_GOT64_LO20:
+      write_j20(loc, (GOT + G + A) >> 32);
+      break;
+    case R_LARCH_GOT64_HI12:
+      write_k12(loc, (GOT + G + A) >> 52);
+      break;
+    case R_LARCH_TLS_LE_LO12:
+      write_k12(loc, S + A - ctx.tp_addr);
+      break;
+    case R_LARCH_TLS_LE_HI20:
+      write_j20(loc, (S + A - ctx.tp_addr) >> 12);
+      break;
+    case R_LARCH_TLS_LE64_LO20:
+      write_j20(loc, (S + A - ctx.tp_addr) >> 32);
+      break;
+    case R_LARCH_TLS_LE64_HI12:
+      write_k12(loc, (S + A - ctx.tp_addr) >> 52);
+      break;
+    case R_LARCH_TLS_IE_PC_LO12:
+      write_k12(loc, sym.get_gottp_addr(ctx) + A);
+      break;
+    case R_LARCH_TLS_IE_PC_HI20:
+      check(sym.get_gottp_addr(ctx) + A - P, -(1LL << 31), 1LL << 31);
+      write_j20(loc, hi20(sym.get_gottp_addr(ctx) + A, P));
+      break;
+    case R_LARCH_TLS_IE64_PC_LO20:
+      write_j20(loc, higher20(sym.get_gottp_addr(ctx) + A, P));
+      break;
+    case R_LARCH_TLS_IE64_PC_HI12:
+      write_k12(loc, highest12(sym.get_gottp_addr(ctx) + A, P));
+      break;
+    case R_LARCH_TLS_IE_LO12:
+      write_k12(loc, sym.get_gottp_addr(ctx) + A);
+      break;
+    case R_LARCH_TLS_IE_HI20:
+      write_j20(loc, (sym.get_gottp_addr(ctx) + A) >> 12);
+      break;
+    case R_LARCH_TLS_IE64_LO20:
+      write_j20(loc, (sym.get_gottp_addr(ctx) + A) >> 32);
+      break;
+    case R_LARCH_TLS_IE64_HI12:
+      write_k12(loc, (sym.get_gottp_addr(ctx) + A) >> 52);
+      break;
+    case R_LARCH_TLS_LD_PC_HI20:
+    case R_LARCH_TLS_GD_PC_HI20:
+      check(sym.get_tlsgd_addr(ctx) + A - P, -(1LL << 31), 1LL << 31);
+      write_j20(loc, hi20(sym.get_tlsgd_addr(ctx) + A, P));
+      break;
+    case R_LARCH_TLS_LD_HI20:
+    case R_LARCH_TLS_GD_HI20:
+      write_j20(loc, (sym.get_tlsgd_addr(ctx) + A) >> 12);
+      break;
+    case R_LARCH_ADD6:
+      *loc = (*loc & 0b1100'0000) | ((*loc + S + A) & 0b0011'1111);
+      break;
+    case R_LARCH_ADD8:
+      *loc += S + A;
+      break;
+    case R_LARCH_ADD16:
+      *(ul16 *)loc += S + A;
+      break;
+    case R_LARCH_ADD32:
+      *(ul32 *)loc += S + A;
+      break;
+    case R_LARCH_ADD64:
+      *(ul64 *)loc += S + A;
+      break;
+    case R_LARCH_SUB6:
+      *loc = (*loc & 0b1100'0000) | ((*loc - S - A) & 0b0011'1111);
+      break;
+    case R_LARCH_SUB8:
+      *loc -= S + A;
+      break;
+    case R_LARCH_SUB16:
+      *(ul16 *)loc -= S + A;
+      break;
+    case R_LARCH_SUB32:
+      *(ul32 *)loc -= S + A;
+      break;
+    case R_LARCH_SUB64:
+      *(ul64 *)loc -= S + A;
+      break;
+    case R_LARCH_32_PCREL:
+      *(ul32 *)loc = S + A - P;
+      break;
+    case R_LARCH_64_PCREL:
+      *(ul64 *)loc = S + A - P;
+      break;
+    case R_LARCH_ADD_ULEB128:
+      overwrite_uleb(loc, read_uleb(loc) + S + A);
+      break;
+    case R_LARCH_SUB_ULEB128:
+      overwrite_uleb(loc, read_uleb(loc) - S - A);
+      break;
+    default:
+      unreachable();
+    }
+  }
+}
+
+template <>
+void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
+  std::span<const ElfRel<E>> rels = get_rels(ctx);
+
+  for (i64 i = 0; i < rels.size(); i++) {
+    const ElfRel<E> &rel = rels[i];
+    if (rel.r_type == R_NONE)
+      continue;
+
+    Symbol<E> &sym = *file.symbols[rel.r_sym];
+    u8 *loc = base + rel.r_offset;
+
+    if (!sym.file) {
+      record_undef_error(ctx, rel);
+      continue;
+    }
+
+    SectionFragment<E> *frag;
+    i64 frag_addend;
+    std::tie(frag, frag_addend) = get_fragment(ctx, rel);
+
+    u64 S = frag ? frag->get_addr(ctx) : sym.get_addr(ctx);
+    u64 A = frag ? frag_addend : (i64)rel.r_addend;
+
+    switch (rel.r_type) {
+    case R_LARCH_32:
+      *(ul32 *)loc = S + A;
+      break;
+    case R_LARCH_64:
+      if (std::optional<u64> val = get_tombstone(sym, frag))
+        *(ul64 *)loc = *val;
+      else
+        *(ul64 *)loc = S + A;
+      break;
+    case R_LARCH_ADD6:
+      *loc = (*loc & 0b1100'0000) | ((*loc + S + A) & 0b0011'1111);
+      break;
+    case R_LARCH_ADD8:
+      *loc += S + A;
+      break;
+    case R_LARCH_ADD16:
+      *(ul16 *)loc += S + A;
+      break;
+    case R_LARCH_ADD32:
+      *(ul32 *)loc += S + A;
+      break;
+    case R_LARCH_ADD64:
+      *(ul64 *)loc += S + A;
+      break;
+    case R_LARCH_SUB6:
+      *loc = (*loc & 0b1100'0000) | ((*loc - S - A) & 0b0011'1111);
+      break;
+    case R_LARCH_SUB8:
+      *loc -= S + A;
+      break;
+    case R_LARCH_SUB16:
+      *(ul16 *)loc -= S + A;
+      break;
+    case R_LARCH_SUB32:
+      *(ul32 *)loc -= S + A;
+      break;
+    case R_LARCH_SUB64:
+      *(ul64 *)loc -= S + A;
+      break;
+    case R_LARCH_TLS_DTPREL32:
+      if (std::optional<u64> val = get_tombstone(sym, frag))
+        *(ul32 *)loc = *val;
+      else
+        *(ul32 *)loc = S + A - ctx.dtp_addr;
+      break;
+    case R_LARCH_TLS_DTPREL64:
+      if (std::optional<u64> val = get_tombstone(sym, frag))
+        *(ul64 *)loc = *val;
+      else
+        *(ul64 *)loc = S + A - ctx.dtp_addr;
+      break;
+    case R_LARCH_ADD_ULEB128:
+      overwrite_uleb(loc, read_uleb(loc) + S + A);
+      break;
+    case R_LARCH_SUB_ULEB128:
+      overwrite_uleb(loc, read_uleb(loc) - S - A);
+      break;
+    default:
+      Fatal(ctx) << *this << ": invalid relocation for non-allocated sections: "
+                 << rel;
+      break;
+    }
+  }
+}
+
+template <>
+void InputSection<E>::scan_relocations(Context<E> &ctx) {
+  assert(shdr().sh_flags & SHF_ALLOC);
+
+  this->reldyn_offset = file.num_dynrel * sizeof(ElfRel<E>);
+  std::span<const ElfRel<E>> rels = get_rels(ctx);
+
+  // Scan relocations
+  for (i64 i = 0; i < rels.size(); i++) {
+    const ElfRel<E> &rel = rels[i];
+
+    if (rel.r_type == R_NONE || rel.r_type == R_LARCH_RELAX ||
+        rel.r_type == R_LARCH_MARK_LA || rel.r_type == R_LARCH_MARK_PCREL ||
+        rel.r_type == R_LARCH_ALIGN)
+      continue;
+
+    if (record_undef_error(ctx, rel))
+      continue;
+
+    Symbol<E> &sym = *file.symbols[rel.r_sym];
+
+    if (!sym.file) {
+      record_undef_error(ctx, rel);
+      continue;
+    }
+
+    if (sym.is_ifunc())
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
+
+    switch (rel.r_type) {
+    case R_LARCH_32:
+      if constexpr (E::is_64)
+        scan_absrel(ctx, sym, rel);
+      else
+        scan_dyn_absrel(ctx, sym, rel);
+      break;
+    case R_LARCH_64:
+      assert(E::is_64);
+      scan_dyn_absrel(ctx, sym, rel);
+      break;
+    case R_LARCH_B26:
+      if (sym.is_imported)
+        sym.flags |= NEEDS_PLT;
+      break;
+    case R_LARCH_GOT_HI20:
+    case R_LARCH_GOT_PC_HI20:
+      sym.flags |= NEEDS_GOT;
+      break;
+    case R_LARCH_TLS_IE_HI20:
+    case R_LARCH_TLS_IE_PC_HI20:
+      sym.flags |= NEEDS_GOTTP;
+      break;
+    case R_LARCH_TLS_LD_PC_HI20:
+    case R_LARCH_TLS_GD_PC_HI20:
+    case R_LARCH_TLS_LD_HI20:
+    case R_LARCH_TLS_GD_HI20:
+      sym.flags |= NEEDS_TLSGD;
+      break;
+    case R_LARCH_32_PCREL:
+    case R_LARCH_64_PCREL:
+      scan_pcrel(ctx, sym, rel);
+      break;
+    case R_LARCH_TLS_LE_HI20:
+    case R_LARCH_TLS_LE_LO12:
+    case R_LARCH_TLS_LE64_LO20:
+    case R_LARCH_TLS_LE64_HI12:
+      check_tlsle(ctx, sym, rel);
+      break;
+    case R_LARCH_B16:
+    case R_LARCH_B21:
+    case R_LARCH_ABS_HI20:
+    case R_LARCH_ABS_LO12:
+    case R_LARCH_ABS64_LO20:
+    case R_LARCH_ABS64_HI12:
+    case R_LARCH_PCALA_HI20:
+    case R_LARCH_PCALA_LO12:
+    case R_LARCH_PCALA64_LO20:
+    case R_LARCH_PCALA64_HI12:
+    case R_LARCH_GOT_PC_LO12:
+    case R_LARCH_GOT64_PC_LO20:
+    case R_LARCH_GOT64_PC_HI12:
+    case R_LARCH_GOT_LO12:
+    case R_LARCH_GOT64_LO20:
+    case R_LARCH_GOT64_HI12:
+    case R_LARCH_TLS_IE_PC_LO12:
+    case R_LARCH_TLS_IE64_PC_LO20:
+    case R_LARCH_TLS_IE64_PC_HI12:
+    case R_LARCH_TLS_IE_LO12:
+    case R_LARCH_TLS_IE64_LO20:
+    case R_LARCH_TLS_IE64_HI12:
+    case R_LARCH_ADD6:
+    case R_LARCH_SUB6:
+    case R_LARCH_ADD8:
+    case R_LARCH_SUB8:
+    case R_LARCH_ADD16:
+    case R_LARCH_SUB16:
+    case R_LARCH_ADD32:
+    case R_LARCH_SUB32:
+    case R_LARCH_ADD64:
+    case R_LARCH_SUB64:
+    case R_LARCH_ADD_ULEB128:
+    case R_LARCH_SUB_ULEB128:
+      break;
+    default:
+      Error(ctx) << *this << ": unknown relocation: " << rel;
+    }
+  }
+}
+
+template <>
+void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
+  static const ul32 insn[] = {
+    0x1a00'000c, // pcalau12i $t0, 0
+    0x02c0'018c, // addi.d    $t0, $t0, 0
+    0x4c00'0180, // jirl      $zero, $t0, 0
+    0x0340'0000, // nop
+  };
+
+  static_assert(E::thunk_size == sizeof(insn));
+
+  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
+  u64 P = output_section.shdr.sh_addr + offset;
+
+  for (Symbol<E> *sym : symbols) {
+    u64 S = sym->get_addr(ctx);
+
+    memcpy(buf, insn, sizeof(insn));
+    write_j20(buf, hi20(S, P));
+    write_k12(buf + 4, S);
+
+    buf += sizeof(insn);
+    P += sizeof(insn);
+  }
+}
+
+} // namespace mold::elf
+
+#endif

elf/arch-m68k.cc

@@ -12,7 +12,7 @@
 // access instructions and therefore can support position-independent
 // code without too much hassle.
 //
-// https://github.com/rui314/mold/wiki/m68k-psabi.pdf
+// https://github.com/rui314/psabi/blob/main/m68k.pdf
 
 #include "mold.h"
 
@@ -56,8 +56,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 }
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -126,7 +126,7 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
 
     switch (rel.r_type) {
     case R_68K_32:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
     case R_68K_16:
       write16(S + A);
@@ -221,17 +221,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     SectionFragment<E> *frag;
     i64 frag_addend;
     std::tie(frag, frag_addend) = get_fragment(ctx, rel);
@@ -262,16 +257,11 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
       Error(ctx) << sym << ": GNU ifunc symbol is not supported on m68k";
 
@@ -294,35 +284,37 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_68K_GOTOFF32:
     case R_68K_GOTOFF16:
     case R_68K_GOTOFF8:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
     case R_68K_PLT32:
     case R_68K_PLT16:
     case R_68K_PLT8:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_68K_TLS_GD32:
     case R_68K_TLS_GD16:
     case R_68K_TLS_GD8:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_TLSGD;
       break;
     case R_68K_TLS_LDM32:
     case R_68K_TLS_LDM16:
     case R_68K_TLS_LDM8:
-      ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+      ctx.needs_tlsld = true;
       break;
     case R_68K_TLS_IE32:
     case R_68K_TLS_IE16:
     case R_68K_TLS_IE8:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
+      break;
+    case R_68K_TLS_LE32:
+    case R_68K_TLS_LE16:
+    case R_68K_TLS_LE8:
+      check_tlsle(ctx, sym, rel);
       break;
     case R_68K_TLS_LDO32:
     case R_68K_TLS_LDO16:
     case R_68K_TLS_LDO8:
-    case R_68K_TLS_LE32:
-    case R_68K_TLS_LE16:
-    case R_68K_TLS_LE8:
       break;
     default:
       Error(ctx) << *this << ": unknown relocation: " << rel;

elf/arch-ppc32.cc

@@ -30,7 +30,7 @@
 // 0x10000 (or 65536) bytes.
 //
 // Since each object file has its own .got2, %r30 refers to different
-// places in a merged .got2 for two functions that are came from different
+// places in a merged .got2 for two functions that came from different
 // input files. Therefore, %r30 makes sense only within a single function.
 //
 // Technically, we can reuse a %r30 value in our PLT if we create a PLT
@@ -38,7 +38,7 @@
 // doesn't seems to be worth its complexity. Our PLT simply doesn't rely
 // on a %r30 value.
 //
-// https://github.com/rui314/mold/wiki/ppc32-psabi.pdf
+// https://github.com/rui314/psabi/blob/main/ppc32.pdf
 
 #include "mold.h"
 
@@ -46,15 +46,11 @@ namespace mold::elf {
 
 using E = PPC32;
 
-static u64 lo(u64 x)       { return x & 0xffff; }
-static u64 hi(u64 x)       { return x >> 16; }
-static u64 ha(u64 x)       { return (x + 0x8000) >> 16; }
-static u64 high(u64 x)     { return (x >> 16) & 0xffff; }
-static u64 higha(u64 x)    { return ((x + 0x8000) >> 16) & 0xffff; }
-static u64 higher(u64 x)   { return (x >> 32) & 0xffff; }
-static u64 highera(u64 x)  { return ((x + 0x8000) >> 32) & 0xffff; }
-static u64 highest(u64 x)  { return x >> 48; }
-static u64 highesta(u64 x) { return (x + 0x8000) >> 48; }
+static u64 lo(u64 x)    { return x & 0xffff; }
+static u64 hi(u64 x)    { return x >> 16; }
+static u64 ha(u64 x)    { return (x + 0x8000) >> 16; }
+static u64 high(u64 x)  { return (x >> 16) & 0xffff; }
+static u64 higha(u64 x) { return ((x + 0x8000) >> 16) & 0xffff; }
 
 template <>
 void write_plt_header(Context<E> &ctx, u8 *buf) {
@@ -130,8 +126,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 }
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -157,9 +153,7 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     dynrel = (ElfRel<E> *)(ctx.buf + ctx.reldyn->shdr.sh_offset +
                            file.reldyn_offset + this->reldyn_offset);
 
-  u64 got2 = 0;
-  if (file.ppc32_got2)
-    got2 = file.ppc32_got2->output_section->shdr.sh_addr + file.ppc32_got2->offset;
+  u64 GOT2 = file.extra.got2 ? file.extra.got2->get_addr() : 0;
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
@@ -178,7 +172,7 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     switch (rel.r_type) {
     case R_PPC_ADDR32:
     case R_PPC_UADDR32:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
     case R_PPC_ADDR14:
       *(ub32 *)loc |= bits(S + A, 15, 2) << 2;
@@ -201,16 +195,16 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ub32 *)loc |= bits(S + A, 31, 2) << 2;
       break;
     case R_PPC_PLT16_LO:
-      *(ub16 *)loc = lo(G + GOT - got2 - A);
+      *(ub16 *)loc = lo(G + GOT - A - GOT2);
       break;
     case R_PPC_PLT16_HI:
-      *(ub16 *)loc = hi(G + GOT - got2 - A);
+      *(ub16 *)loc = hi(G + GOT - A - GOT2);
       break;
     case R_PPC_PLT16_HA:
-      *(ub16 *)loc = ha(G + GOT - got2 - A);
+      *(ub16 *)loc = ha(G + GOT - A - GOT2);
       break;
     case R_PPC_PLT32:
-      *(ub32 *)loc = G + GOT - got2 - A;
+      *(ub32 *)loc = G + GOT - A - GOT2;
       break;
     case R_PPC_REL14:
       *(ub32 *)loc |= bits(S + A - P, 15, 2) << 2;
@@ -299,17 +293,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     SectionFragment<E> *frag;
     i64 frag_addend;
     std::tie(frag, frag_addend) = get_fragment(ctx, rel);
@@ -341,18 +330,13 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
 
     switch (rel.r_type) {
     case R_PPC_ADDR32:
@@ -385,30 +369,32 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_PPC_PLT16_HI:
     case R_PPC_PLT16_HA:
     case R_PPC_PLT32:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
     case R_PPC_REL24:
     case R_PPC_PLTREL24:
     case R_PPC_PLTREL32:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_PPC_GOT_TLSGD16:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_TLSGD;
       break;
     case R_PPC_GOT_TLSLD16:
-      ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+      ctx.needs_tlsld = true;
       break;
     case R_PPC_GOT_TPREL16:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
+      break;
+    case R_PPC_TPREL16_LO:
+    case R_PPC_TPREL16_HI:
+    case R_PPC_TPREL16_HA:
+      check_tlsle(ctx, sym, rel);
       break;
     case R_PPC_LOCAL24PC:
     case R_PPC_TLS:
     case R_PPC_TLSGD:
     case R_PPC_TLSLD:
-    case R_PPC_TPREL16_LO:
-    case R_PPC_TPREL16_HI:
-    case R_PPC_TPREL16_HA:
     case R_PPC_DTPREL16_LO:
     case R_PPC_DTPREL16_HI:
     case R_PPC_DTPREL16_HA:
@@ -423,8 +409,6 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
 template <>
 void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
-  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
-
   static const ub32 local_thunk[] = {
     // Get this thunk's address
     0x7c08'02a6, // mflr    r0
@@ -443,22 +427,26 @@ void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
   static_assert(E::thunk_size == sizeof(plt_entry));
   static_assert(E::thunk_size == sizeof(local_thunk));
 
-  for (i64 i = 0; i < symbols.size(); i++) {
-    ub32 *loc = (ub32 *)(buf + i * E::thunk_size);
-    Symbol<E> &sym = *symbols[i];
+  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
+  u64 P = output_section.shdr.sh_addr + offset;
 
-    if (sym.has_plt(ctx)) {
-      memcpy(loc, plt_entry, sizeof(plt_entry));
-      u64 got = sym.has_got(ctx) ? sym.get_got_addr(ctx) : sym.get_gotplt_addr(ctx);
-      i64 val = got - get_addr(i) - 8;
-      loc[4] |= higha(val);
-      loc[5] |= lo(val);
+  for (Symbol<E> *sym : symbols) {
+    if (sym->has_plt(ctx)) {
+      u64 got =
+        sym->has_got(ctx) ? sym->get_got_addr(ctx) : sym->get_gotplt_addr(ctx);
+      i64 val = got - P - 8;
+      memcpy(buf, plt_entry, sizeof(plt_entry));
+      *(ub32 *)(buf + 16) |= higha(val);
+      *(ub32 *)(buf + 20) |= lo(val);
     } else {
-      memcpy(loc, local_thunk, sizeof(local_thunk));
-      i64 val = sym.get_addr(ctx) - get_addr(i) - 8;
-      loc[4] |= higha(val);
-      loc[5] |= lo(val);
+      i64 val = sym->get_addr(ctx) - P - 8;
+      memcpy(buf, local_thunk, sizeof(local_thunk));
+      *(ub32 *)(buf + 16) |= higha(val);
+      *(ub32 *)(buf + 20) |= lo(val);
     }
+
+    buf += E::thunk_size;
+    P += E::thunk_size;
   }
 }
 

elf/arch-ppc64v1.cc

@@ -5,14 +5,15 @@
 //
 // Even though they are similiar, ELFv1 isn't only different from ELFv2 in
 // endianness. The most notable difference is, in ELFv1, a function
-// pointer doesn't directly refer the entry point of a function but
-// instead refers a data structure so-called "function descriptor".
+// pointer doesn't directly refer to the entry point of a function but
+// instead refers to a data structure so-called "function descriptor".
 //
 // The function descriptor is essentially a pair of a function entry point
 // address and a value that should be set to %r2 before calling that
-// function. There is actually a third member, but we don't need to know
-// what that is. In total, the function descriptor is 24 bytes long. Here
-// is why we need it.
+// function. There is also a third member for "the environment pointer for
+// languages such as Pascal and PL/1" according to the psABI, but it looks
+// like no one acutally uses it. In total, the function descriptor is 24
+// bytes long. Here is why we need it.
 //
 // PPC generally lacks PC-relative data access instructions. Position-
 // independent code sets GOT + 0x8000 to %r2 and access global variables
@@ -25,8 +26,8 @@
 //
 // In this way, you can't call a function just by knowing the function's
 // entry point address. You also need to know a proper %r2 value for the
-// function. This is why a function pointer refers a tuple of an address
-// and a %r2 value.
+// function. This is why a function pointer refers to a tuple of an
+// address and a %r2 value.
 //
 // If a function call is made through PLT, PLT takes care of restoring %r2.
 // Therefore, the caller has to restore %r2 only for function calls
@@ -39,8 +40,10 @@
 // few different addresses for different purposes. It may not only have an
 // entry point address but may also have PLT and/or GOT addresses.
 // In PPCV1, it may have an OPD address in addition to these. OPD address
-// is used for relocations that refers the address of a function as a
+// is used for relocations that refers to the address of a function as a
 // function pointer.
+//
+// https://github.com/rui314/psabi/blob/main/ppc64v1.pdf
 
 #include "mold.h"
 
@@ -51,15 +54,11 @@ namespace mold::elf {
 
 using E = PPC64V1;
 
-static u64 lo(u64 x)       { return x & 0xffff; }
-static u64 hi(u64 x)       { return x >> 16; }
-static u64 ha(u64 x)       { return (x + 0x8000) >> 16; }
-static u64 high(u64 x)     { return (x >> 16) & 0xffff; }
-static u64 higha(u64 x)    { return ((x + 0x8000) >> 16) & 0xffff; }
-static u64 higher(u64 x)   { return (x >> 32) & 0xffff; }
-static u64 highera(u64 x)  { return ((x + 0x8000) >> 32) & 0xffff; }
-static u64 highest(u64 x)  { return x >> 48; }
-static u64 highesta(u64 x) { return (x + 0x8000) >> 48; }
+static u64 lo(u64 x)    { return x & 0xffff; }
+static u64 hi(u64 x)    { return x >> 16; }
+static u64 ha(u64 x)    { return (x + 0x8000) >> 16; }
+static u64 high(u64 x)  { return (x >> 16) & 0xffff; }
+static u64 higha(u64 x) { return ((x + 0x8000) >> 16) & 0xffff; }
 
 // .plt is used only for lazy symbol resolution on PPC64. All PLT
 // calls are made via range extension thunks even if they are within
@@ -73,29 +72,55 @@ void write_plt_header(Context<E> &ctx, u8 *buf) {
     0x7d88'02a6, // mflr    r12
     0x429f'0005, // bcl     20, 31, 4 // obtain PC
     0x7d68'02a6, // mflr    r11
-    0xe84b'0024, // ld      r2,36(r11)
     0x7d88'03a6, // mtlr    r12
-    0x7d62'5a14, // add     r11,r2,r11
+    0x3d6b'0000, // addis   r11, r11, GOTPLT_OFFSET@ha
+    0x396b'0000, // addi    r11, r11, GOTPLT_OFFSET@lo
     0xe98b'0000, // ld      r12,0(r11)
     0xe84b'0008, // ld      r2,8(r11)
     0x7d89'03a6, // mtctr   r12
     0xe96b'0010, // ld      r11,16(r11)
     0x4e80'0420, // bctr
-    // .quad .got.plt - .plt - 8
-    0x0000'0000,
-    0x0000'0000,
   };
 
   static_assert(sizeof(insn) == E::plt_hdr_size);
   memcpy(buf, insn, sizeof(insn));
-  *(ub64 *)(buf + 44) = ctx.gotplt->shdr.sh_addr - ctx.plt->shdr.sh_addr - 8;
+
+  i64 val = ctx.gotplt->shdr.sh_addr - ctx.plt->shdr.sh_addr - 8;
+  *(ub32 *)(buf + 16) |= higha(val);
+  *(ub32 *)(buf + 20) |= lo(val);
 }
 
 template <>
 void write_plt_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
-  i64 offset = ctx.plt->shdr.sh_addr - sym.get_plt_addr(ctx) - 4;
-  *(ub32 *)(buf + 0) = 0x3800'0000 | sym.get_plt_idx(ctx);   // li %r0, PLT_INDEX
-  *(ub32 *)(buf + 4) = 0x4b00'0000 | (offset & 0x00ff'ffff); // b  plt0
+  ub32 *loc = (ub32 *)buf;
+  i64 idx = sym.get_plt_idx(ctx);
+
+  // The PPC64 ELFv1 ABI requires PLT entries to be vary in size depending
+  // on their indices. Unlike other targets, .got.plt is filled not by us
+  // but by the loader, so we don't have a control over where the initial
+  // call to the PLT entry jumps to. So we need to strictly follow the PLT
+  // section layout as the loader expect it to be.
+  if (idx < 0x8000) {
+    static const ub32 insn[] = {
+      0x3800'0000, // li      r0, PLT_INDEX
+      0x4b00'0000, // b       plt0
+    };
+
+    memcpy(loc, insn, sizeof(insn));
+    loc[0] |= idx;
+    loc[1] |= (ctx.plt->shdr.sh_addr - sym.get_plt_addr(ctx) - 4) & 0x00ff'ffff;
+  } else {
+    static const ub32 insn[] = {
+      0x3c00'0000, // lis     r0, PLT_INDEX@high
+      0x6000'0000, // ori     r0, r0, PLT_INDEX@lo
+      0x4b00'0000, // b       plt0
+    };
+
+    memcpy(loc, insn, sizeof(insn));
+    loc[0] |= high(idx);
+    loc[1] |= lo(idx);
+    loc[2] |= (ctx.plt->shdr.sh_addr - sym.get_plt_addr(ctx) - 8) & 0x00ff'ffff;
+  }
 }
 
 // .plt.got is not necessary on PPC64 because range extension thunks
@@ -104,8 +129,8 @@ template <>
 void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {}
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -158,10 +183,10 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
 
     switch (rel.r_type) {
     case R_PPC64_ADDR64:
-      apply_toc_rel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_toc_rel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
     case R_PPC64_TOC:
-      apply_toc_rel(ctx, *ctx.extra.TOC, rel, loc, TOC, A, P, dynrel);
+      apply_toc_rel(ctx, *ctx.extra.TOC, rel, loc, TOC, A, P, &dynrel);
       break;
     case R_PPC64_TOC16_HA:
       *(ub16 *)loc = ha(S + A - TOC);
@@ -169,12 +194,10 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_PPC64_TOC16_LO:
       *(ub16 *)loc = lo(S + A - TOC);
       break;
-    case R_PPC64_TOC16_DS: {
-      i64 val = S + A - TOC;
-      check(val, -(1 << 15), 1 << 15);
-      *(ub16 *)loc |= val & 0xfffc;
+    case R_PPC64_TOC16_DS:
+      check(S + A - TOC, -(1 << 15), 1 << 15);
+      *(ub16 *)loc |= (S + A - TOC) & 0xfffc;
       break;
-    }
     case R_PPC64_TOC16_LO_DS:
       *(ub16 *)loc |= (S + A - TOC) & 0xfffc;
       break;
@@ -194,6 +217,9 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
         *(ub32 *)(loc + 4) = 0xe841'0028; // ld r2, 40(r1)
       break;
     }
+    case R_PPC64_REL32:
+      *(ub32 *)loc = S + A - P;
+      break;
     case R_PPC64_REL64:
       *(ub64 *)loc = S + A - P;
       break;
@@ -233,12 +259,12 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_PPC64_DTPREL16_HA:
       *(ub16 *)loc = ha(S + A - ctx.dtp_addr);
       break;
-    case R_PPC64_TPREL16_HA:
-      *(ub16 *)loc = ha(S + A - ctx.tp_addr);
-      break;
     case R_PPC64_DTPREL16_LO:
       *(ub16 *)loc = lo(S + A - ctx.dtp_addr);
       break;
+    case R_PPC64_TPREL16_HA:
+      *(ub16 *)loc = ha(S + A - ctx.tp_addr);
+      break;
     case R_PPC64_TPREL16_LO:
       *(ub16 *)loc = lo(S + A - ctx.tp_addr);
       break;
@@ -263,17 +289,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     auto check = [&](i64 val, i64 lo, i64 hi) {
       if (val < lo || hi <= val)
         Error(ctx) << *this << ": relocation " << rel << " against "
@@ -321,48 +342,45 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT | NEEDS_PPC_OPD,
-                         std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT | NEEDS_PPC_OPD;
 
     // Any relocation except R_PPC64_REL24 is considered as an
     // address-taking relocation.
     if (rel.r_type != R_PPC64_REL24 && sym.get_type() == STT_FUNC)
-      sym.flags.fetch_or(NEEDS_PPC_OPD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_PPC_OPD;
 
     switch (rel.r_type) {
     case R_PPC64_ADDR64:
-      scan_toc_rel(ctx, sym, rel);
-      break;
     case R_PPC64_TOC:
       scan_toc_rel(ctx, sym, rel);
       break;
     case R_PPC64_GOT_TPREL16_HA:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
       break;
     case R_PPC64_REL24:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_PPC64_PLT16_HA:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
     case R_PPC64_GOT_TLSGD16_HA:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_TLSGD;
       break;
     case R_PPC64_GOT_TLSLD16_HA:
-      ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+      ctx.needs_tlsld = true;
       break;
+    case R_PPC64_TPREL16_HA:
+    case R_PPC64_TPREL16_LO:
+      check_tlsle(ctx, sym, rel);
+      break;
+    case R_PPC64_REL32:
     case R_PPC64_REL64:
     case R_PPC64_TOC16_HA:
     case R_PPC64_TOC16_LO:
@@ -375,8 +393,6 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_PPC64_PLT16_LO_DS:
     case R_PPC64_PLTSEQ:
     case R_PPC64_PLTCALL:
-    case R_PPC64_TPREL16_HA:
-    case R_PPC64_TPREL16_LO:
     case R_PPC64_GOT_TPREL16_LO_DS:
     case R_PPC64_GOT_TLSGD16_LO:
     case R_PPC64_GOT_TLSLD16_LO:
@@ -394,8 +410,6 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
 template <>
 void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
-  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
-
   // If the destination is .plt.got, we save the current r2, read an
   // address of a function descriptor from .got, restore %r2 and jump
   // to the function.
@@ -450,26 +464,29 @@ void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
   static_assert(E::thunk_size == sizeof(plt_thunk));
   static_assert(E::thunk_size == sizeof(local_thunk));
 
-  for (i64 i = 0; i < symbols.size(); i++) {
-    Symbol<E> &sym = *symbols[i];
-    ub32 *loc = (ub32 *)(buf + i * E::thunk_size);
+  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
+  u64 P = output_section.shdr.sh_addr + offset;
 
-    if (sym.has_got(ctx)) {
-      memcpy(loc, pltgot_thunk, sizeof(pltgot_thunk));
-      i64 val = sym.get_got_addr(ctx) - ctx.extra.TOC->value;
-      loc[1] |= higha(val);
-      loc[2] |= lo(val);
-    } else if(sym.has_plt(ctx)) {
-      memcpy(loc, plt_thunk, sizeof(plt_thunk));
-      i64 val = sym.get_gotplt_addr(ctx) - ctx.extra.TOC->value;
-      loc[1] |= higha(val);
-      loc[2] |= lo(val);
+  for (Symbol<E> *sym : symbols) {
+    if (sym->has_got(ctx)) {
+      i64 val = sym->get_got_addr(ctx) - ctx.extra.TOC->value;
+      memcpy(buf, pltgot_thunk, sizeof(pltgot_thunk));
+      *(ub32 *)(buf + 4) |= higha(val);
+      *(ub32 *)(buf + 8) |= lo(val);
+    } else if(sym->has_plt(ctx)) {
+      i64 val = sym->get_gotplt_addr(ctx) - ctx.extra.TOC->value;
+      memcpy(buf, plt_thunk, sizeof(plt_thunk));
+      *(ub32 *)(buf + 4) |= higha(val);
+      *(ub32 *)(buf + 8) |= lo(val);
     } else {
-      memcpy(loc, local_thunk, sizeof(local_thunk));
-      i64 val = sym.get_addr(ctx, NO_OPD) - ctx.extra.TOC->value;
-      loc[0] |= higha(val);
-      loc[1] |= lo(val);
+      i64 val = sym->get_addr(ctx, NO_OPD) - ctx.extra.TOC->value;
+      memcpy(buf, local_thunk, sizeof(local_thunk));
+      *(ub32 *)buf |= higha(val);
+      *(ub32 *)(buf + 4) |= lo(val);
     }
+
+    buf += E::thunk_size;
+    P += E::thunk_size;
   }
 }
 
@@ -521,10 +538,10 @@ get_opd_sym_at(Context<E> &ctx, std::span<OpdSymbol> syms, u64 offset) {
 // However, in reality, .opd isn't a normal input section. It needs many
 // special treatments as follows:
 //
-// 1. A function symbol refers not a .text but an .opd. Its address works
-//    fine for address-taking relocations such as R_PPC64_ADDR64. However,
-//    R_PPC64_REL24 (which is used for branch instruction) needs a
-//    function's real address instead of the function's .opd address.
+// 1. A function symbol refers to not a .text but an .opd. Its address
+//    works fine for address-taking relocations such as R_PPC64_ADDR64.
+//    However, R_PPC64_REL24 (which is used for branch instruction) needs
+//    a function's real address instead of the function's .opd address.
 //    We need to read .opd contents to find out a function entry point
 //    address to apply R_PPC64_REL24.
 //
@@ -532,9 +549,9 @@ get_opd_sym_at(Context<E> &ctx, std::span<OpdSymbol> syms, u64 offset) {
 //    are taken. Just copying input .opd sections to an output would
 //    produces lots of dead .opd entries.
 //
-// 3. In this design, all function symbols refer an .opd section, and that
-//    doesn't work well with graph traversal optimizations such as garbage
-//    collection or identical comdat folding. For example, garbage
+// 3. In this design, all function symbols refer to an .opd section, and
+//    that doesn't work well with graph traversal optimizations such as
+//    garbage collection or identical comdat folding. For example, garbage
 //    collector would mark an .opd alive which in turn mark all functions
 //    thatare referenced by .opd as alive, effectively keeping all
 //    functions as alive.
@@ -547,9 +564,9 @@ get_opd_sym_at(Context<E> &ctx, std::span<OpdSymbol> syms, u64 offset) {
 //
 // So, in this function, we undo what the compiler did to .opd. We remove
 // function symbols from .opd and reattach them to their function entry
-// points. We also rewrite relocations that directly refer an input .opd
-// section so that they refer function symbols instead. We then mark input
-// .opd sections as dead.
+// points. We also rewrite relocations that directly refer to an input
+// .opd  section so that they refer to function symbols instead. We then
+// mark input .opd sections as dead.
 //
 // After this function, we mark symbols with the NEEDS_PPC_OPD flag if the
 // symbol needs an .opd entry. We then create an output .opd just like we
@@ -612,8 +629,8 @@ void ppc64v1_rewrite_opd(Context<E> &ctx) {
 }
 
 // When a function is exported, the dynamic symbol for the function should
-// refers the function's .opd entry. This function marks such symbols with
-// NEEDS_PPC_OPD.
+// refers to the function's .opd entry. This function marks such symbols
+// with NEEDS_PPC_OPD.
 void ppc64v1_scan_symbols(Context<E> &ctx) {
   tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
     for (Symbol<E> *sym : file->symbols)
@@ -640,13 +657,30 @@ void PPC64OpdSection::add_symbol(Context<E> &ctx, Symbol<E> *sym) {
   this->shdr.sh_size += ENTRY_SIZE;
 }
 
+i64 PPC64OpdSection::get_reldyn_size(Context<E> &ctx) const {
+  if (ctx.arg.pic)
+    return symbols.size() * 2;
+  return 0;
+}
+
 void PPC64OpdSection::copy_buf(Context<E> &ctx) {
   ub64 *buf = (ub64 *)(ctx.buf + this->shdr.sh_offset);
 
+  ElfRel<E> *rel = nullptr;
+  if (ctx.arg.pic)
+    rel = (ElfRel<E> *)(ctx.buf + ctx.reldyn->shdr.sh_offset + reldyn_offset);
+
   for (Symbol<E> *sym : symbols) {
-    *buf++ = sym->get_addr(ctx, NO_PLT | NO_OPD);
+    u64 addr = sym->get_addr(ctx, NO_PLT | NO_OPD);
+    *buf++ = addr;
     *buf++ = ctx.extra.TOC->value;
     *buf++ = 0;
+
+    if (ctx.arg.pic) {
+      u64 loc = sym->get_opd_addr(ctx);
+      *rel++ = ElfRel<E>(loc, E::R_RELATIVE, 0, addr);
+      *rel++ = ElfRel<E>(loc + 8, E::R_RELATIVE, 0, ctx.extra.TOC->value);
+    }
   }
 }
 

elf/arch-ppc64v2.cc

@@ -1,21 +1,29 @@
 // This file implements the PowerPC ELFv2 ABI which was standardized in
 // 2014. Modern little-endian PowerPC systems are based on this ABI.
 // The ABI is often referred to as "ppc64le". This shouldn't be confused
-// with "ppc64" which refers the original, big-endian PowerPC systems.
+// with "ppc64" which refers to the original, big-endian PowerPC systems.
 //
 // PPC64 is a bit tricky to support because PC-relative load/store
-// instructions are generally not available. Therefore, it's not easy
-// for position-independent code to load a value from, for example,
-// .got, as we can't do that with [PC + the offset to the .got entry].
+// instructions hadn't been available until Power10 which debuted in 2021.
+// Prior to Power10, it wasn't trivial for position-independent code (PIC)
+// to load a value from, for example, .got, as we can't do that with [PC +
+// the offset to the .got entry].
 //
-// We can get the program counter by the following four instructions
+// In the following, I'll explain how PIC is supported on pre-Power10
+// systems first and then explain what has changed with Power10.
+//
+//
+// Position-independent code on Power9 or earlier:
+//
+// We can get the program counter on older PPC64 systems with the
+// following four instructions
 //
 //   mflr  r1  // save the current link register to r1
 //   bl    .+4 // branch to the next instruction as if it were a function
 //   mflr  r0  // copy the return address to r0
 //   mtlr  r1  // restore the original link register value
 //
-// , but that's too expensive to do if we do this for each load/store.
+// , but it's too expensive to do if we do this for each load/store.
 //
 // As a workaround, most functions are compiled in such a way that r2 is
 // assumed to always contain the address of .got + 0x8000. With this, we
@@ -27,9 +35,9 @@
 // calls are usually within the same ELF module, so this mechanism is
 // efficient.
 //
-// In PPC64, a function usually have two entry points, global and local.
-// The global entry point is usually 8 bytes precedes the local entry
-// point. In between is the following instructions:
+// A function compiled for pre-Power10 usually has two entry points,
+// global and local. The global entry point usually 8 bytes precedes
+// the local entry point. In between is the following instructions:
 //
 //   addis r2, r12, .TOC.@ha
 //   addi  r2, r2,  .TOC.@lo + 4;
@@ -46,12 +54,31 @@
 // to that address. Then the callee computes its own TOC pointer using
 // r12.
 //
+//
+// Position-independent code on Power10:
+//
+// Power10 added 8-bytes-long instructions to the ISA. Some of them are
+// PC-relative load/store instructions that take 34 bits offsets.
+// Functions compiled with `-mcpu=power10` use these instructions for PIC.
+// r2 does not have a special meaning in such fucntions.
+//
+// When a fucntion compiled for Power10 calls a function that uses the TOC
+// pointer, we need to compute a correct value for TOC and set it to r2
+// before transferring the control to the callee. Thunks are responsible
+// for doing it.
+//
+// `_NOTOC` relocations such as `R_PPC64_REL24_NOTOC` indicate that the
+// callee does not use TOC (i.e. compiled with `-mcpu=power10`). If a
+// function using TOC is referenced via a `_NOTOC` relocation, that call
+// is made through a range extension thunk.
+//
+//
 // Note on section names: the PPC64 psABI uses a weird naming convention
 // which calls .got.plt .plt. We ignored that part because it's just
 // confusing. Since the runtime only cares about segments, we should be
 // able to name sections whatever we want.
 //
-// https://openpowerfoundation.org/specifications/64bitelfabi/
+// https://github.com/rui314/psabi/blob/main/ppc64v2.pdf
 
 #include "mold.h"
 
@@ -59,15 +86,15 @@ namespace mold::elf {
 
 using E = PPC64V2;
 
-static u64 lo(u64 x)       { return x & 0xffff; }
-static u64 hi(u64 x)       { return x >> 16; }
-static u64 ha(u64 x)       { return (x + 0x8000) >> 16; }
-static u64 high(u64 x)     { return (x >> 16) & 0xffff; }
-static u64 higha(u64 x)    { return ((x + 0x8000) >> 16) & 0xffff; }
-static u64 higher(u64 x)   { return (x >> 32) & 0xffff; }
-static u64 highera(u64 x)  { return ((x + 0x8000) >> 32) & 0xffff; }
-static u64 highest(u64 x)  { return x >> 48; }
-static u64 highesta(u64 x) { return (x + 0x8000) >> 48; }
+static u64 lo(u64 x)    { return x & 0xffff; }
+static u64 hi(u64 x)    { return x >> 16; }
+static u64 ha(u64 x)    { return (x + 0x8000) >> 16; }
+static u64 high(u64 x)  { return (x >> 16) & 0xffff; }
+static u64 higha(u64 x) { return ((x + 0x8000) >> 16) & 0xffff; }
+
+static u64 prefix34(u64 x) {
+  return bits(x, 33, 16) | (bits(x, 15, 0) << 32);
+}
 
 // .plt is used only for lazy symbol resolution on PPC64. All PLT
 // calls are made via range extension thunks even if they are within
@@ -85,25 +112,26 @@ void write_plt_header(Context<E> &ctx, u8 *buf) {
     0x7c08'03a6, // mtlr    r0
 
     // Compute the PLT entry index
-    0xe80b'002c, // ld      r0, 44(r11)
-    0x7d8b'6050, // subf    r12, r11, r12
-    0x7d60'5a14, // add     r11, r0, r11
-    0x380c'ffcc, // addi    r0, r12, -52
+    0x398c'ffd4, // addi    r12, r12, -44
+    0x7c0b'6050, // subf    r0, r11, r12
     0x7800'f082, // rldicl  r0, r0, 62, 2
 
+    // Compute the address of .got.plt
+    0x3d6b'0000, // addis   r11, r11, GOTPLT_OFFSET@ha
+    0x396b'0000, // addi    r11, r11, GOTPLT_OFFSET@lo
+
     // Load .got.plt[0] and .got.plt[1] and branch to .got.plt[0]
     0xe98b'0000, // ld      r12, 0(r11)
     0x7d89'03a6, // mtctr   r12
     0xe96b'0008, // ld      r11, 8(r11)
     0x4e80'0420, // bctr
-
-    // .quad .got.plt - .plt - 8
-    0x0000'0000,
-    0x0000'0000,
   };
 
   memcpy(buf, insn, sizeof(insn));
-  *(ul64 *)(buf + 52) = ctx.gotplt->shdr.sh_addr - ctx.plt->shdr.sh_addr - 8;
+
+  i64 val = ctx.gotplt->shdr.sh_addr - ctx.plt->shdr.sh_addr - 8;
+  *(ul32 *)(buf + 28) |= higha(val);
+  *(ul32 *)(buf + 32) |= lo(val);
 }
 
 template <>
@@ -120,8 +148,8 @@ template <>
 void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {}
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -144,10 +172,11 @@ void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
 static u64 get_local_entry_offset(Context<E> &ctx, Symbol<E> &sym) {
   i64 val = sym.esym().ppc_local_entry;
   assert(val <= 7);
-  if (val == 0 || val == 1)
-    return 0;
   if (val == 7)
     Fatal(ctx) << sym << ": local entry offset 7 is reserved";
+
+  if (val == 0 || val == 1)
+    return 0;
   return 1 << val;
 }
 
@@ -168,13 +197,6 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    auto check = [&](i64 val, i64 lo, i64 hi) {
-      if (val < lo || hi <= val)
-        Error(ctx) << *this << ": relocation " << rel << " against "
-                   << sym << " out of range: " << val << " is not in ["
-                   << lo << ", " << hi << ")";
-    };
-
     u64 S = sym.get_addr(ctx);
     u64 A = rel.r_addend;
     u64 P = get_addr() + rel.r_offset;
@@ -182,12 +204,15 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     u64 GOT = ctx.got->shdr.sh_addr;
     u64 TOC = ctx.extra.TOC->value;
 
+    auto r2save_thunk_addr = [&] { return get_thunk_addr(i); };
+    auto no_r2save_thunk_addr = [&] { return get_thunk_addr(i) + 4; };
+
     switch (rel.r_type) {
     case R_PPC64_ADDR64:
       if (name() == ".toc")
-        apply_toc_rel(ctx, sym, rel, loc, S, A, P, dynrel);
+        apply_toc_rel(ctx, sym, rel, loc, S, A, P, &dynrel);
       else
-        apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+        apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
     case R_PPC64_TOC16_HA:
       *(ul16 *)loc = ha(S + A - TOC);
@@ -199,22 +224,37 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_PPC64_TOC16_LO_DS:
       *(ul16 *)loc |= (S + A - TOC) & 0xfffc;
       break;
-    case R_PPC64_REL24: {
-      i64 val = S + A - P + get_local_entry_offset(ctx, sym);
-      if (sym.has_plt(ctx) || sign_extend(val, 25) != val)
-        val = get_thunk_addr(i) + A - P;
+    case R_PPC64_REL24:
+      if (sym.has_plt(ctx) || !sym.esym().preserves_r2()) {
+        i64 val = r2save_thunk_addr() + A - P;
+        *(ul32 *)loc |= bits(val, 25, 2) << 2;
 
-      check(val, -(1 << 25), 1 << 25);
-      *(ul32 *)loc |= bits(val, 25, 2) << 2;
-
-      // If a callee is an external function, PLT saves %r2 to the
-      // caller's r2 save slot. We need to restore it after function
-      // return. To do so, there's usually a NOP as a placeholder
-      // after a BL. 0x6000'0000 is a NOP.
-      if (sym.has_plt(ctx) && *(ul32 *)(loc + 4) == 0x6000'0000)
-        *(ul32 *)(loc + 4) = 0xe841'0018; // ld r2, 24(r1)
+        // The thunk saves %r2 to the caller's r2 save slot. We need to
+        // restore it after function return. To do so, there's usually a
+        // NOP as a placeholder after a BL. 0x6000'0000 is a NOP.
+        if (*(ul32 *)(loc + 4) == 0x6000'0000)
+          *(ul32 *)(loc + 4) = 0xe841'0018; // ld r2, 24(r1)
+      } else {
+        i64 val = S + get_local_entry_offset(ctx, sym) + A - P;
+        if (sign_extend(val, 25) != val)
+          val = no_r2save_thunk_addr() + A - P;
+        *(ul32 *)loc |= bits(val, 25, 2) << 2;
+      }
+      break;
+    case R_PPC64_REL24_NOTOC:
+      if (sym.has_plt(ctx) || sym.esym().uses_toc()) {
+        i64 val = no_r2save_thunk_addr() + A - P;
+        *(ul32 *)loc |= bits(val, 25, 2) << 2;
+      } else {
+        i64 val = S + A - P;
+        if (sign_extend(val, 25) != val)
+          val = no_r2save_thunk_addr() + A - P;
+        *(ul32 *)loc |= bits(val, 25, 2) << 2;
+      }
+      break;
+    case R_PPC64_REL32:
+      *(ul32 *)loc = S + A - P;
       break;
-    }
     case R_PPC64_REL64:
       *(ul64 *)loc = S + A - P;
       break;
@@ -236,38 +276,60 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_PPC64_PLT16_LO_DS:
       *(ul16 *)loc |= (G + GOT - TOC) & 0xfffc;
       break;
+    case R_PPC64_PLT_PCREL34:
+    case R_PPC64_PLT_PCREL34_NOTOC:
+    case R_PPC64_GOT_PCREL34:
+      *(ul64 *)loc |= prefix34(G + GOT - P);
+      break;
+    case R_PPC64_PCREL34:
+      *(ul64 *)loc |= prefix34(S + A - P);
+      break;
     case R_PPC64_GOT_TPREL16_HA:
       *(ul16 *)loc = ha(sym.get_gottp_addr(ctx) - TOC);
       break;
+    case R_PPC64_GOT_TPREL16_LO_DS:
+      *(ul16 *)loc |= (sym.get_gottp_addr(ctx) - TOC) & 0xfffc;
+      break;
+    case R_PPC64_GOT_TPREL_PCREL34:
+      *(ul64 *)loc |= prefix34(sym.get_gottp_addr(ctx) - P);
+      break;
     case R_PPC64_GOT_TLSGD16_HA:
       *(ul16 *)loc = ha(sym.get_tlsgd_addr(ctx) - TOC);
       break;
     case R_PPC64_GOT_TLSGD16_LO:
       *(ul16 *)loc = lo(sym.get_tlsgd_addr(ctx) - TOC);
       break;
+    case R_PPC64_GOT_TLSGD_PCREL34:
+      *(ul64 *)loc |= prefix34(sym.get_tlsgd_addr(ctx) - P);
+      break;
     case R_PPC64_GOT_TLSLD16_HA:
       *(ul16 *)loc = ha(ctx.got->get_tlsld_addr(ctx) - TOC);
       break;
     case R_PPC64_GOT_TLSLD16_LO:
       *(ul16 *)loc = lo(ctx.got->get_tlsld_addr(ctx) - TOC);
       break;
+    case R_PPC64_GOT_TLSLD_PCREL34:
+      *(ul64 *)loc |= prefix34(ctx.got->get_tlsld_addr(ctx) - P);
+      break;
     case R_PPC64_DTPREL16_HA:
       *(ul16 *)loc = ha(S + A - ctx.dtp_addr);
       break;
-    case R_PPC64_TPREL16_HA:
-      *(ul16 *)loc = ha(S + A - ctx.tp_addr);
-      break;
     case R_PPC64_DTPREL16_LO:
       *(ul16 *)loc = lo(S + A - ctx.dtp_addr);
       break;
+    case R_PPC64_DTPREL34:
+      *(ul64 *)loc |= prefix34(S + A - ctx.dtp_addr);
+      break;
+    case R_PPC64_TPREL16_HA:
+      *(ul16 *)loc = ha(S + A - ctx.tp_addr);
+      break;
     case R_PPC64_TPREL16_LO:
       *(ul16 *)loc = lo(S + A - ctx.tp_addr);
       break;
-    case R_PPC64_GOT_TPREL16_LO_DS:
-      *(ul16 *)loc |= (sym.get_gottp_addr(ctx) - TOC) & 0xfffc;
-      break;
     case R_PPC64_PLTSEQ:
+    case R_PPC64_PLTSEQ_NOTOC:
     case R_PPC64_PLTCALL:
+    case R_PPC64_PLTCALL_NOTOC:
     case R_PPC64_TLS:
     case R_PPC64_TLSGD:
     case R_PPC64_TLSLD:
@@ -284,17 +346,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     auto check = [&](i64 val, i64 lo, i64 hi) {
       if (val < lo || hi <= val)
         Error(ctx) << *this << ": relocation " << rel << " against "
@@ -342,18 +399,13 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
 
     switch (rel.r_type) {
     case R_PPC64_ADDR64:
@@ -363,21 +415,37 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
         scan_dyn_absrel(ctx, sym, rel);
       break;
     case R_PPC64_GOT_TPREL16_HA:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+    case R_PPC64_GOT_TPREL_PCREL34:
+      sym.flags |= NEEDS_GOTTP;
       break;
     case R_PPC64_REL24:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
+      break;
+    case R_PPC64_REL24_NOTOC:
+      if (sym.is_imported)
+        sym.flags |= NEEDS_PLT;
+      ctx.extra.is_power10 = true;
       break;
     case R_PPC64_PLT16_HA:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+    case R_PPC64_PLT_PCREL34:
+    case R_PPC64_PLT_PCREL34_NOTOC:
+    case R_PPC64_GOT_PCREL34:
+      sym.flags |= NEEDS_GOT;
       break;
     case R_PPC64_GOT_TLSGD16_HA:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+    case R_PPC64_GOT_TLSGD_PCREL34:
+      sym.flags |= NEEDS_TLSGD;
       break;
     case R_PPC64_GOT_TLSLD16_HA:
-      ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+    case R_PPC64_GOT_TLSLD_PCREL34:
+      ctx.needs_tlsld = true;
       break;
+    case R_PPC64_TPREL16_HA:
+    case R_PPC64_TPREL16_LO:
+      check_tlsle(ctx, sym, rel);
+      break;
+    case R_PPC64_REL32:
     case R_PPC64_REL64:
     case R_PPC64_TOC16_HA:
     case R_PPC64_TOC16_LO:
@@ -388,10 +456,11 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_PPC64_PLT16_HI:
     case R_PPC64_PLT16_LO:
     case R_PPC64_PLT16_LO_DS:
+    case R_PPC64_PCREL34:
     case R_PPC64_PLTSEQ:
+    case R_PPC64_PLTSEQ_NOTOC:
     case R_PPC64_PLTCALL:
-    case R_PPC64_TPREL16_HA:
-    case R_PPC64_TPREL16_LO:
+    case R_PPC64_PLTCALL_NOTOC:
     case R_PPC64_GOT_TPREL16_LO_DS:
     case R_PPC64_GOT_TLSGD16_LO:
     case R_PPC64_GOT_TLSLD16_LO:
@@ -400,6 +469,7 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_PPC64_TLSLD:
     case R_PPC64_DTPREL16_HA:
     case R_PPC64_DTPREL16_LO:
+    case R_PPC64_DTPREL34:
       break;
     default:
       Error(ctx) << *this << ": unknown relocation: " << rel;
@@ -409,53 +479,84 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
 template <>
 void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
-  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
-
   // If the destination is PLT, we read an address from .got.plt or .got
   // and jump there.
   static const ul32 plt_thunk[] = {
-    // Save r2 to the r2 save slot reserved in the caller's stack frame
     0xf841'0018, // std   r2, 24(r1)
-
-    // Jump to a PLT entry
     0x3d82'0000, // addis r12, r2, foo@gotplt@toc@ha
     0xe98c'0000, // ld    r12, foo@gotplt@toc@lo(r12)
     0x7d89'03a6, // mtctr r12
     0x4e80'0420, // bctr
   };
 
+  static const ul32 plt_thunk_power10[] = {
+    0xf841'0018, // std   r2, 24(r1)
+    0x0410'0000, // pld   r12, foo@gotplt@pcrel
+    0xe580'0000,
+    0x7d89'03a6, // mtctr r12
+    0x4e80'0420, // bctr
+  };
+
   // If the destination is a non-imported function, we directly jump
   // to its local entry point.
   static const ul32 local_thunk[] = {
-    // Jump to a local entry point
+    0xf841'0018, // std   r2, 24(r1)
     0x3d82'0000, // addis r12, r2,  foo@toc@ha
     0x398c'0000, // addi  r12, r12, foo@toc@lo
     0x7d89'03a6, // mtctr r12
     0x4e80'0420, // bctr
-    0x6000'0000, // nop
+  };
+
+  static const ul32 local_thunk_power10[] = {
+    0xf841'0018, // std   r2, 24(r1)
+    0x0610'0000, // pla   r12, foo@pcrel
+    0x3980'0000,
+    0x7d89'03a6, // mtctr r12
+    0x4e80'0420, // bctr
   };
 
   static_assert(E::thunk_size == sizeof(plt_thunk));
+  static_assert(E::thunk_size == sizeof(plt_thunk_power10));
   static_assert(E::thunk_size == sizeof(local_thunk));
+  static_assert(E::thunk_size == sizeof(local_thunk_power10));
 
-  for (i64 i = 0; i < symbols.size(); i++) {
-    Symbol<E> &sym = *symbols[i];
-    ul32 *loc = (ul32 *)(buf + i * E::thunk_size);
+  u8 *buf = ctx.buf + output_section.shdr.sh_offset + offset;
+  u64 P = output_section.shdr.sh_addr + offset;
 
-    if (sym.has_plt(ctx)) {
-      memcpy(loc, plt_thunk, sizeof(plt_thunk));
-      u64 got = sym.has_got(ctx) ? sym.get_got_addr(ctx) : sym.get_gotplt_addr(ctx);
-      i64 val = got - ctx.extra.TOC->value;
-      loc[1] |= higha(val);
-      loc[2] |= lo(val);
+  for (Symbol<E> *sym : symbols) {
+    if (sym->has_plt(ctx)) {
+      u64 got =
+        sym->has_got(ctx) ? sym->get_got_addr(ctx) : sym->get_gotplt_addr(ctx);
+
+      if (ctx.extra.is_power10) {
+        memcpy(buf, plt_thunk_power10, E::thunk_size);
+        *(ul64 *)(buf + 1) |= prefix34(got - P - 4);
+      } else {
+        i64 val = got - ctx.extra.TOC->value;
+        memcpy(buf, plt_thunk, E::thunk_size);
+        *(ul32 *)(buf + 4) |= higha(val);
+        *(ul32 *)(buf + 8) |= lo(val);
+      }
     } else {
-      memcpy(loc, local_thunk, sizeof(local_thunk));
-      i64 val = sym.get_addr(ctx) + get_local_entry_offset(ctx, sym) -
-                ctx.extra.TOC->value;
-      loc[0] |= higha(val);
-      loc[1] |= lo(val);
+      if (ctx.extra.is_power10) {
+        memcpy(buf, local_thunk_power10, E::thunk_size);
+        *(ul64 *)(buf + 1) |= prefix34(sym->get_addr(ctx) - P - 4);
+      } else {
+        i64 val = sym->get_addr(ctx) - ctx.extra.TOC->value;
+        memcpy(buf, local_thunk, E::thunk_size);
+        *(ul32 *)(buf + 4) |= higha(val);
+        *(ul32 *)(buf + 8) |= lo(val);
+      }
     }
+
+    buf += E::thunk_size;
+    P += E::thunk_size;
   }
 }
 
+template <>
+u64 get_eflags(Context<E> &ctx) {
+  return 2;
+}
+
 } // namespace mold::elf

elf/arch-s390x.cc

@@ -12,7 +12,7 @@
 // As an instruction set, s390x isn't particularly odd. It has 16 general-
 // purpose registers. Instructions are 2, 4 or 6 bytes long and always
 // aligned to 2 bytes boundaries. Despite unfamiliarty, I found that it
-// just feels like a 64-bit i386 in a parallel universe.
+// just feels like an x86-64 in a parallel universe.
 //
 // Here is the register usage in this ABI:
 //
@@ -25,15 +25,15 @@
 //   a1:    upper 32 bits of TP (thread pointer)
 //   a2:    lower 32 bits of TP (thread pointer)
 //
-// TLS is supported on s390x in the same way as it is on other targets
-// with one exeption. On other targets, __tls_get_addr is used to get an
-// address of a thread-local variable. On s390x, __tls_get_offset is used
-// instead. The difference is __tls_get_offset returns an address of a
-// thread-local variable as an offset from TP. So we need to add TP to a
-// return value before use. I don't know why it is different, but that is
-// the way it is.
+// Thread-local storage (TLS) is supported on s390x in the same way as it
+// is on other targets with one exeption. On other targets, __tls_get_addr
+// is used to get an address of a thread-local variable. On s390x,
+// __tls_get_offset is used instead. The difference is __tls_get_offset
+// returns an address of a thread-local variable as an offset from TP. So
+// we need to add TP to a return value before use. I don't know why it is
+// different, but that is the way it is.
 //
-// https://github.com/IBM/s390x-abi/releases/download/v1.6/lzsabi_s390x.pdf
+// https://github.com/rui314/psabi/blob/main/s390x.pdf
 
 #include "mold.h"
 
@@ -48,16 +48,23 @@ static void write_mid20(u8 *loc, u64 val) {
 template <>
 void write_plt_header(Context<E> &ctx, u8 *buf) {
   static u8 insn[] = {
+    // Compute PLT_INDEX
+    0xb9, 0x09, 0x00, 0x01,             // sgr   %r0, %r1
+    0xa7, 0x0b, 0xff, 0xc2,             // aghi  %r0, -62
+    0xeb, 0x10, 0x00, 0x01, 0x00, 0x0c, // srlg  %r1, %r0, 1
+    0xb9, 0x08, 0x00, 0x01,             // agr   %r0, %r1
     0xe3, 0x00, 0xf0, 0x38, 0x00, 0x24, // stg   %r0, 56(%r15)
+    // Branch to _dl_runtime_resolve
     0xc0, 0x10, 0, 0, 0, 0,             // larl  %r1, GOTPLT_OFFSET
     0xd2, 0x07, 0xf0, 0x30, 0x10, 0x08, // mvc   48(8, %r15), 8(%r1)
     0xe3, 0x10, 0x10, 0x10, 0x00, 0x04, // lg    %r1, 16(%r1)
     0x07, 0xf1,                         // br    %r1
-    0x07, 0x00, 0x07, 0x00, 0x07, 0x00, // nopr; nopr; nopr
+    0x07, 0x00, 0x07, 0x00,             // nopr; nopr
   };
 
   memcpy(buf, insn, sizeof(insn));
-  *(ub32 *)(buf + 8) = (ctx.gotplt->shdr.sh_addr - ctx.plt->shdr.sh_addr - 6) >> 1;
+  *(ub32 *)(buf + 26) =
+    (ctx.gotplt->shdr.sh_addr - ctx.plt->shdr.sh_addr - 24) >> 1;
 }
 
 template <>
@@ -65,15 +72,12 @@ void write_plt_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
   static u8 insn[] = {
     0xc0, 0x10, 0, 0, 0, 0,             // larl  %r1, GOTPLT_ENTRY_OFFSET
     0xe3, 0x10, 0x10, 0x00, 0x00, 0x04, // lg    %r1, (%r1)
-    0xc0, 0x01, 0, 0, 0, 0,             // lgfi  %r0, PLT_INDEX
-    0x07, 0xf1,                         // br    %r1
-    0x07, 0x00, 0x07, 0x00, 0x07, 0x00, // nopr; nopr; nopr
-    0x07, 0x00, 0x07, 0x00, 0x07, 0x00, // nopr; nopr; nopr
+    0x0d, 0x01,                         // basr  %r0, %r1
+    0x07, 0x00,                         // nopr
   };
 
   memcpy(buf, insn, sizeof(insn));
   *(ub32 *)(buf + 2) = (sym.get_gotplt_addr(ctx) - sym.get_plt_addr(ctx)) >> 1;
-  *(ub32 *)(buf + 14) = sym.get_plt_idx(ctx) * sizeof(ElfRel<E>);
 }
 
 template <>
@@ -90,8 +94,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 }
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -149,136 +153,97 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
 
     switch (rel.r_type) {
     case R_390_64:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
-    case R_390_8: {
-      i64 val = S + A;
-      check(val, 0, 1 << 8);
-      *loc = val;
+    case R_390_8:
+      check(S + A, 0, 1 << 8);
+      *loc = S + A;
       break;
-    }
-    case R_390_12: {
-      i64 val = S + A;
-      check(val, 0, 1 << 12);
-      *(ul16 *)loc = bits(val, 11, 0);
+    case R_390_12:
+      check(S + A, 0, 1 << 12);
+      *(ul16 *)loc |= bits(S + A, 11, 0);
       break;
-    }
-    case R_390_16: {
-      i64 val = S + A;
-      check(val, 0, 1 << 16);
-      *(ub16 *)loc = val;
+    case R_390_16:
+      check(S + A, 0, 1 << 16);
+      *(ub16 *)loc = S + A;
       break;
-    }
-    case R_390_20: {
-      i64 val = S + A;
-      check(val, 0, 1 << 20);
-      write_mid20(loc, val);
+    case R_390_20:
+      check(S + A, 0, 1 << 20);
+      write_mid20(loc, S + A);
       break;
-    }
     case R_390_32:
-    case R_390_PLT32: {
-      i64 val = S + A;
-      check(val, 0, 1LL << 32);
-      *(ub32 *)loc = val;
+    case R_390_PLT32:
+      check(S + A, 0, 1LL << 32);
+      *(ub32 *)loc = S + A;
       break;
-    }
     case R_390_PLT64:
       *(ub64 *)loc = S + A;
       break;
     case R_390_PC12DBL:
-    case R_390_PLT12DBL: {
-      i64 val = S + A - P;
-      check_dbl(val, -(1 << 12), 1 << 12);
-      *(ul16 *)loc = (val >> 1) & 0x0fff;
+    case R_390_PLT12DBL:
+      check_dbl(S + A - P, -(1 << 12), 1 << 12);
+      *(ul16 *)loc |= bits(S + A - P, 12, 1);
       break;
-    }
-    case R_390_PC16: {
-      i64 val = S + A - P;
-      check(val, -(1 << 15), 1 << 15);
-      *(ub16 *)loc = val;
+    case R_390_PC16:
+      check(S + A - P, -(1 << 15), 1 << 15);
+      *(ub16 *)loc = S + A - P;
       break;
-    }
-    case R_390_PC32: {
-      i64 val = S + A - P;
-      check(val, -(1LL << 31), 1LL << 31);
-      *(ub32 *)loc = val;
+    case R_390_PC32:
+      check(S + A - P, -(1LL << 31), 1LL << 31);
+      *(ub32 *)loc = S + A - P;
       break;
-    }
     case R_390_PC64:
       *(ub64 *)loc = S + A - P;
       break;
     case R_390_PC16DBL:
-    case R_390_PLT16DBL: {
-      i64 val = S + A - P;
-      check_dbl(val, -(1 << 16), 1 << 16);
-      *(ub16 *)loc = val >> 1;
+    case R_390_PLT16DBL:
+      check_dbl(S + A - P, -(1 << 16), 1 << 16);
+      *(ub16 *)loc = (S + A - P) >> 1;
       break;
-    }
     case R_390_PC24DBL:
-    case R_390_PLT24DBL: {
-      i64 val = S + A - P;
-      check_dbl(val, -(1 << 24), 1 << 24);
-      *(ub32 *)loc |= bits(val, 24, 1);
+    case R_390_PLT24DBL:
+      check_dbl(S + A - P, -(1 << 24), 1 << 24);
+      *(ub32 *)loc |= bits(S + A - P, 24, 1);
       break;
-    }
     case R_390_PC32DBL:
     case R_390_PLT32DBL:
-      if (ctx.is_static && &sym == ctx.tls_get_offset) {
-        // __tls_get_offset() in libc.a is stub code that calls abort().
-        // So we provide a replacement function.
-        *(ub32 *)loc = (ctx.extra.tls_get_offset->shdr.sh_addr - P) >> 1;
-      } else {
-        i64 val = S + A - P;
-        check_dbl(val, -(1LL << 32), 1LL << 32);
-        *(ub32 *)loc = val >> 1;
-      }
+      check_dbl(S + A - P, -(1LL << 32), 1LL << 32);
+      *(ub32 *)loc = (S + A - P) >> 1;
       break;
     case R_390_GOT12:
-    case R_390_GOTPLT12: {
-      i64 val = G + A;
-      check(val, 0, 1 << 12);
-      *(ul16 *)loc = bits(val, 11, 0);
+    case R_390_GOTPLT12:
+      check(G + A, 0, 1 << 12);
+      *(ul16 *)loc |= bits(G + A, 11, 0);
       break;
-    }
     case R_390_GOT16:
-    case R_390_GOTPLT16: {
-      i64 val = G + A;
-      check(val, 0, 1 << 16);
-      *(ub16 *)loc = val;
+    case R_390_GOTPLT16:
+      check(G + A, 0, 1 << 16);
+      *(ub16 *)loc = G + A;
       break;
-    }
     case R_390_GOT20:
-    case R_390_GOTPLT20: {
-      i64 val = G + A;
-      check(val, 0, 1 << 20);
-      write_mid20(loc, val);
+    case R_390_GOTPLT20:
+      check(G + A, 0, 1 << 20);
+      write_mid20(loc, G + A);
       break;
-    }
     case R_390_GOT32:
-    case R_390_GOTPLT32: {
-      i64 val = G + A;
-      check(val, 0, 1LL << 32);
-      *(ub32 *)loc = val;
+    case R_390_GOTPLT32:
+      check(G + A, 0, 1LL << 32);
+      *(ub32 *)loc = G + A;
       break;
-    }
     case R_390_GOT64:
     case R_390_GOTPLT64:
       *(ub64 *)loc = G + A;
       break;
     case R_390_GOTOFF16:
-    case R_390_PLTOFF16: {
-      i64 val = S + A - GOT;
-      check(val, -(1 << 15), 1 << 15);
-      *(ub16 *)loc = val;
+    case R_390_PLTOFF16:
+      check(S + A - GOT, -(1 << 15), 1 << 15);
+      *(ub16 *)loc = S + A - GOT;
       break;
-    }
     case R_390_GOTOFF32:
-    case R_390_PLTOFF32: {
-      i64 val = S + A - GOT;
-      check(val, -(1LL << 31), 1LL << 31);
-      *(ub32 *)loc = val;
+    case R_390_PLTOFF32:
+      check(S + A - GOT, -(1LL << 31), 1LL << 31);
+      *(ub32 *)loc = S + A - GOT;
       break;
-    }
     case R_390_GOTOFF64:
     case R_390_PLTOFF64:
       *(ub64 *)loc = S + A - GOT;
@@ -286,18 +251,14 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_390_GOTPC:
       *(ub64 *)loc = GOT + A - P;
       break;
-    case R_390_GOTPCDBL: {
-      i64 val = GOT + A - P;
-      check_dbl(val, -(1LL << 32), 1LL << 32);
-      *(ub32 *)loc = val >> 1;
+    case R_390_GOTPCDBL:
+      check_dbl(GOT + A - P, -(1LL << 32), 1LL << 32);
+      *(ub32 *)loc = (GOT + A - P) >> 1;
       break;
-    }
-    case R_390_GOTENT: {
-      i64 val = GOT + G + A - P;
-      check(val, -(1LL << 32), 1LL << 32);
-      *(ub32 *)loc = val >> 1;
+    case R_390_GOTENT:
+      check(GOT + G + A - P, -(1LL << 32), 1LL << 32);
+      *(ub32 *)loc = (GOT + G + A - P) >> 1;
       break;
-    }
     case R_390_TLS_LE32:
       *(ub32 *)loc = S + A - ctx.tp_addr;
       break;
@@ -313,19 +274,30 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_390_TLS_GD32:
       if (sym.has_tlsgd(ctx))
         *(ub32 *)loc = sym.get_tlsgd_addr(ctx) + A - GOT;
+      else if (sym.has_gottp(ctx))
+        *(ub32 *)loc = sym.get_gottp_addr(ctx) + A - GOT;
       else
         *(ub32 *)loc = S + A - ctx.tp_addr;
       break;
     case R_390_TLS_GD64:
       if (sym.has_tlsgd(ctx))
         *(ub64 *)loc = sym.get_tlsgd_addr(ctx) + A - GOT;
+      else if (sym.has_gottp(ctx))
+        *(ub64 *)loc = sym.get_gottp_addr(ctx) + A - GOT;
       else
         *(ub64 *)loc = S + A - ctx.tp_addr;
       break;
     case R_390_TLS_GDCALL:
-      if (!sym.has_tlsgd(ctx)) {
-        static u8 nop[] = { 0xc0, 0x04, 0x00, 0x00, 0x00, 0x00 };
-        memcpy(loc, nop, sizeof(nop));
+      if (sym.has_tlsgd(ctx)) {
+        // do nothing
+      } else if (sym.has_gottp(ctx)) {
+        // lg %r2, 0(%r2, %r12)
+        static u8 insn[] = { 0xe3, 0x22, 0xc0, 0x00, 0x00, 0x04 };
+        memcpy(loc, insn, sizeof(insn));
+      } else {
+        // nop
+        static u8 insn[] = { 0xc0, 0x04, 0x00, 0x00, 0x00, 0x00 };
+        memcpy(loc, insn, sizeof(insn));
       }
       break;
     case R_390_TLS_LDM32:
@@ -350,8 +322,9 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       break;
     case R_390_TLS_LDCALL:
       if (!ctx.got->has_tlsld(ctx)) {
-        static u8 nop[] = { 0xc0, 0x04, 0x00, 0x00, 0x00, 0x00 };
-        memcpy(loc, nop, sizeof(nop));
+        // nop
+        static u8 insn[] = { 0xc0, 0x04, 0x00, 0x00, 0x00, 0x00 };
+        memcpy(loc, insn, sizeof(insn));
       }
       break;
     default:
@@ -366,17 +339,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     auto check = [&](i64 val, i64 lo, i64 hi) {
       if (val < lo || hi <= val)
         Error(ctx) << *this << ": relocation " << rel << " against "
@@ -392,12 +360,10 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
     u64 A = frag ? frag_addend : (i64)rel.r_addend;
 
     switch (rel.r_type) {
-    case R_390_32: {
-      i64 val = S + A;
-      check(val, 0, 1LL << 32);
-      *(ub32 *)loc = val;
+    case R_390_32:
+      check(S + A, 0, 1LL << 32);
+      *(ub32 *)loc = S + A;
       break;
-    }
     case R_390_64:
       if (std::optional<u64> val = get_tombstone(sym, frag))
         *(ub64 *)loc = *val;
@@ -426,18 +392,13 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
 
     switch (rel.r_type) {
     case R_390_64:
@@ -473,7 +434,7 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_390_GOTPC:
     case R_390_GOTPCDBL:
     case R_390_GOTENT:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
     case R_390_PLT12DBL:
     case R_390_PLT16DBL:
@@ -485,57 +446,47 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_390_PLTOFF32:
     case R_390_PLTOFF64:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_390_TLS_GOTIE20:
     case R_390_TLS_IEENT:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
       break;
     case R_390_TLS_GD32:
     case R_390_TLS_GD64:
-      if (!relax_tlsgd(ctx, sym))
-        sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      // We always want to relax calls to __tls_get_offset() in statically-
+      // linked executables because __tls_get_offset() in libc.a just calls
+      // abort().
+      if ((ctx.arg.relax && sym.is_tprel_linktime_const(ctx)) ||
+          ctx.arg.is_static) {
+        // Do nothing
+      } else if (ctx.arg.relax && sym.is_tprel_runtime_const(ctx)) {
+        sym.flags |= NEEDS_GOTTP;
+      } else {
+        sym.flags |= NEEDS_TLSGD;
+      }
       break;
     case R_390_TLS_LDM32:
     case R_390_TLS_LDM64:
-      if (!relax_tlsld(ctx))
-        ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+      if (ctx.arg.is_static || (ctx.arg.relax && !ctx.arg.shared)) {
+        // Do nothing
+      } else {
+        ctx.needs_tlsld = true;
+      }
       break;
     case R_390_TLS_LE32:
     case R_390_TLS_LE64:
+      check_tlsle(ctx, sym, rel);
+      break;
     case R_390_TLS_LDO32:
     case R_390_TLS_LDO64:
     case R_390_TLS_GDCALL:
     case R_390_TLS_LDCALL:
       break;
     default:
-      Fatal(ctx) << *this << ": scan_relocations: " << rel;
+      Error(ctx) << *this << ": unknown relocation: " << rel;
     }
   }
 }
 
-// __tls_get_offset() in libc.a just calls abort(), assuming that the
-// linker always relaxes TLS calls for statically-linkd executables.
-// We don't always do that because we believe --relax and --static
-// should be orthogonal.
-//
-// This section provides a replacement for __tls_get_offset() in libc.a.
-void S390XTlsGetOffsetSection::copy_buf(Context<E> &ctx) {
-  static const u8 insn[] = {
-    0xc0, 0x10, 0, 0, 0, 0,             // larl %r1, GOT
-    0xb9, 0x08, 0x00, 0x21,             // agr  %r2, %r1
-    0xe3, 0x20, 0x20, 0x08, 0x00, 0x04, // lg   %r2, 8(%r2)
-    0xc0, 0x11, 0, 0, 0, 0,             // lgfi %r1, TLS_BLOCK_SIZE
-    0xb9, 0x09, 0x00, 0x21,             // sgr  %r2, %r1
-    0x07, 0xfe,                         // br   %r14
-  };
-
-  assert(this->shdr.sh_size == sizeof(insn));
-
-  u8 *loc = ctx.buf + this->shdr.sh_offset;
-  memcpy(loc, insn, sizeof(insn));
-  *(ub32 *)(loc + 2) = (ctx.got->shdr.sh_addr - this->shdr.sh_addr) >> 1;
-  *(ub32 *)(loc + 18) = ctx.tp_addr - ctx.tls_begin;
-}
-
 } // namespace mold::elf

elf/arch-sh4.cc

@@ -26,6 +26,13 @@
 //    small to contain large immediates. On SH-4, large immediates are
 //    loaded from memory using `mov.l` PC-relative load instruction.
 //
+//  - Many RISC ISAs are, despite their name, actually fairly complex.
+//    They tend to have hundreds if not thousands of different instructions.
+//    SH-4 doesn't really have that many instructions because its 16-bit
+//    machine code simply can't encode many different opcodes. As a
+//    result, the number of relocations the linker has to support is also
+//    small.
+//
 // Beside these, SH-4 has a delay branch slot just like contemporary MIPS
 // and SPARC. That is, one instruction after a branch instruction will
 // always be executed even if the branch is taken. Delay branch slot allows
@@ -179,8 +186,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 }
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -222,7 +229,7 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
 
     switch (rel.r_type) {
     case R_SH_DIR32:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
     case R_SH_REL32:
     case R_SH_PLT32:
@@ -264,17 +271,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     SectionFragment<E> *frag;
     i64 frag_addend;
     std::tie(frag, frag_addend) = get_fragment(ctx, rel);
@@ -305,16 +307,11 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
       Error(ctx) << sym << ": GNU ifunc symbol is not supported on sh4";
 
@@ -326,25 +323,27 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
       scan_pcrel(ctx, sym, rel);
       break;
     case R_SH_GOT32:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
     case R_SH_PLT32:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_SH_TLS_GD_32:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_TLSGD;
       break;
     case R_SH_TLS_LD_32:
-      ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+      ctx.needs_tlsld = true;
       break;
     case R_SH_TLS_IE_32:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
+      break;
+    case R_SH_TLS_LE_32:
+      check_tlsle(ctx, sym, rel);
       break;
     case R_SH_GOTPC:
     case R_SH_GOTOFF:
     case R_SH_TLS_LDO_32:
-    case R_SH_TLS_LE_32:
       break;
     default:
       Fatal(ctx) << *this << ": unknown relocation: " << rel;

elf/arch-sparc64.cc

@@ -54,8 +54,7 @@
 // This scheme is very similar to i386. That may not be a coincidence
 // because the i386 ELF psABI is created by Sun Microsystems too.
 //
-// https://docs.oracle.com/cd/E36784_01/html/E36857/chapter6-62988.html
-// https://docs.oracle.com/cd/E19120-01/open.solaris/819-0690/chapter8-40/index.html
+// https://github.com/rui314/psabi/blob/main/sparc.pdf
 
 #include "mold.h"
 
@@ -120,8 +119,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 }
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -171,99 +170,73 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
 
     switch (rel.r_type) {
     case R_SPARC_64:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
-    case R_SPARC_5: {
-      i64 val = S + A;
-      check(val, 0, 1 << 5);
-      *(ub32 *)loc |= bits(val, 4, 0);
+    case R_SPARC_5:
+      check(S + A, 0, 1 << 5);
+      *(ub32 *)loc |= bits(S + A, 4, 0);
       break;
-    }
-    case R_SPARC_6: {
-      i64 val = S + A;
-      check(val, 0, 1 << 6);
-      *(ub32 *)loc |= bits(val, 5, 0);
+    case R_SPARC_6:
+      check(S + A, 0, 1 << 6);
+      *(ub32 *)loc |= bits(S + A, 5, 0);
       break;
-    }
-    case R_SPARC_7: {
-      i64 val = S + A;
-      check(val, 0, 1 << 7);
-      *(ub32 *)loc |= bits(val, 6, 0);
+    case R_SPARC_7:
+      check(S + A, 0, 1 << 7);
+      *(ub32 *)loc |= bits(S + A, 6, 0);
       break;
-    }
-    case R_SPARC_8: {
-      i64 val = S + A;
-      check(val, 0, 1 << 8);
-      *(u8 *)loc = val;
+    case R_SPARC_8:
+      check(S + A, 0, 1 << 8);
+      *(u8 *)loc = S + A;
       break;
-    }
-    case R_SPARC_10: {
-      i64 val = S + A;
-      check(val, 0, 1 << 10);
-      *(ub32 *)loc |= bits(val, 9, 0);
+    case R_SPARC_10:
+      check(S + A, 0, 1 << 10);
+      *(ub32 *)loc |= bits(S + A, 9, 0);
       break;
-    }
     case R_SPARC_LO10:
     case R_SPARC_LOPLT10:
       *(ub32 *)loc |= bits(S + A, 9, 0);
       break;
-    case R_SPARC_11: {
-      i64 val = S + A;
-      check(val, 0, 1 << 11);
-      *(ub32 *)loc |= bits(val, 10, 0);
+    case R_SPARC_11:
+      check(S + A, 0, 1 << 11);
+      *(ub32 *)loc |= bits(S + A, 10, 0);
       break;
-    }
-    case R_SPARC_13: {
-      i64 val = S + A;
-      check(val, 0, 1 << 13);
-      *(ub32 *)loc |= bits(val, 12, 0);
+    case R_SPARC_13:
+      check(S + A, 0, 1 << 13);
+      *(ub32 *)loc |= bits(S + A, 12, 0);
       break;
-    }
     case R_SPARC_16:
-    case R_SPARC_UA16: {
-      i64 val = S + A;
-      check(val, 0, 1 << 16);
-      *(ub16 *)loc = val;
+    case R_SPARC_UA16:
+      check(S + A, 0, 1 << 16);
+      *(ub16 *)loc = S + A;
       break;
-    }
-    case R_SPARC_22: {
-      i64 val = S + A;
-      check(val, 0, 1 << 22);
-      *(ub32 *)loc |= bits(val, 21, 0);
+    case R_SPARC_22:
+      check(S + A, 0, 1 << 22);
+      *(ub32 *)loc |= bits(S + A, 21, 0);
       break;
-    }
     case R_SPARC_32:
     case R_SPARC_UA32:
-    case R_SPARC_PLT32: {
-      i64 val = S + A;
-      check(val, 0, 1LL << 32);
-      *(ub32 *)loc = val;
+    case R_SPARC_PLT32:
+      check(S + A, 0, 1LL << 32);
+      *(ub32 *)loc = S + A;
       break;
-    }
     case R_SPARC_PLT64:
     case R_SPARC_UA64:
     case R_SPARC_REGISTER:
       *(ub64 *)loc = S + A;
       break;
-    case R_SPARC_DISP8: {
-      i64 val = S + A - P;
-      check(val, -(1 << 7), 1 << 7);
-      *(u8 *)loc = val;
+    case R_SPARC_DISP8:
+      check(S + A - P, -(1 << 7), 1 << 7);
+      *(u8 *)loc = S + A - P;
       break;
-    }
-    case R_SPARC_DISP16: {
-      i64 val = S + A - P;
-      check(val, -(1 << 15), 1 << 15);
-      *(ub16 *)loc = val;
+    case R_SPARC_DISP16:
+      check(S + A - P, -(1 << 15), 1 << 15);
+      *(ub16 *)loc = S + A - P;
       break;
-    }
     case R_SPARC_DISP32:
-    case R_SPARC_PCPLT32: {
-      i64 val = S + A - P;
-      check(val, -(1LL << 31), 1LL << 31);
-      *(ub32 *)loc = val;
+    case R_SPARC_PCPLT32:
+      check(S + A - P, -(1LL << 31), 1LL << 31);
+      *(ub32 *)loc = S + A - P;
       break;
-    }
     case R_SPARC_DISP64:
       *(ub64 *)loc = S + A - P;
       break;
@@ -273,25 +246,19 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ub16 *)loc |= (bit(val, 16) << 21) | bits(val, 15, 2);
       break;
     }
-    case R_SPARC_WDISP19: {
-      i64 val = S + A - P;
-      check(val, -(1 << 20), 1 << 20);
-      *(ub32 *)loc |= bits(val, 20, 2);
+    case R_SPARC_WDISP19:
+      check(S + A - P, -(1 << 20), 1 << 20);
+      *(ub32 *)loc |= bits(S + A - P, 20, 2);
       break;
-    }
-    case R_SPARC_WDISP22: {
-      i64 val = S + A - P;
-      check(val, -(1 << 23), 1 << 23);
-      *(ub32 *)loc |= bits(val, 23, 2);
+    case R_SPARC_WDISP22:
+      check(S + A - P, -(1 << 23), 1 << 23);
+      *(ub32 *)loc |= bits(S + A - P, 23, 2);
       break;
-    }
     case R_SPARC_WDISP30:
-    case R_SPARC_WPLT30: {
-      i64 val = S + A - P;
-      check(val, -(1LL << 31), 1LL << 31);
-      *(ub32 *)loc |= bits(val, 31, 2);
+    case R_SPARC_WPLT30:
+      check(S + A - P, -(1LL << 31), 1LL << 31);
+      *(ub32 *)loc |= bits(S + A - P, 31, 2);
       break;
-    }
     case R_SPARC_HI22:
     case R_SPARC_HIPLT22:
     case R_SPARC_LM22:
@@ -321,36 +288,32 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       // We always have to relax a GOT load to a load immediate if a
       // symbol is local, because R_SPARC_GOTDATA_OP cannot represent
       // an addend for a local symbol.
-      if (sym.is_imported || sym.is_ifunc()) {
+      if (sym.is_absolute()) {
+        i64 val = S + A;
+        *(ub32 *)loc |= bits(val < 0 ? ~val : val, 31, 10);
+      } else if (sym.is_pcrel_linktime_const(ctx)) {
+        i64 val = S + A - GOT;
+        *(ub32 *)loc |= bits(val < 0 ? ~val : val, 31, 10);
+      } else {
         *(ub32 *)loc |= bits(G, 31, 10);
-      } else if (sym.is_absolute()) {
-        i64 val = S + A;
-        *(ub32 *)loc |= bits(val < 0 ? ~val : val, 31, 10);
-      } else {
-        i64 val = S + A - GOT;
-        *(ub32 *)loc |= bits(val < 0 ? ~val : val, 31, 10);
       }
       break;
-    case R_SPARC_GOTDATA_OP_LOX10: {
-      if (sym.is_imported || sym.is_ifunc()) {
+    case R_SPARC_GOTDATA_OP_LOX10:
+      if (sym.is_absolute()) {
+        i64 val = S + A;
+        *(ub32 *)loc |= bits(val, 9, 0) | (val < 0 ? 0b1'1100'0000'0000 : 0);
+      } else if (sym.is_pcrel_linktime_const(ctx)) {
+        i64 val = S + A - GOT;
+        *(ub32 *)loc |= bits(val, 9, 0) | (val < 0 ? 0b1'1100'0000'0000 : 0);
+      } else {
         *(ub32 *)loc |= bits(G, 9, 0);
-      } else if (sym.is_absolute()) {
-        i64 val = S + A;
-        *(ub32 *)loc |= bits(val, 9, 0) | (val < 0 ? 0b1'1100'0000'0000 : 0);
-      } else {
-        i64 val = S + A - GOT;
-        *(ub32 *)loc |= bits(val, 9, 0) | (val < 0 ? 0b1'1100'0000'0000 : 0);
       }
       break;
-    }
     case R_SPARC_GOTDATA_OP:
-      if (sym.is_imported || sym.is_ifunc())
-        break;
-
       if (sym.is_absolute()) {
         // ldx [ %g2 + %g1 ], %g1  →  nop
         *(ub32 *)loc = 0x0100'0000;
-      } else {
+      } else if (sym.is_pcrel_linktime_const(ctx)) {
         // ldx [ %g2 + %g1 ], %g1  →  add %g2, %g1, %g1
         *(ub32 *)loc &= 0b00'11111'000000'11111'1'11111111'11111;
         *(ub32 *)loc |= 0b10'00000'000000'00000'0'00000000'00000;
@@ -402,12 +365,16 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ub32 *)loc |= bits(sym.get_tlsgd_addr(ctx) + A - GOT, 9, 0);
       break;
     case R_SPARC_TLS_GD_CALL:
-    case R_SPARC_TLS_LDM_CALL:
+    case R_SPARC_TLS_LDM_CALL: {
+      u64 addr;
       if (ctx.arg.is_static)
-        *(ub32 *)loc |= bits(ctx.extra.tls_get_addr->shdr.sh_addr + A - P, 31, 2);
+        addr = ctx.extra.tls_get_addr_sec->shdr.sh_addr;
       else
-        *(ub32 *)loc |= bits(ctx.tls_get_addr->get_addr(ctx) + A - P, 31, 2);
+        addr = ctx.extra.tls_get_addr_sym->get_addr(ctx);
+
+      *(ub32 *)loc |= bits(addr + A - P, 31, 2);
       break;
+    }
     case R_SPARC_TLS_LDM_HI22:
       *(ub32 *)loc |= bits(ctx.got->get_tlsld_addr(ctx) + A - GOT, 31, 10);
       break;
@@ -454,17 +421,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     auto check = [&](i64 val, i64 lo, i64 hi) {
       if (val < lo || hi <= val)
         Error(ctx) << *this << ": relocation " << rel << " against "
@@ -516,18 +478,13 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
 
     switch (rel.r_type) {
     case R_SPARC_64:
@@ -571,17 +528,17 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_SPARC_PCPLT10:
     case R_SPARC_PLT64:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
     case R_SPARC_GOT13:
     case R_SPARC_GOT10:
     case R_SPARC_GOT22:
     case R_SPARC_GOTDATA_HIX22:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT;
       break;
     case R_SPARC_GOTDATA_OP_HIX22:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_GOT;
       break;
     case R_SPARC_DISP16:
     case R_SPARC_DISP32:
@@ -597,18 +554,23 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
       scan_pcrel(ctx, sym, rel);
       break;
     case R_SPARC_TLS_GD_HI22:
-      sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+      sym.flags |= NEEDS_TLSGD;
       break;
     case R_SPARC_TLS_LDM_HI22:
-      ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+      ctx.needs_tlsld = true;
       break;
     case R_SPARC_TLS_IE_HI22:
-      sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOTTP;
       break;
     case R_SPARC_TLS_GD_CALL:
     case R_SPARC_TLS_LDM_CALL:
-      if (!ctx.arg.is_static && ctx.tls_get_addr->is_imported)
-        ctx.tls_get_addr->flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+      if (!ctx.arg.is_static)
+        if (Symbol<E> &sym = *ctx.extra.tls_get_addr_sym; sym.is_imported)
+          sym.flags |= NEEDS_PLT;
+      break;
+    case R_SPARC_TLS_LE_HIX22:
+    case R_SPARC_TLS_LE_LOX10:
+      check_tlsle(ctx, sym, rel);
       break;
     case R_SPARC_GOTDATA_OP_LOX10:
     case R_SPARC_GOTDATA_OP:
@@ -624,12 +586,10 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_SPARC_TLS_IE_LD:
     case R_SPARC_TLS_IE_LDX:
     case R_SPARC_TLS_IE_LO10:
-    case R_SPARC_TLS_LE_HIX22:
-    case R_SPARC_TLS_LE_LOX10:
     case R_SPARC_SIZE32:
       break;
     default:
-      Fatal(ctx) << *this << ": scan_relocations: " << rel;
+      Error(ctx) << *this << ": unknown relocation: " << rel;
     }
   }
 }

elf/arch-x86-64.cc

@@ -23,8 +23,8 @@
 // TLS block as TP (with some addend). As a result, offsets from TP to
 // thread-local variables (TLVs) in the main executable are all negative.
 //
-// https://github.com/rui314/mold/wiki/i386-psabi.pdf
-// https://github.com/rui314/mold/wiki/x86-64-psabi.pdf
+// https://github.com/rui314/psabi/blob/main/x86-64.pdf
+// https://github.com/rui314/psabi/blob/main/i386.pdf
 
 #include "mold.h"
 
@@ -89,8 +89,8 @@ void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
 }
 
 template <>
-void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,
-                                    u64 offset, u64 val) {
+void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
+                                       u64 offset, u64 val) {
   u8 *loc = ctx.buf + this->shdr.sh_offset + offset;
 
   switch (rel.r_type) {
@@ -165,26 +165,143 @@ static u32 relax_gottpoff(u8 *loc) {
   return 0;
 }
 
-static u32 relax_gotpc32_tlsdesc(u8 *loc) {
-  switch ((loc[0] << 16) | (loc[1] << 8) | loc[2]) {
-  case 0x488d05: return 0x48c7c0; // lea 0(%rip), %rax -> mov $0, %rax
-  case 0x488d0d: return 0x48c7c1; // lea 0(%rip), %rcx -> mov $0, %rcx
-  case 0x488d15: return 0x48c7c2; // lea 0(%rip), %rdx -> mov $0, %rdx
-  case 0x488d1d: return 0x48c7c3; // lea 0(%rip), %rbx -> mov $0, %rbx
-  case 0x488d25: return 0x48c7c4; // lea 0(%rip), %rsp -> mov $0, %rsp
-  case 0x488d2d: return 0x48c7c5; // lea 0(%rip), %rbp -> mov $0, %rbp
-  case 0x488d35: return 0x48c7c6; // lea 0(%rip), %rsi -> mov $0, %rsi
-  case 0x488d3d: return 0x48c7c7; // lea 0(%rip), %rdi -> mov $0, %rdi
-  case 0x4c8d05: return 0x49c7c0; // lea 0(%rip), %r8  -> mov $0, %r8
-  case 0x4c8d0d: return 0x49c7c1; // lea 0(%rip), %r9  -> mov $0, %r9
-  case 0x4c8d15: return 0x49c7c2; // lea 0(%rip), %r10 -> mov $0, %r10
-  case 0x4c8d1d: return 0x49c7c3; // lea 0(%rip), %r11 -> mov $0, %r11
-  case 0x4c8d25: return 0x49c7c4; // lea 0(%rip), %r12 -> mov $0, %r12
-  case 0x4c8d2d: return 0x49c7c5; // lea 0(%rip), %r13 -> mov $0, %r13
-  case 0x4c8d35: return 0x49c7c6; // lea 0(%rip), %r14 -> mov $0, %r14
-  case 0x4c8d3d: return 0x49c7c7; // lea 0(%rip), %r15 -> mov $0, %r15
+// Rewrite a function call to __tls_get_addr to a cheaper instruction
+// sequence. We can do this when we know the thread-local variable's TP-
+// relative address at link-time.
+static void relax_gd_to_le(u8 *loc, ElfRel<E> rel, u64 val) {
+  switch (rel.r_type) {
+  case R_X86_64_PLT32:
+  case R_X86_64_PC32:
+  case R_X86_64_GOTPCREL:
+  case R_X86_64_GOTPCRELX: {
+    // The original instructions are the following:
+    //
+    //  66 48 8d 3d 00 00 00 00    lea  foo@tlsgd(%rip), %rdi
+    //  66 66 48 e8 00 00 00 00    call __tls_get_addr
+    //
+    // or
+    //
+    //  66 48 8d 3d 00 00 00 00    lea foo@tlsgd(%rip), %rdi
+    //  66 48 ff 15 00 00 00 00    call *__tls_get_addr@GOT(%rip)
+    static const u8 insn[] = {
+      0x64, 0x48, 0x8b, 0x04, 0x25, 0, 0, 0, 0, // mov %fs:0, %rax
+      0x48, 0x81, 0xc0, 0, 0, 0, 0,             // add $tp_offset, %rax
+    };
+    memcpy(loc - 4, insn, sizeof(insn));
+    *(ul32 *)(loc + 8) = val;
+    break;
+  }
+  case R_X86_64_PLTOFF64: {
+    // The original instructions are the following:
+    //
+    //  48 8d 3d 00 00 00 00           lea    foo@tlsgd(%rip), %rdi
+    //  48 b8 00 00 00 00 00 00 00 00  movabs __tls_get_addr, %rax
+    //  48 01 d8                       add    %rbx, %rax
+    //  ff d0                          call   *%rax
+    static const u8 insn[] = {
+      0x64, 0x48, 0x8b, 0x04, 0x25, 0, 0, 0, 0, // mov %fs:0, %rax
+      0x48, 0x81, 0xc0, 0, 0, 0, 0,             // add $tp_offset, %rax
+      0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,       // nop
+    };
+    memcpy(loc - 3, insn, sizeof(insn));
+    *(ul32 *)(loc + 9) = val;
+    break;
+  }
+  default:
+    unreachable();
+  }
+}
+
+static void relax_gd_to_ie(u8 *loc, ElfRel<E> rel, u64 val) {
+  switch (rel.r_type) {
+  case R_X86_64_PLT32:
+  case R_X86_64_PC32:
+  case R_X86_64_GOTPCREL:
+  case R_X86_64_GOTPCRELX: {
+    static const u8 insn[] = {
+      0x64, 0x48, 0x8b, 0x04, 0x25, 0, 0, 0, 0, // mov %fs:0, %rax
+      0x48, 0x03, 0x05, 0, 0, 0, 0,             // add foo@gottpoff(%rip), %rax
+    };
+    memcpy(loc - 4, insn, sizeof(insn));
+    *(ul32 *)(loc + 8) = val - 12;
+    break;
+  }
+  case R_X86_64_PLTOFF64: {
+    static const u8 insn[] = {
+      0x64, 0x48, 0x8b, 0x04, 0x25, 0, 0, 0, 0, // mov %fs:0, %rax
+      0x48, 0x03, 0x05, 0, 0, 0, 0,             // add foo@gottpoff(%rip), %rax
+      0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,       // nop
+    };
+    memcpy(loc - 3, insn, sizeof(insn));
+    *(ul32 *)(loc + 9) = val - 13;
+    break;
+  }
+  default:
+    unreachable();
+  }
+}
+
+// Rewrite a function call to __tls_get_addr to a cheaper instruction
+// sequence. The difference from relax_gd_to_le is that we are
+// materializing a Dynamic Thread Pointer for the current ELF module
+// instead of an address for a particular thread-local variable.
+static void relax_ld_to_le(u8 *loc, ElfRel<E> rel, u64 tls_size) {
+  switch (rel.r_type) {
+  case R_X86_64_PLT32:
+  case R_X86_64_PC32: {
+    // The original instructions are the following:
+    //
+    //  48 8d 3d 00 00 00 00    lea    foo@tlsld(%rip), %rdi
+    //  e8 00 00 00 00          call   __tls_get_addr
+    //
+    // The instructions are so short that we cannot rewrite them with
+    // "mov %fs:0, %rax" which is 9 bytes long. We use a shorter code
+    // sequence instead. Since "xor %eax, %eax" zero-clears %rax, the
+    // meaning is equivalent.
+    static const u8 insn[] = {
+      0x31, 0xc0,                   // xor %eax, %eax
+      0x64, 0x48, 0x8b, 0x00,       // mov %fs:(%rax), %rax
+      0x48, 0x2d, 0, 0, 0, 0,       // sub $tls_size, %rax
+    };
+    memcpy(loc - 3, insn, sizeof(insn));
+    *(ul32 *)(loc + 5) = tls_size;
+    break;
+  }
+  case R_X86_64_GOTPCREL:
+  case R_X86_64_GOTPCRELX: {
+    // The original instructions are the following:
+    //
+    //  48 8d 3d 00 00 00 00    lea    foo@tlsld(%rip), %rdi
+    //  ff 15 00 00 00 00       call   *__tls_get_addr@GOT(%rip)
+    static const u8 insn[] = {
+      0x31, 0xc0,                   // xor %eax, %eax
+      0x64, 0x48, 0x8b, 0x00,       // mov %fs:(%rax), %rax
+      0x48, 0x2d, 0, 0, 0, 0,       // sub $tls_size, %rax
+      0x90,                         // nop
+    };
+    memcpy(loc - 3, insn, sizeof(insn));
+    *(ul32 *)(loc + 5) = tls_size;
+    break;
+  }
+  case R_X86_64_PLTOFF64: {
+    // The original instructions are the following:
+    //
+    //  48 8d 3d 00 00 00 00           lea    foo@tlsld(%rip), %rdi
+    //  48 b8 00 00 00 00 00 00 00 00  movabs __tls_get_addr@GOTOFF, %rax
+    //  48 01 d8                       add    %rbx, %rax
+    //  ff d0                          call   *%rax
+    static const u8 insn[] = {
+      0x64, 0x48, 0x8b, 0x04, 0x25, 0, 0, 0, 0, // mov %fs:0, %rax
+      0x48, 0x2d, 0, 0, 0, 0,                   // sub $tls_size, %rax
+      0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00, // nop
+    };
+    memcpy(loc - 3, insn, sizeof(insn));
+    *(ul32 *)(loc + 8) = tls_size;
+    break;
+  }
+  default:
+    unreachable();
   }
-  return 0;
 }
 
 // Apply relocations to SHF_ALLOC sections (i.e. sections that are
@@ -231,18 +348,14 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     u64 GOTPLT = ctx.gotplt->shdr.sh_addr;
 
     switch (rel.r_type) {
-    case R_X86_64_8: {
-      i64 val = S + A;
-      check(val, 0, 1 << 8);
-      *loc = val;
+    case R_X86_64_8:
+      check(S + A, 0, 1 << 8);
+      *loc = S + A;
       break;
-    }
-    case R_X86_64_16: {
-      i64 val = S + A;
-      check(val, 0, 1 << 16);
-      *(ul16 *)loc = val;
+    case R_X86_64_16:
+      check(S + A, 0, 1 << 16);
+      *(ul16 *)loc = S + A;
       break;
-    }
     case R_X86_64_32:
       write32(S + A);
       break;
@@ -250,20 +363,16 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       write32s(S + A);
       break;
     case R_X86_64_64:
-      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, dynrel);
+      apply_dyn_absrel(ctx, sym, rel, loc, S, A, P, &dynrel);
       break;
-    case R_X86_64_PC8: {
-      i64 val = S + A - P;
-      check(val, -(1 << 7), 1 << 7);
-      *loc = val;
+    case R_X86_64_PC8:
+      check(S + A - P, -(1 << 7), 1 << 7);
+      *loc = S + A - P;
       break;
-    }
-    case R_X86_64_PC16: {
-      i64 val = S + A - P;
-      check(val, -(1 << 15), 1 << 15);
-      *(ul16 *)loc = val;
+    case R_X86_64_PC16:
+      check(S + A - P, -(1 << 15), 1 << 15);
+      *(ul16 *)loc = S + A - P;
       break;
-    }
     case R_X86_64_PC32:
     case R_X86_64_PLT32:
       write32s(S + A - P);
@@ -272,7 +381,7 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ul64 *)loc = S + A - P;
       break;
     case R_X86_64_GOT32:
-      write32s(G + A);
+      write32(G + A);
       break;
     case R_X86_64_GOT64:
       *(ul64 *)loc = G + A;
@@ -294,146 +403,48 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       *(ul64 *)loc = G + GOTPLT + A - P;
       break;
     case R_X86_64_GOTPCRELX:
-      if (sym.has_got(ctx)) {
-        write32s(G + GOTPLT + A - P);
-      } else {
+      // We always want to relax GOTPCRELX relocs even if --no-relax
+      // was given because some static PIE runtime code depends on these
+      // relaxations.
+      if (sym.is_pcrel_linktime_const(ctx)) {
         u32 insn = relax_gotpcrelx(loc - 2);
-        loc[-2] = insn >> 8;
-        loc[-1] = insn;
-        write32s(S + A - P);
+        i64 val = S + A - P;
+        if (insn && (i32)val == val) {
+          loc[-2] = insn >> 8;
+          loc[-1] = insn;
+          *(ul32 *)loc = val;
+          break;
+        }
       }
+      write32s(G + GOTPLT + A - P);
       break;
     case R_X86_64_REX_GOTPCRELX:
-      if (sym.has_got(ctx)) {
-        write32s(G + GOTPLT + A - P);
-      } else {
+      if (sym.is_pcrel_linktime_const(ctx)) {
         u32 insn = relax_rex_gotpcrelx(loc - 3);
-        loc[-3] = insn >> 16;
-        loc[-2] = insn >> 8;
-        loc[-1] = insn;
-        write32s(S + A - P);
+        i64 val = S + A - P;
+        if (insn && (i32)val == val) {
+          loc[-3] = insn >> 16;
+          loc[-2] = insn >> 8;
+          loc[-1] = insn;
+          *(ul32 *)loc = val;
+          break;
+        }
       }
+      write32s(G + GOTPLT + A - P);
       break;
     case R_X86_64_TLSGD:
-      if (sym.has_tlsgd(ctx)) {
+      if (sym.has_tlsgd(ctx))
         write32s(sym.get_tlsgd_addr(ctx) + A - P);
-      } else {
-        // Relax GD to LE. If we are creating an exectuable, the offset of
-        // a thread-local variable from TP is a link-time constant. So we
-        // don't need to call __tls_get_addr to obtain the address of a TLV.
-        i64 val = S - ctx.tp_addr;
-        assert(A == -4);
-        check(val, -(1LL << 31), 1LL << 31);
-
-        switch (rels[i + 1].r_type) {
-        case R_X86_64_PLT32:
-        case R_X86_64_PC32:
-        case R_X86_64_GOTPCREL:
-        case R_X86_64_GOTPCRELX: {
-          // The original instructions are the following:
-          //
-          //  66 48 8d 3d 00 00 00 00    lea  foo@tlsgd(%rip), %rdi
-          //  66 66 48 e8 00 00 00 00    call __tls_get_addr
-          //
-          // or
-          //
-          //  66 48 8d 3d 00 00 00 00    lea foo@tlsgd(%rip), %rdi
-          //  66 48 ff 15 00 00 00 00    call *__tls_get_addr@GOT(%rip)
-          static const u8 insn[] = {
-            0x64, 0x48, 0x8b, 0x04, 0x25, 0, 0, 0, 0, // mov %fs:0, %rax
-            0x48, 0x81, 0xc0, 0, 0, 0, 0,             // add $val, %rax
-          };
-          memcpy(loc - 4, insn, sizeof(insn));
-          *(ul32 *)(loc + 8) = val;
-          break;
-        }
-        case R_X86_64_PLTOFF64: {
-          // The original instructions are the following:
-          //
-          //  48 8d 3d 00 00 00 00           lea    foo@tlsgd(%rip), %rdi
-          //  48 b8 00 00 00 00 00 00 00 00  movabs __tls_get_addr, %rax
-          //  48 01 d8                       add    %rbx, %rax
-          //  ff d0                          call   *%rax
-          static const u8 insn[] = {
-            0x64, 0x48, 0x8b, 0x04, 0x25, 0, 0, 0, 0, // mov %fs:0, %rax
-            0x48, 0x81, 0xc0, 0, 0, 0, 0,             // add $val, %rax
-            0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,       // nop
-          };
-          memcpy(loc - 3, insn, sizeof(insn));
-          *(ul32 *)(loc + 9) = val;
-          break;
-        }
-        default:
-          unreachable();
-        }
-
-        i++;
-      }
+      else if (sym.has_gottp(ctx))
+        relax_gd_to_ie(loc, rels[++i], sym.get_gottp_addr(ctx) - P);
+      else
+        relax_gd_to_le(loc, rels[++i], S - ctx.tp_addr);
       break;
     case R_X86_64_TLSLD:
-      if (ctx.got->has_tlsld(ctx)) {
+      if (ctx.got->has_tlsld(ctx))
         write32s(ctx.got->get_tlsld_addr(ctx) + A - P);
-      } else {
-        // Relax LD to LE. If we are creating an executable, we don't need
-        // to call __tls_get_addr to obtain the address of the beginning
-        // of the current TLS block. TP points past the end of the TLS
-        // block, and we know the exact size of the TLS block we have
-        // created, so we can just subtract it from TP.
-        switch (rels[i + 1].r_type) {
-        case R_X86_64_PLT32:
-        case R_X86_64_PC32: {
-          // The original instructions are the following:
-          //
-          //  48 8d 3d 00 00 00 00    lea    foo@tlsld(%rip), %rdi
-          //  e8 00 00 00 00          call   __tls_get_addr
-          static const u8 insn[] = {
-            0x31, 0xc0,                   // xor %eax, %eax
-            0x64, 0x48, 0x8b, 0x00,       // mov %fs:(%rax), %rax
-            0x48, 0x2d, 0, 0, 0, 0,       // sub $tls_size, %rax
-          };
-          memcpy(loc - 3, insn, sizeof(insn));
-          break;
-        }
-        case R_X86_64_GOTPCREL:
-        case R_X86_64_GOTPCRELX: {
-          // The original instructions are the following:
-          //
-          //  48 8d 3d 00 00 00 00    lea    foo@tlsld(%rip), %rdi
-          //  ff 15 00 00 00 00       call   *__tls_get_addr@GOT(%rip)
-          static const u8 insn[] = {
-            0x31, 0xc0,                   // xor %eax, %eax
-            0x64, 0x48, 0x8b, 0x00,       // mov %fs:(%rax), %rax
-            0x48, 0x2d, 0, 0, 0, 0,       // sub $tls_size, %rax
-            0x90,                         // nop
-          };
-          memcpy(loc - 3, insn, sizeof(insn));
-          break;
-        }
-        case R_X86_64_PLTOFF64: {
-          // The original instructions are the following:
-          //
-          //  48 8d 3d 00 00 00 00           lea    foo@tlsld(%rip), %rdi
-          //  48 b8 00 00 00 00 00 00 00 00  movabs __tls_get_addr@GOTOFF, %rax
-          //  48 01 d8                       add    %rbx, %rax
-          //  ff d0                          call   *%rax
-          static const u8 insn[] = {
-            0x31, 0xc0,                   // xor %eax, %eax
-            0x64, 0x48, 0x8b, 0x00,       // mov %fs:(%rax), %rax
-            0x48, 0x2d, 0, 0, 0, 0,       // sub $tls_size, %rax
-            0x0f, 0x1f, 0x44, 0x00, 0x00, // nop
-            0x0f, 0x1f, 0x44, 0x00, 0x00, // nop
-          };
-          memcpy(loc - 3, insn, sizeof(insn));
-          break;
-        }
-        default:
-          unreachable();
-        }
-
-        *(ul32 *)(loc + 5) = ctx.tp_addr - ctx.tls_begin;
-        assert(A == -4);
-        i++;
-      }
+      else
+        relax_ld_to_le(loc, rels[++i], ctx.tp_addr - ctx.tls_begin);
       break;
     case R_X86_64_DTPOFF32:
       write32s(S + A - ctx.dtp_addr);
@@ -460,15 +471,46 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       }
       break;
     case R_X86_64_GOTPC32_TLSDESC:
+      // x86-64 TLSDESC uses the following code sequence to materialize
+      // a TP-relative address in %rax.
+      //
+      //   lea    0(%rip), %rax
+      //       R_X86_64_GOTPC32_TLSDESC    foo
+      //   call   *(%rax)
+      //       R_X86_64_TLSDESC_CALL       foo
+      //
+      // We may relax the instructions to the following for non-dlopen'd DSO
+      //
+      //   mov     foo@GOTTPOFF(%rip), %rax
+      //   nop
+      //
+      // or to the following for executable.
+      //
+      //   mov     $foo@TPOFF, %rax
+      //   nop
       if (sym.has_tlsdesc(ctx)) {
         write32s(sym.get_tlsdesc_addr(ctx) + A - P);
+      } else if (sym.has_gottp(ctx)) {
+        // mov foo@gottpoff(%rip), %rax
+        loc[-3] = 0x48;
+        loc[-2] = 0x8b;
+        loc[-1] = 0x05;
+        write32s(sym.get_gottp_addr(ctx) + A - P);
       } else {
-        u32 insn = relax_gotpc32_tlsdesc(loc - 3);
-        loc[-3] = insn >> 16;
-        loc[-2] = insn >> 8;
-        loc[-1] = insn;
+        // mov $foo@tpoff, %rax
+        loc[-3] = 0x48;
+        loc[-2] = 0xc7;
+        loc[-1] = 0xc0;
         write32s(S - ctx.tp_addr);
-        assert(A == -4);
+      }
+      break;
+    case R_X86_64_TLSDESC_CALL:
+      if (sym.has_tlsdesc(ctx)) {
+        // Do nothing
+      } else {
+        // call *(%rax) -> nop
+        loc[0] = 0x66;
+        loc[1] = 0x90;
       }
       break;
     case R_X86_64_SIZE32:
@@ -477,13 +519,6 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_X86_64_SIZE64:
       *(ul64 *)loc = sym.esym().st_size + A;
       break;
-    case R_X86_64_TLSDESC_CALL:
-      if (!sym.has_tlsdesc(ctx)) {
-        // call *(%rax) -> nop
-        loc[0] = 0x66;
-        loc[1] = 0x90;
-      }
-      break;
     default:
       unreachable();
     }
@@ -508,17 +543,12 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
 
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = base + rel.r_offset;
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     auto check = [&](i64 val, i64 lo, i64 hi) {
       if (val < lo || hi <= val)
         Error(ctx) << *this << ": relocation " << rel << " against "
@@ -544,18 +574,14 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
     u64 A = frag ? frag_addend : (i64)rel.r_addend;
 
     switch (rel.r_type) {
-    case R_X86_64_8: {
-      i64 val = S + A;
-      check(val, 0, 1 << 8);
-      *loc = val;
+    case R_X86_64_8:
+      check(S + A, 0, 1 << 8);
+      *loc = S + A;
       break;
-    }
-    case R_X86_64_16: {
-      i64 val = S + A;
-      check(val, 0, 1 << 16);
-      *(ul16 *)loc = val;
+    case R_X86_64_16:
+      check(S + A, 0, 1 << 16);
+      *(ul16 *)loc = S + A;
       break;
-    }
     case R_X86_64_32:
       write32(S + A);
       break;
@@ -580,6 +606,15 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
       else
         *(ul64 *)loc = S + A - ctx.dtp_addr;
       break;
+    case R_X86_64_GOTOFF64:
+      *(ul64 *)loc = S + A - ctx.gotplt->shdr.sh_addr;
+      break;
+    case R_X86_64_GOTPC64:
+      // PC-relative relocation doesn't make sense for non-memory-allocated
+      // section, but GCC 6.3.0 seems to create this reloc for
+      // _GLOBAL_OFFSET_TABLE_.
+      *(ul64 *)loc = ctx.gotplt->shdr.sh_addr + A;
+      break;
     case R_X86_64_SIZE32:
       write32(sym.esym().st_size + A);
       break;
@@ -609,19 +644,27 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations
   for (i64 i = 0; i < rels.size(); i++) {
     const ElfRel<E> &rel = rels[i];
-    if (rel.r_type == R_NONE)
+    if (rel.r_type == R_NONE || record_undef_error(ctx, rel))
       continue;
 
     Symbol<E> &sym = *file.symbols[rel.r_sym];
     u8 *loc = (u8 *)(contents.data() + rel.r_offset);
 
-    if (!sym.file) {
-      record_undef_error(ctx, rel);
-      continue;
-    }
-
     if (sym.is_ifunc())
-      sym.flags.fetch_or(NEEDS_GOT | NEEDS_PLT, std::memory_order_relaxed);
+      sym.flags |= NEEDS_GOT | NEEDS_PLT;
+
+    if (rel.r_type == R_X86_64_TLSGD || rel.r_type == R_X86_64_TLSLD) {
+      if (i + 1 == rels.size())
+        Fatal(ctx) << *this << ": " << rel
+                   << " must be followed by PLT or GOTPCREL";
+
+      if (u32 ty = rels[i + 1].r_type;
+          ty != R_X86_64_PLT32 && ty != R_X86_64_PC32 &&
+          ty != R_X86_64_PLTOFF64 && ty != R_X86_64_GOTPCREL &&
+          ty != R_X86_64_GOTPCRELX)
+        Fatal(ctx) << *this << ": " << rel
+                   << " must be followed by PLT or GOTPCREL";
+    }
 
     switch (rel.r_type) {
     case R_X86_64_8:
@@ -645,103 +688,57 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_X86_64_GOTPC64:
     case R_X86_64_GOTPCREL:
     case R_X86_64_GOTPCREL64:
-      sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
+    case R_X86_64_GOTPCRELX:
+    case R_X86_64_REX_GOTPCRELX:
+      sym.flags |= NEEDS_GOT;
       break;
-    case R_X86_64_GOTPCRELX: {
-      if (rel.r_addend != -4)
-        Fatal(ctx) << *this << ": bad r_addend for R_X86_64_GOTPCRELX";
-
-      bool do_relax = ctx.arg.relax && !sym.is_imported &&
-                      sym.is_relative() && relax_gotpcrelx(loc - 2);
-      if (!do_relax)
-        sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
-      break;
-    }
-    case R_X86_64_REX_GOTPCRELX: {
-      if (rel.r_addend != -4)
-        Fatal(ctx) << *this << ": bad r_addend for R_X86_64_REX_GOTPCRELX";
-
-      bool do_relax = ctx.arg.relax && !sym.is_imported &&
-                      sym.is_relative() && relax_rex_gotpcrelx(loc - 3);
-      if (!do_relax)
-        sym.flags.fetch_or(NEEDS_GOT, std::memory_order_relaxed);
-      break;
-    }
     case R_X86_64_PLT32:
     case R_X86_64_PLTOFF64:
       if (sym.is_imported)
-        sym.flags.fetch_or(NEEDS_PLT, std::memory_order_relaxed);
+        sym.flags |= NEEDS_PLT;
       break;
-    case R_X86_64_TLSGD: {
-      if (rel.r_addend != -4)
-        Fatal(ctx) << *this << ": bad r_addend for R_X86_64_TLSGD";
-
-      if (i + 1 == rels.size())
-        Fatal(ctx) << *this << ": TLSGD reloc must be followed by PLT or GOTPCREL";
-
-      if (u32 ty = rels[i + 1].r_type;
-          ty != R_X86_64_PLT32 && ty != R_X86_64_PC32 &&
-          ty != R_X86_64_PLTOFF64 && ty != R_X86_64_GOTPCREL &&
-          ty != R_X86_64_GOTPCRELX)
-        Fatal(ctx) << *this << ": TLSGD reloc must be followed by PLT or GOTPCREL";
-
-      if (relax_tlsgd(ctx, sym))
+    case R_X86_64_TLSGD:
+      if ((ctx.arg.relax && sym.is_tprel_linktime_const(ctx)) ||
+          ctx.arg.is_static) {
+        // We always relax if -static because libc.a doesn't contain
+        // __tls_get_addr().
+        i++;
+      } else if (ctx.arg.relax && sym.is_tprel_runtime_const(ctx)) {
+        sym.flags |= NEEDS_GOTTP;
+        i++;
+      } else {
+        sym.flags |= NEEDS_TLSGD;
+      }
+      break;
+    case R_X86_64_TLSLD:
+      // We always relax if -static because libc.a doesn't contain
+      // __tls_get_addr().
+      if (ctx.arg.is_static || (ctx.arg.relax && !ctx.arg.shared))
         i++;
       else
-        sym.flags.fetch_or(NEEDS_TLSGD, std::memory_order_relaxed);
+        ctx.needs_tlsld = true;
       break;
-    }
-    case R_X86_64_TLSLD: {
-      if (rel.r_addend != -4)
-        Fatal(ctx) << *this << ": bad r_addend for R_X86_64_TLSLD";
-
-      if (i + 1 == rels.size())
-        Fatal(ctx) << *this << ": TLSLD reloc must be followed by PLT or GOTPCREL";
-
-      if (u32 ty = rels[i + 1].r_type;
-          ty != R_X86_64_PLT32 && ty != R_X86_64_PC32 &&
-          ty != R_X86_64_PLTOFF64 && ty != R_X86_64_GOTPCREL &&
-          ty != R_X86_64_GOTPCRELX)
-        Fatal(ctx) << *this << ": TLSLD reloc must be followed by PLT or GOTPCREL";
-
-      if (relax_tlsld(ctx))
-        i++;
-      else
-        ctx.needs_tlsld.store(true, std::memory_order_relaxed);
+    case R_X86_64_GOTTPOFF:
+      if (ctx.arg.relax && relax_gottpoff(loc - 3) &&
+          sym.is_tprel_linktime_const(ctx)) {
+        // do nothing
+      } else {
+        sym.flags |= NEEDS_GOTTP;
+      }
       break;
-    }
-    case R_X86_64_GOTTPOFF: {
-      if (rel.r_addend != -4)
-        Fatal(ctx) << *this << ": bad r_addend for R_X86_64_GOTTPOFF";
-
-      ctx.has_gottp_rel.store(true, std::memory_order_relaxed);
-
-      bool do_relax = ctx.arg.relax && !ctx.arg.shared &&
-                      !sym.is_imported && relax_gottpoff(loc - 3);
-      if (!do_relax)
-        sym.flags.fetch_or(NEEDS_GOTTP, std::memory_order_relaxed);
+    case R_X86_64_TLSDESC_CALL:
+      scan_tlsdesc(ctx, sym);
       break;
-    }
-    case R_X86_64_GOTPC32_TLSDESC: {
-      if (rel.r_addend != -4)
-        Fatal(ctx) << *this << ": bad r_addend for R_X86_64_GOTPC32_TLSDESC";
-
-      if (relax_gotpc32_tlsdesc(loc - 3) == 0)
-        Fatal(ctx) << *this << ": GOTPC32_TLSDESC relocation is used"
-                   << " against an invalid code sequence";
-
-      if (!relax_tlsdesc(ctx, sym))
-        sym.flags.fetch_or(NEEDS_TLSDESC, std::memory_order_relaxed);
+    case R_X86_64_TPOFF32:
+    case R_X86_64_TPOFF64:
+      check_tlsle(ctx, sym, rel);
       break;
-    }
     case R_X86_64_GOTOFF64:
     case R_X86_64_DTPOFF32:
     case R_X86_64_DTPOFF64:
-    case R_X86_64_TPOFF32:
-    case R_X86_64_TPOFF64:
     case R_X86_64_SIZE32:
     case R_X86_64_SIZE64:
-    case R_X86_64_TLSDESC_CALL:
+    case R_X86_64_GOTPC32_TLSDESC:
       break;
     default:
       Error(ctx) << *this << ": unknown relocation: " << rel;

elf/cmdline.cc

@@ -1,6 +1,7 @@
 #include "mold.h"
-#include "../cmdline.h"
+#include "../common/cmdline.h"
 
+#include <random>
 #include <regex>
 #include <sstream>
 #include <sys/stat.h>
@@ -98,7 +99,7 @@ Options:
   --gc-sections               Remove unreferenced sections
     --no-gc-sections
   --gdb-index                 Create .gdb_index for faster gdb startup
-  --hash-style [sysv,gnu,both]
+  --hash-style [sysv,gnu,both,none]
                               Set hash style
   --icf=[all,safe,none]       Fold identical code
     --no-icf
@@ -187,18 +188,21 @@ Options:
   -z origin                   Mark object requiring immediate $ORIGIN processing at runtime
   -z pack-relative-relocs     Alias for --pack-dyn-relocs=relr
     -z nopack-relative-relocs
+  -z sectionheader            Do not omit section header (default)
+    -z nosectionheader        Omit section header
   -z separate-loadable-segments
                               Separate all loadable segments to different pages
     -z separate-code          Separate code and data into different pages
     -z noseparate-code        Allow overlap in pages
+  -z stack-size=VALUE         Set size of stack segment
   -z relro                    Make some sections read-only after relocation (default)
     -z norelro
   -z text                     Report error if DT_TEXTREL is set
     -z notext
     -z textoff
 
-mold: supported targets: elf32-i386 elf64-x86-64 elf32-littlearm elf64-littleaarch64 elf32-littleriscv elf32-bigriscv elf64-littleriscv elf64-bigriscv elf32-powerpc elf64-powerpc elf64-powerpc elf64-powerpcle elf64-s390 elf64-sparc elf32-m68k elf32-sh-linux elf64-alpha
-mold: supported emulations: elf_i386 elf_x86_64 armelf_linux_eabi aarch64linux aarch64elf elf32lriscv elf32briscv elf64lriscv elf64briscv elf32ppc elf32ppclinux elf64ppc elf64lppc elf64_s390 elf64_sparc m68kelf shlelf_linux elf64alpha)";
+mold: supported targets: elf32-i386 elf64-x86-64 elf32-littlearm elf64-littleaarch64 elf32-littleriscv elf32-bigriscv elf64-littleriscv elf64-bigriscv elf32-powerpc elf64-powerpc elf64-powerpc elf64-powerpcle elf64-s390 elf64-sparc elf32-m68k elf32-sh-linux elf64-alpha elf64-loongarch elf32-loongarch
+mold: supported emulations: elf_i386 elf_x86_64 armelf_linux_eabi aarch64linux aarch64elf elf32lriscv elf32briscv elf64lriscv elf64briscv elf32ppc elf32ppclinux elf64ppc elf64lppc elf64_s390 elf64_sparc m68kelf shlelf_linux elf64alpha elf64loongarch elf32loongarch)";
 
 static std::vector<std::string> add_dashes(std::string name) {
   // Single-letter option
@@ -398,13 +402,13 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
   std::vector<std::string> remaining;
   std::string_view arg;
 
-  ctx.page_size = E::page_size;
   ctx.arg.color_diagnostics = isatty(STDERR_FILENO);
 
   bool version_shown = false;
   bool warn_shared_textrel = false;
   std::optional<SeparateCodeKind> z_separate_code;
   std::optional<bool> z_relro;
+  std::optional<u64> shuffle_sections_seed;
   std::unordered_set<std::string_view> rpaths;
 
   auto add_rpath = [&](std::string_view arg) {
@@ -423,7 +427,7 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
   // It looks like the SPARC's dynamic linker takes both RELA's r_addend
   // and the value at the relocated place. So we don't want to write
   // values to relocated places.
-  if (is_sparc<E>)
+  if constexpr (is_sparc<E>)
     ctx.arg.apply_dynamic_relocs = false;
 
   auto read_arg = [&](std::string name) {
@@ -520,7 +524,8 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
                    << "   aarch64linux\n   armelf_linux_eabi\n   elf64lriscv\n"
                    << "   elf64briscv\n   elf32lriscv\n   elf32briscv\n"
                    << "   elf32ppc\n   elf64ppc\n   elf64lppc\n   elf64_s390\n"
-                   << "   elf64_sparc\n   m68kelf\n   shlelf_linux\n   elf64alpha";
+                   << "   elf64_sparc\n   m68kelf\n   shlelf_linux\n"
+                   << "   elf64alpha\n   elf64loongarch\n   elf32loongarch";
       version_shown = true;
     } else if (read_arg("m")) {
       if (arg == "elf_x86_64") {
@@ -555,6 +560,10 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
         ctx.arg.emulation = SH4::target_name;
       } else if (arg == "elf64alpha") {
         ctx.arg.emulation = ALPHA::target_name;
+      } else if (arg == "elf64loongarch") {
+        ctx.arg.emulation = LOONGARCH64::target_name;
+      } else if (arg == "elf32loongarch") {
+        ctx.arg.emulation = LOONGARCH32::target_name;
       } else {
         Fatal(ctx) << "unknown -m argument: " << arg;
       }
@@ -582,9 +591,7 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       ctx.arg.Map = arg;
       ctx.arg.print_map = true;
     } else if (read_flag("print-dependencies")) {
-      ctx.arg.print_dependencies = 1;
-    } else if (read_flag("print-dependencies=full")) {
-      ctx.arg.print_dependencies = 2;
+      ctx.arg.print_dependencies = true;
     } else if (read_flag("print-map") || read_flag("M")) {
       ctx.arg.print_map = true;
     } else if (read_flag("Bstatic") || read_flag("dn") || read_flag("static")) {
@@ -625,7 +632,7 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       ctx.arg.shuffle_sections = SHUFFLE_SECTIONS_SHUFFLE;
     } else if (read_eq("shuffle-sections")) {
       ctx.arg.shuffle_sections = SHUFFLE_SECTIONS_SHUFFLE;
-      ctx.arg.shuffle_sections_seed = parse_number(ctx, "shuffle-sections", arg);
+      shuffle_sections_seed = parse_number(ctx, "shuffle-sections", arg);
     } else if (read_flag("reverse-sections")) {
       ctx.arg.shuffle_sections = SHUFFLE_SECTIONS_REVERSE;
     } else if (read_flag("rosegment")) {
@@ -652,7 +659,7 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
         ctx.arg.unresolved_symbols = UNRESOLVED_IGNORE;
       else
         Fatal(ctx) << "unknown --unresolved-symbols argument: " << arg;
-    } else if (read_arg("u") || read_arg("undefined")) {
+    } else if (read_arg("undefined") || read_arg("u")) {
       ctx.arg.undefined.push_back(arg);
     } else if (read_arg("require-defined")) {
       ctx.arg.require_defined.push_back(arg);
@@ -670,6 +677,9 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       } else if (arg == "both") {
         ctx.arg.hash_style_sysv = true;
         ctx.arg.hash_style_gnu = true;
+      } else if (arg == "none") {
+        ctx.arg.hash_style_sysv = false;
+        ctx.arg.hash_style_gnu = false;
       } else {
         Fatal(ctx) << "invalid --hash-style argument: " << arg;
       }
@@ -691,7 +701,8 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
     } else if (read_flag("pie") || read_flag("pic-executable")) {
       ctx.arg.pic = true;
       ctx.arg.pie = true;
-    } else if (read_flag("no-pie") || read_flag("no-pic-executable")) {
+    } else if (read_flag("no-pie") || read_flag("no-pic-executable") ||
+               read_flag("nopie")) {
       ctx.arg.pic = false;
       ctx.arg.pie = false;
     } else if (read_flag("relax")) {
@@ -710,9 +721,11 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       ctx.arg.relocatable_merge_sections = true;
     } else if (read_flag("perf")) {
       ctx.arg.perf = true;
-    } else if (read_flag("pack-dyn-relocs=relr")) {
+    } else if (read_flag("pack-dyn-relocs=relr") ||
+               read_z_flag("pack-relative-relocs")) {
       ctx.arg.pack_dyn_relocs_relr = true;
-    } else if (read_flag("pack-dyn-relocs=none")) {
+    } else if (read_flag("pack-dyn-relocs=none") ||
+               read_z_flag("nopack-relative-relocs")) {
       ctx.arg.pack_dyn_relocs_relr = false;
     } else if (read_arg("package-metadata")) {
       ctx.arg.package_metadata = arg;
@@ -854,20 +867,22 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       ctx.arg.z_origin = true;
     } else if (read_z_flag("nodefaultlib")) {
       ctx.arg.z_nodefaultlib = true;
-    } else if (read_z_flag("pack-relative-relocs")) {
-      ctx.arg.pack_dyn_relocs_relr = true;
-    } else if (read_z_flag("nopack-relative-relocs")) {
-      ctx.arg.pack_dyn_relocs_relr = false;
     } else if (read_z_flag("separate-loadable-segments")) {
       z_separate_code = SEPARATE_LOADABLE_SEGMENTS;
     } else if (read_z_flag("separate-code")) {
       z_separate_code = SEPARATE_CODE;
     } else if (read_z_flag("noseparate-code")) {
       z_separate_code = NOSEPARATE_CODE;
+    } else if (read_z_arg("stack-size")) {
+      ctx.arg.z_stack_size = parse_number(ctx, "-z stack-size", arg);
     } else if (read_z_flag("dynamic-undefined-weak")) {
       ctx.arg.z_dynamic_undefined_weak = true;
     } else if (read_z_flag("nodynamic-undefined-weak")) {
       ctx.arg.z_dynamic_undefined_weak = false;
+    } else if (read_z_flag("sectionheader")) {
+      ctx.arg.z_sectionheader = true;
+    } else if (read_z_flag("nosectionheader")) {
+      ctx.arg.z_sectionheader = false;
     } else if (read_flag("no-undefined")) {
       ctx.arg.z_defs = true;
     } else if (read_flag("fatal-warnings")) {
@@ -993,7 +1008,7 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
     } else if (read_flag("strip-all") || read_flag("s")) {
       ctx.arg.strip_all = true;
     } else if (read_flag("strip-debug") || read_flag("S")) {
-      ctx.arg.strip_all = true;
+      ctx.arg.strip_debug = true;
     } else if (read_flag("warn-unresolved-symbols")) {
       ctx.arg.unresolved_symbols = UNRESOLVED_WARN;
     } else if (read_flag("error-unresolved-symbols")) {
@@ -1046,6 +1061,8 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
     } else if (read_arg("filter") || read_arg("F")) {
       ctx.arg.filter.push_back(arg);
     } else if (read_arg("O")) {
+    } else if (read_flag("EB")) {
+    } else if (read_flag("EL")) {
     } else if (read_flag("O0")) {
     } else if (read_flag("O1")) {
     } else if (read_flag("O2")) {
@@ -1086,7 +1103,7 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
     } else if (read_arg("max-cache-size")) {
     } else if (read_arg("version-script")) {
       // --version-script, --dynamic-list and --export-dynamic-symbol[-list]
-      // are treated as positional arguments even if they are actually not
+      // are treated as positional arguments even though they are actually not
       // positional. This is because linker scripts (a positional argument)
       // can also specify a version script, and it's better to consolidate
       // parsing in read_input_files. In particular, version scripts can
@@ -1123,10 +1140,8 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       Warn(ctx) << "unknown command line option: -z " << args[1];
       args = args.subspan(2);
     } else if (args[0] == "-dynamic") {
-      Fatal(ctx) << "unknown command line option: -dynamic;"
-                 << " -dynamic is a macOS linker's option. If you are trying"
-                 << " to build a binary for an Apple platform, you need to use"
-                 << " ld64.mold instead of mold or ld.mold.";
+      Fatal(ctx) << "unknown command line option: -dynamic; -dynamic is a "
+                 << "macOS linker's option. mold does not support macOS.";
     } else {
       if (args[0][0] == '-')
         Fatal(ctx) << "unknown command line option: " << args[0];
@@ -1163,6 +1178,14 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
   if (ctx.arg.relocatable)
     ctx.arg.is_static = true;
 
+  if (ctx.arg.shuffle_sections == SHUFFLE_SECTIONS_SHUFFLE) {
+    if (shuffle_sections_seed)
+      ctx.arg.shuffle_sections_seed = *shuffle_sections_seed;
+    else
+      ctx.arg.shuffle_sections_seed =
+        ((u64)std::random_device()() << 32) | std::random_device()();
+  }
+
   // --section-order implies `-z separate-loadable-segments`
   if (z_separate_code)
     ctx.arg.z_separate_code = *z_separate_code;
@@ -1182,12 +1205,14 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       Fatal(ctx) << "-auxiliary may not be used without -shared";
   }
 
-  if (!ctx.arg.apply_dynamic_relocs && !E::is_rela)
-    Fatal(ctx) << "--no-apply-dynamic-relocs may not be used on "
-               << E::target_name;
+  if constexpr (!E::is_rela)
+    if (!ctx.arg.apply_dynamic_relocs)
+      Fatal(ctx) << "--no-apply-dynamic-relocs may not be used on "
+                 << E::target_name;
 
-  if (is_sparc<E> && ctx.arg.apply_dynamic_relocs)
-    Fatal(ctx) << "--apply-dynamic-relocs may not be used on SPARC64";
+  if constexpr (is_sparc<E>)
+    if (ctx.arg.apply_dynamic_relocs)
+      Fatal(ctx) << "--apply-dynamic-relocs may not be used on SPARC64";
 
   if (!ctx.arg.section_start.empty() && !ctx.arg.section_order.empty())
     Fatal(ctx) << "--section-start may not be used with --section-order";

elf/dwarf.cc

@@ -77,7 +77,7 @@ read_compunits(Context<E> &ctx, ObjectFile<E> &file) {
   while (!data.empty()) {
     if (data.size() < 4)
       Fatal(ctx) << *file.debug_info << ": corrupted .debug_info";
-    if (*(U32<E> *)data.data() == 0xffffffff)
+    if (*(U32<E> *)data.data() == 0xffff'ffff)
       Fatal(ctx) << *file.debug_info << ": --gdb-index: DWARF64 not supported";
     i64 len = *(U32<E> *)data.data() + 4;
     vec.push_back(data.substr(0, len));
@@ -216,7 +216,7 @@ find_compunit(Context<E> &ctx, ObjectFile<E> &file, i64 offset) {
                << dwarf_version;
   }
 
-  u32 abbrev_code = read_uleb(cu);
+  u32 abbrev_code = read_uleb(&cu);
 
   // Find a .debug_abbrev record corresponding to the .debug_info record.
   // We assume the .debug_info record at a given offset is of
@@ -224,14 +224,14 @@ find_compunit(Context<E> &ctx, ObjectFile<E> &file, i64 offset) {
   u8 *abbrev = get_buffer(ctx, ctx.debug_abbrev) + abbrev_offset;
 
   for (;;) {
-    u32 code = read_uleb(abbrev);
+    u32 code = read_uleb(&abbrev);
     if (code == 0)
       Fatal(ctx) << file << ": --gdb-index: .debug_abbrev does not contain"
                  << " a record for the first .debug_info record";
 
     if (code == abbrev_code) {
       // Found a record
-      u64 abbrev_tag = read_uleb(abbrev);
+      u64 abbrev_tag = read_uleb(&abbrev);
       if (abbrev_tag != DW_TAG_compile_unit && abbrev_tag != DW_TAG_skeleton_unit)
         Fatal(ctx) << file << ": --gdb-index: the first entry's tag is not"
                    << " DW_TAG_compile_unit/DW_TAG_skeleton_unit but 0x"
@@ -240,15 +240,15 @@ find_compunit(Context<E> &ctx, ObjectFile<E> &file, i64 offset) {
     }
 
     // Skip an uninteresting record
-    read_uleb(abbrev); // tag
+    read_uleb(&abbrev); // tag
     abbrev++; // has_children byte
     for (;;) {
-      u64 name = read_uleb(abbrev);
-      u64 form = read_uleb(abbrev);
+      u64 name = read_uleb(&abbrev);
+      u64 form = read_uleb(&abbrev);
       if (name == 0 && form == 0)
         break;
       if (form == DW_FORM_implicit_const)
-        read_uleb(abbrev);
+        read_uleb(&abbrev);
     }
   }
 
@@ -295,7 +295,7 @@ public:
 // Read value of the given DW_FORM_* form. If a value is not scalar,
 // returns a dummy value 0.
 template <typename E>
-inline u64 DebugInfoReader<E>::read(u64 form) {
+u64 DebugInfoReader<E>::read(u64 form) {
   switch (form) {
   case DW_FORM_flag_present:
     return 0;
@@ -348,7 +348,7 @@ inline u64 DebugInfoReader<E>::read(u64 form) {
   case DW_FORM_ref_udata:
   case DW_FORM_loclistx:
   case DW_FORM_rnglistx:
-    return read_uleb(cu);
+    return read_uleb(&cu);
   case DW_FORM_string:
     cu += strlen((char *)cu) + 1;
     return 0;
@@ -391,19 +391,19 @@ read_rnglist_range(Context<E> &ctx, ObjectFile<E> &file, u8 *rnglist,
     case DW_RLE_end_of_list:
       return vec;
     case DW_RLE_base_addressx:
-      base = addrx[read_uleb(rnglist)];
+      base = addrx[read_uleb(&rnglist)];
       break;
     case DW_RLE_startx_endx:
-      vec.push_back(addrx[read_uleb(rnglist)]);
-      vec.push_back(addrx[read_uleb(rnglist)]);
+      vec.push_back(addrx[read_uleb(&rnglist)]);
+      vec.push_back(addrx[read_uleb(&rnglist)]);
       break;
     case DW_RLE_startx_length:
-      vec.push_back(addrx[read_uleb(rnglist)]);
-      vec.push_back(vec.back() + read_uleb(rnglist));
+      vec.push_back(addrx[read_uleb(&rnglist)]);
+      vec.push_back(vec.back() + read_uleb(&rnglist));
       break;
     case DW_RLE_offset_pair:
-      vec.push_back(base + read_uleb(rnglist));
-      vec.push_back(base + read_uleb(rnglist));
+      vec.push_back(base + read_uleb(&rnglist));
+      vec.push_back(base + read_uleb(&rnglist));
       break;
     case DW_RLE_base_address:
       base = *(Word<E> *)rnglist;
@@ -418,7 +418,7 @@ read_rnglist_range(Context<E> &ctx, ObjectFile<E> &file, u8 *rnglist,
     case DW_RLE_start_length:
       vec.push_back(*(Word<E> *)rnglist);
       rnglist += sizeof(Word<E>);
-      vec.push_back(vec.back() + read_uleb(rnglist));
+      vec.push_back(vec.back() + read_uleb(&rnglist));
       break;
     }
   }
@@ -455,8 +455,8 @@ read_address_areas(Context<E> &ctx, ObjectFile<E> &file, i64 offset) {
 
   // Read all interesting debug records.
   for (;;) {
-    u64 name = read_uleb(abbrev);
-    u64 form = read_uleb(abbrev);
+    u64 name = read_uleb(&abbrev);
+    u64 form = read_uleb(&abbrev);
     if (name == 0 && form == 0)
       break;
 

elf/elf.h

@@ -1,8 +1,8 @@
 #pragma once
 
-#include "../integers.h"
+#include "../common/integers.h"
 
-#include <cstdint>
+#include <concepts>
 #include <ostream>
 #include <string>
 #include <type_traits>
@@ -25,6 +25,8 @@ struct SPARC64;
 struct M68K;
 struct SH4;
 struct ALPHA;
+struct LOONGARCH64;
+struct LOONGARCH32;
 
 template <typename E> struct ElfSym;
 template <typename E> struct ElfShdr;
@@ -54,9 +56,9 @@ enum : u32 {
 
 enum : u32 {
   SHN_UNDEF = 0,
+  SHN_LORESERVE = 0xff00,
   SHN_ABS = 0xfff1,
   SHN_COMMON = 0xfff2,
-  SHN_LORESERVE = 0xff00,
   SHN_XINDEX = 0xffff,
 };
 
@@ -87,6 +89,7 @@ enum : u32 {
   SHT_X86_64_UNWIND = 0x70000001,
   SHT_ARM_EXIDX = 0x70000001,
   SHT_ARM_ATTRIBUTES = 0x70000003,
+  SHT_RISCV_ATTRIBUTES = 0x70000003,
 };
 
 enum : u32 {
@@ -182,7 +185,9 @@ enum : u32 {
   PT_GNU_EH_FRAME = 0x6474e550,
   PT_GNU_STACK = 0x6474e551,
   PT_GNU_RELRO = 0x6474e552,
+  PT_OPENBSD_RANDOMIZE = 0x65a3dbe6,
   PT_ARM_EXIDX = 0x70000001,
+  PT_RISCV_ATTRIBUTES = 0x70000003,
 };
 
 enum : u32 {
@@ -226,6 +231,7 @@ enum : u32 {
   EM_X86_64 = 62,
   EM_AARCH64 = 183,
   EM_RISCV = 243,
+  EM_LOONGARCH = 258,
   EM_ALPHA = 0x9026,
 };
 
@@ -370,6 +376,24 @@ enum : u32 {
   STO_ALPHA_STD_GPLOAD = 0x22,
 };
 
+enum : u32 {
+  ELF_TAG_FILE = 1,
+  ELF_TAG_SECTION = 2,
+  ELF_TAG_SYMBOL = 3,
+  ELF_TAG_RISCV_STACK_ALIGN = 4,
+  ELF_TAG_RISCV_ARCH = 5,
+  ELF_TAG_RISCV_UNALIGNED_ACCESS = 6,
+};
+
+enum : u32 {
+  EF_LOONGARCH_ABI_SOFT_FLOAT = 0x1,
+  EF_LOONGARCH_ABI_SINGLE_FLOAT = 0x2,
+  EF_LOONGARCH_ABI_DOUBLE_FLOAT = 0x3,
+  EF_LOONGARCH_ABI_MODIFIER_MASK = 0x7,
+  EF_LOONGARCH_OBJABI_V1 = 0x40,
+  EF_LOONGARCH_OBJABI_MASK = 0xC0,
+};
+
 //
 // Relocation types
 //
@@ -720,6 +744,7 @@ enum : u32 {
   R_RISCV_TLS_DTPREL64 = 9,
   R_RISCV_TLS_TPREL32 = 10,
   R_RISCV_TLS_TPREL64 = 11,
+  R_RISCV_TLSDESC = 12,
   R_RISCV_BRANCH = 16,
   R_RISCV_JAL = 17,
   R_RISCV_CALL = 18,
@@ -757,6 +782,13 @@ enum : u32 {
   R_RISCV_SET32 = 56,
   R_RISCV_32_PCREL = 57,
   R_RISCV_IRELATIVE = 58,
+  R_RISCV_PLT32 = 59,
+  R_RISCV_SET_ULEB128 = 60,
+  R_RISCV_SUB_ULEB128 = 61,
+  R_RISCV_TLSDESC_HI20 = 62,
+  R_RISCV_TLSDESC_LOAD_LO12 = 63,
+  R_RISCV_TLSDESC_ADD_LO12 = 64,
+  R_RISCV_TLSDESC_CALL = 65,
 };
 
 enum : u32 {
@@ -933,9 +965,13 @@ enum : u32 {
   R_PPC64_REL24_NOTOC = 116,
   R_PPC64_PLTSEQ = 119,
   R_PPC64_PLTCALL = 120,
+  R_PPC64_PLTSEQ_NOTOC = 121,
+  R_PPC64_PLTCALL_NOTOC = 122,
   R_PPC64_PCREL_OPT = 123,
   R_PPC64_PCREL34 = 132,
   R_PPC64_GOT_PCREL34 = 133,
+  R_PPC64_PLT_PCREL34 = 134,
+  R_PPC64_PLT_PCREL34_NOTOC = 135,
   R_PPC64_TPREL34 = 146,
   R_PPC64_DTPREL34 = 147,
   R_PPC64_GOT_TLSGD_PCREL34 = 148,
@@ -1217,6 +1253,107 @@ enum : u32 {
   R_ALPHA_TPREL16 = 41,
 };
 
+enum : u32 {
+  R_LARCH_NONE = 0,
+  R_LARCH_32 = 1,
+  R_LARCH_64 = 2,
+  R_LARCH_RELATIVE = 3,
+  R_LARCH_COPY = 4,
+  R_LARCH_JUMP_SLOT = 5,
+  R_LARCH_TLS_DTPMOD32 = 6,
+  R_LARCH_TLS_DTPMOD64 = 7,
+  R_LARCH_TLS_DTPREL32 = 8,
+  R_LARCH_TLS_DTPREL64 = 9,
+  R_LARCH_TLS_TPREL32 = 10,
+  R_LARCH_TLS_TPREL64 = 11,
+  R_LARCH_IRELATIVE = 12,
+  R_LARCH_MARK_LA = 20,
+  R_LARCH_MARK_PCREL = 21,
+  R_LARCH_SOP_PUSH_PCREL = 22,
+  R_LARCH_SOP_PUSH_ABSOLUTE = 23,
+  R_LARCH_SOP_PUSH_DUP = 24,
+  R_LARCH_SOP_PUSH_GPREL = 25,
+  R_LARCH_SOP_PUSH_TLS_TPREL = 26,
+  R_LARCH_SOP_PUSH_TLS_GOT = 27,
+  R_LARCH_SOP_PUSH_TLS_GD = 28,
+  R_LARCH_SOP_PUSH_PLT_PCREL = 29,
+  R_LARCH_SOP_ASSERT = 30,
+  R_LARCH_SOP_NOT = 31,
+  R_LARCH_SOP_SUB = 32,
+  R_LARCH_SOP_SL = 33,
+  R_LARCH_SOP_SR = 34,
+  R_LARCH_SOP_ADD = 35,
+  R_LARCH_SOP_AND = 36,
+  R_LARCH_SOP_IF_ELSE = 37,
+  R_LARCH_SOP_POP_32_S_10_5 = 38,
+  R_LARCH_SOP_POP_32_U_10_12 = 39,
+  R_LARCH_SOP_POP_32_S_10_12 = 40,
+  R_LARCH_SOP_POP_32_S_10_16 = 41,
+  R_LARCH_SOP_POP_32_S_10_16_S2 = 42,
+  R_LARCH_SOP_POP_32_S_5_20 = 43,
+  R_LARCH_SOP_POP_32_S_0_5_10_16_S2 = 44,
+  R_LARCH_SOP_POP_32_S_0_10_10_16_S2 = 45,
+  R_LARCH_SOP_POP_32_U = 46,
+  R_LARCH_ADD8 = 47,
+  R_LARCH_ADD16 = 48,
+  R_LARCH_ADD24 = 49,
+  R_LARCH_ADD32 = 50,
+  R_LARCH_ADD64 = 51,
+  R_LARCH_SUB8 = 52,
+  R_LARCH_SUB16 = 53,
+  R_LARCH_SUB24 = 54,
+  R_LARCH_SUB32 = 55,
+  R_LARCH_SUB64 = 56,
+  R_LARCH_GNU_VTINHERIT = 57,
+  R_LARCH_GNU_VTENTRY = 58,
+  R_LARCH_B16 = 64,
+  R_LARCH_B21 = 65,
+  R_LARCH_B26 = 66,
+  R_LARCH_ABS_HI20 = 67,
+  R_LARCH_ABS_LO12 = 68,
+  R_LARCH_ABS64_LO20 = 69,
+  R_LARCH_ABS64_HI12 = 70,
+  R_LARCH_PCALA_HI20 = 71,
+  R_LARCH_PCALA_LO12 = 72,
+  R_LARCH_PCALA64_LO20 = 73,
+  R_LARCH_PCALA64_HI12 = 74,
+  R_LARCH_GOT_PC_HI20 = 75,
+  R_LARCH_GOT_PC_LO12 = 76,
+  R_LARCH_GOT64_PC_LO20 = 77,
+  R_LARCH_GOT64_PC_HI12 = 78,
+  R_LARCH_GOT_HI20 = 79,
+  R_LARCH_GOT_LO12 = 80,
+  R_LARCH_GOT64_LO20 = 81,
+  R_LARCH_GOT64_HI12 = 82,
+  R_LARCH_TLS_LE_HI20 = 83,
+  R_LARCH_TLS_LE_LO12 = 84,
+  R_LARCH_TLS_LE64_LO20 = 85,
+  R_LARCH_TLS_LE64_HI12 = 86,
+  R_LARCH_TLS_IE_PC_HI20 = 87,
+  R_LARCH_TLS_IE_PC_LO12 = 88,
+  R_LARCH_TLS_IE64_PC_LO20 = 89,
+  R_LARCH_TLS_IE64_PC_HI12 = 90,
+  R_LARCH_TLS_IE_HI20 = 91,
+  R_LARCH_TLS_IE_LO12 = 92,
+  R_LARCH_TLS_IE64_LO20 = 93,
+  R_LARCH_TLS_IE64_HI12 = 94,
+  R_LARCH_TLS_LD_PC_HI20 = 95,
+  R_LARCH_TLS_LD_HI20 = 96,
+  R_LARCH_TLS_GD_PC_HI20 = 97,
+  R_LARCH_TLS_GD_HI20 = 98,
+  R_LARCH_32_PCREL = 99,
+  R_LARCH_RELAX = 100,
+  R_LARCH_DELETE = 101,
+  R_LARCH_ALIGN = 102,
+  R_LARCH_PCREL20_S2 = 103,
+  R_LARCH_CFA = 104,
+  R_LARCH_ADD6 = 105,
+  R_LARCH_SUB6 = 106,
+  R_LARCH_ADD_ULEB128 = 107,
+  R_LARCH_SUB_ULEB128 = 108,
+  R_LARCH_64_PCREL = 109,
+};
+
 //
 // DWARF data types
 //
@@ -1453,14 +1590,8 @@ struct ElfPhdr<E> {
 // with a newly computed value.
 //
 // We don't want to have too many `if (REL)`s and `if (RELA)`s in our
-// codebase, so we write dynamic relocations in the following manner:
-//
-// - We always create a dynamic relocation with an addend. If it's REL,
-//   the addend will be discarded.
-//
-// - We also always write an addend to the relocated place by default
-//   even though it's redundant for RELA. If RELA, the written value
-//   will be ovewritten by the dynamic linker at load-time.
+// codebase, so ElfRel always takes r_addend as a constructor argument.
+// If it's REL, the argument will simply be ignored.
 template <typename E> requires E::is_le && E::is_rela
 struct ElfRel<E> {
   ElfRel() = default;
@@ -1582,6 +1713,9 @@ struct ElfSym<PPC64V2> {
   bool is_weak() const { return st_bind == STB_WEAK; }
   bool is_undef_weak() const { return is_undef() && is_weak(); }
 
+  bool preserves_r2() const { return ppc_local_entry != 1; }
+  bool uses_toc() const { return ppc_local_entry > 1; }
+
   ul32 st_name;
 
 #ifdef __LITTLE_ENDIAN__
@@ -1648,37 +1782,38 @@ struct ElfRel<SPARC64> {
 // Machine descriptions
 //
 
-template <typename E>
-static constexpr bool supports_tlsdesc = requires { E::R_TLSDESC; };
+template <typename E> concept supports_ifunc = requires { E::R_IRELATIVE; };
+template <typename E> concept supports_tlsdesc = requires { E::R_TLSDESC; };
+template <typename E> concept needs_thunk = requires { E::thunk_size; };
 
-template <typename E>
-static constexpr bool needs_thunk = requires { E::thunk_size; };
+template <typename E> concept is_x86_64 = std::same_as<E, X86_64>;
+template <typename E> concept is_i386 = std::same_as<E, I386>;
+template <typename E> concept is_arm64 = std::same_as<E, ARM64>;
+template <typename E> concept is_arm32 = std::same_as<E, ARM32>;
+template <typename E> concept is_rv64le = std::same_as<E, RV64LE>;
+template <typename E> concept is_rv64be = std::same_as<E, RV64BE>;
+template <typename E> concept is_rv32le = std::same_as<E, RV32LE>;
+template <typename E> concept is_rv32be = std::same_as<E, RV32BE>;
+template <typename E> concept is_ppc32 = std::same_as<E, PPC32>;
+template <typename E> concept is_ppc64v1 = std::same_as<E, PPC64V1>;
+template <typename E> concept is_ppc64v2 = std::same_as<E, PPC64V2>;
+template <typename E> concept is_s390x = std::same_as<E, S390X>;
+template <typename E> concept is_sparc64 = std::same_as<E, SPARC64>;
+template <typename E> concept is_m68k = std::same_as<E, M68K>;
+template <typename E> concept is_sh4 = std::same_as<E, SH4>;
+template <typename E> concept is_alpha = std::same_as<E, ALPHA>;
+template <typename E> concept is_loongarch64 = std::same_as<E, LOONGARCH64>;
+template <typename E> concept is_loongarch32 = std::same_as<E, LOONGARCH32>;
 
-template <typename E> static constexpr bool is_x86_64 = std::is_same_v<E, X86_64>;
-template <typename E> static constexpr bool is_i386 = std::is_same_v<E, I386>;
-template <typename E> static constexpr bool is_arm64 = std::is_same_v<E, ARM64>;
-template <typename E> static constexpr bool is_arm32 = std::is_same_v<E, ARM32>;
-template <typename E> static constexpr bool is_rv64le = std::is_same_v<E, RV64LE>;
-template <typename E> static constexpr bool is_rv64be = std::is_same_v<E, RV64BE>;
-template <typename E> static constexpr bool is_rv32le = std::is_same_v<E, RV32LE>;
-template <typename E> static constexpr bool is_rv32be = std::is_same_v<E, RV32BE>;
-template <typename E> static constexpr bool is_ppc32 = std::is_same_v<E, PPC32>;
-template <typename E> static constexpr bool is_ppc64v1 = std::is_same_v<E, PPC64V1>;
-template <typename E> static constexpr bool is_ppc64v2 = std::is_same_v<E, PPC64V2>;
-template <typename E> static constexpr bool is_s390x = std::is_same_v<E, S390X>;
-template <typename E> static constexpr bool is_sparc64 = std::is_same_v<E, SPARC64>;
-template <typename E> static constexpr bool is_m68k = std::is_same_v<E, M68K>;
-template <typename E> static constexpr bool is_sh4 = std::is_same_v<E, SH4>;
-template <typename E> static constexpr bool is_alpha = std::is_same_v<E, ALPHA>;
-
-template <typename E> static constexpr bool is_x86 = is_x86_64<E> || is_i386<E>;
-template <typename E> static constexpr bool is_arm = is_arm64<E> || is_arm32<E>;
-template <typename E> static constexpr bool is_rv64 = is_rv64le<E> || is_rv64be<E>;
-template <typename E> static constexpr bool is_rv32 = is_rv32le<E> || is_rv32be<E>;
-template <typename E> static constexpr bool is_riscv = is_rv64<E> || is_rv32<E>;
-template <typename E> static constexpr bool is_ppc64 = is_ppc64v1<E> || is_ppc64v2<E>;
-template <typename E> static constexpr bool is_ppc = is_ppc64<E> || is_ppc32<E>;
-template <typename E> static constexpr bool is_sparc = is_sparc64<E>;
+template <typename E> concept is_x86 = is_x86_64<E> || is_i386<E>;
+template <typename E> concept is_arm = is_arm64<E> || is_arm32<E>;
+template <typename E> concept is_rv64 = is_rv64le<E> || is_rv64be<E>;
+template <typename E> concept is_rv32 = is_rv32le<E> || is_rv32be<E>;
+template <typename E> concept is_riscv = is_rv64<E> || is_rv32<E>;
+template <typename E> concept is_ppc64 = is_ppc64v1<E> || is_ppc64v2<E>;
+template <typename E> concept is_ppc = is_ppc64<E> || is_ppc32<E>;
+template <typename E> concept is_sparc = is_sparc64<E>;
+template <typename E> concept is_loongarch = is_loongarch64<E> || is_loongarch32<E>;
 
 struct X86_64 {
   static constexpr std::string_view target_name = "x86_64";
@@ -1695,6 +1830,7 @@ struct X86_64 {
   static constexpr u32 R_GLOB_DAT = R_X86_64_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_X86_64_JUMP_SLOT;
   static constexpr u32 R_ABS = R_X86_64_64;
+  static constexpr u32 R_DYNAMIC = R_X86_64_64;
   static constexpr u32 R_RELATIVE = R_X86_64_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_X86_64_IRELATIVE;
   static constexpr u32 R_DTPOFF = R_X86_64_DTPOFF64;
@@ -1718,6 +1854,7 @@ struct I386 {
   static constexpr u32 R_GLOB_DAT = R_386_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_386_JUMP_SLOT;
   static constexpr u32 R_ABS = R_386_32;
+  static constexpr u32 R_DYNAMIC = R_386_32;
   static constexpr u32 R_RELATIVE = R_386_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_386_IRELATIVE;
   static constexpr u32 R_DTPOFF = R_386_TLS_DTPOFF32;
@@ -1737,12 +1874,13 @@ struct ARM64 {
   static constexpr u32 plt_size = 16;
   static constexpr u32 pltgot_size = 16;
   static constexpr u32 thunk_hdr_size = 0;
-  static constexpr u32 thunk_size = 12;
+  static constexpr u32 thunk_size = 16;
 
   static constexpr u32 R_COPY = R_AARCH64_COPY;
   static constexpr u32 R_GLOB_DAT = R_AARCH64_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_AARCH64_JUMP_SLOT;
   static constexpr u32 R_ABS = R_AARCH64_ABS64;
+  static constexpr u32 R_DYNAMIC = R_AARCH64_ABS64;
   static constexpr u32 R_RELATIVE = R_AARCH64_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_AARCH64_IRELATIVE;
   static constexpr u32 R_DTPOFF = R_AARCH64_TLS_DTPREL64;
@@ -1761,13 +1899,14 @@ struct ARM32 {
   static constexpr u32 plt_hdr_size = 32;
   static constexpr u32 plt_size = 16;
   static constexpr u32 pltgot_size = 16;
-  static constexpr u32 thunk_hdr_size = 12;
-  static constexpr u32 thunk_size = 20;
+  static constexpr u32 thunk_hdr_size = 16;
+  static constexpr u32 thunk_size = 16;
 
   static constexpr u32 R_COPY = R_ARM_COPY;
   static constexpr u32 R_GLOB_DAT = R_ARM_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_ARM_JUMP_SLOT;
   static constexpr u32 R_ABS = R_ARM_ABS32;
+  static constexpr u32 R_DYNAMIC = R_ARM_ABS32;
   static constexpr u32 R_RELATIVE = R_ARM_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_ARM_IRELATIVE;
   static constexpr u32 R_DTPOFF = R_ARM_TLS_DTPOFF32;
@@ -1776,32 +1915,40 @@ struct ARM32 {
   static constexpr u32 R_TLSDESC = R_ARM_TLS_DESC;
 };
 
-struct RV64LE {
+struct RV64 {
+  static constexpr bool is_64 = true;
+  static constexpr bool is_rela = true;
+  static constexpr u32 page_size = 4096;
+  static constexpr u32 e_machine = EM_RISCV;
+  static constexpr u32 plt_hdr_size = 32;
+  static constexpr u32 plt_size = 16;
+  static constexpr u32 pltgot_size = 16;
+
+  static constexpr u32 R_COPY = R_RISCV_COPY;
+  static constexpr u32 R_GLOB_DAT = R_RISCV_64;
+  static constexpr u32 R_JUMP_SLOT = R_RISCV_JUMP_SLOT;
+  static constexpr u32 R_ABS = R_RISCV_64;
+  static constexpr u32 R_DYNAMIC = R_RISCV_64;
+  static constexpr u32 R_RELATIVE = R_RISCV_RELATIVE;
+  static constexpr u32 R_IRELATIVE = R_RISCV_IRELATIVE;
+  static constexpr u32 R_DTPOFF = R_RISCV_TLS_DTPREL64;
+  static constexpr u32 R_TPOFF = R_RISCV_TLS_TPREL64;
+  static constexpr u32 R_DTPMOD = R_RISCV_TLS_DTPMOD64;
+  static constexpr u32 R_TLSDESC = R_RISCV_TLSDESC;
+};
+
+struct RV64LE : RV64 {
   static constexpr std::string_view target_name = "riscv64";
-  static constexpr bool is_64 = true;
   static constexpr bool is_le = true;
-  static constexpr bool is_rela = true;
-  static constexpr u32 page_size = 4096;
-  static constexpr u32 e_machine = EM_RISCV;
-  static constexpr u32 plt_hdr_size = 32;
-  static constexpr u32 plt_size = 16;
-  static constexpr u32 pltgot_size = 16;
-
-  static constexpr u32 R_COPY = R_RISCV_COPY;
-  static constexpr u32 R_GLOB_DAT = R_RISCV_64;
-  static constexpr u32 R_JUMP_SLOT = R_RISCV_JUMP_SLOT;
-  static constexpr u32 R_ABS = R_RISCV_64;
-  static constexpr u32 R_RELATIVE = R_RISCV_RELATIVE;
-  static constexpr u32 R_IRELATIVE = R_RISCV_IRELATIVE;
-  static constexpr u32 R_DTPOFF = R_RISCV_TLS_DTPREL64;
-  static constexpr u32 R_TPOFF = R_RISCV_TLS_TPREL64;
-  static constexpr u32 R_DTPMOD = R_RISCV_TLS_DTPMOD64;
 };
 
-struct RV64BE {
+struct RV64BE : RV64 {
   static constexpr std::string_view target_name = "riscv64be";
-  static constexpr bool is_64 = true;
   static constexpr bool is_le = false;
+};
+
+struct RV32 {
+  static constexpr bool is_64 = false;
   static constexpr bool is_rela = true;
   static constexpr u32 page_size = 4096;
   static constexpr u32 e_machine = EM_RISCV;
@@ -1810,58 +1957,26 @@ struct RV64BE {
   static constexpr u32 pltgot_size = 16;
 
   static constexpr u32 R_COPY = R_RISCV_COPY;
-  static constexpr u32 R_GLOB_DAT = R_RISCV_64;
+  static constexpr u32 R_GLOB_DAT = R_RISCV_32;
   static constexpr u32 R_JUMP_SLOT = R_RISCV_JUMP_SLOT;
-  static constexpr u32 R_ABS = R_RISCV_64;
+  static constexpr u32 R_ABS = R_RISCV_32;
+  static constexpr u32 R_DYNAMIC = R_RISCV_32;
   static constexpr u32 R_RELATIVE = R_RISCV_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_RISCV_IRELATIVE;
-  static constexpr u32 R_DTPOFF = R_RISCV_TLS_DTPREL64;
-  static constexpr u32 R_TPOFF = R_RISCV_TLS_TPREL64;
-  static constexpr u32 R_DTPMOD = R_RISCV_TLS_DTPMOD64;
+  static constexpr u32 R_DTPOFF = R_RISCV_TLS_DTPREL32;
+  static constexpr u32 R_TPOFF = R_RISCV_TLS_TPREL32;
+  static constexpr u32 R_DTPMOD = R_RISCV_TLS_DTPMOD32;
+  static constexpr u32 R_TLSDESC = R_RISCV_TLSDESC;
 };
 
-struct RV32LE {
+struct RV32LE : RV32 {
   static constexpr std::string_view target_name = "riscv32";
-  static constexpr bool is_64 = false;
   static constexpr bool is_le = true;
-  static constexpr bool is_rela = true;
-  static constexpr u32 page_size = 4096;
-  static constexpr u32 e_machine = EM_RISCV;
-  static constexpr u32 plt_hdr_size = 32;
-  static constexpr u32 plt_size = 16;
-  static constexpr u32 pltgot_size = 16;
-
-  static constexpr u32 R_COPY = R_RISCV_COPY;
-  static constexpr u32 R_GLOB_DAT = R_RISCV_32;
-  static constexpr u32 R_JUMP_SLOT = R_RISCV_JUMP_SLOT;
-  static constexpr u32 R_ABS = R_RISCV_32;
-  static constexpr u32 R_RELATIVE = R_RISCV_RELATIVE;
-  static constexpr u32 R_IRELATIVE = R_RISCV_IRELATIVE;
-  static constexpr u32 R_DTPOFF = R_RISCV_TLS_DTPREL32;
-  static constexpr u32 R_TPOFF = R_RISCV_TLS_TPREL32;
-  static constexpr u32 R_DTPMOD = R_RISCV_TLS_DTPMOD32;
 };
 
-struct RV32BE {
+struct RV32BE : RV32 {
   static constexpr std::string_view target_name = "riscv32be";
-  static constexpr bool is_64 = false;
   static constexpr bool is_le = false;
-  static constexpr bool is_rela = true;
-  static constexpr u32 page_size = 4096;
-  static constexpr u32 e_machine = EM_RISCV;
-  static constexpr u32 plt_hdr_size = 32;
-  static constexpr u32 plt_size = 16;
-  static constexpr u32 pltgot_size = 16;
-
-  static constexpr u32 R_COPY = R_RISCV_COPY;
-  static constexpr u32 R_GLOB_DAT = R_RISCV_32;
-  static constexpr u32 R_JUMP_SLOT = R_RISCV_JUMP_SLOT;
-  static constexpr u32 R_ABS = R_RISCV_32;
-  static constexpr u32 R_RELATIVE = R_RISCV_RELATIVE;
-  static constexpr u32 R_IRELATIVE = R_RISCV_IRELATIVE;
-  static constexpr u32 R_DTPOFF = R_RISCV_TLS_DTPREL32;
-  static constexpr u32 R_TPOFF = R_RISCV_TLS_TPREL32;
-  static constexpr u32 R_DTPMOD = R_RISCV_TLS_DTPMOD32;
 };
 
 struct PPC32 {
@@ -1881,6 +1996,7 @@ struct PPC32 {
   static constexpr u32 R_GLOB_DAT = R_PPC_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_PPC_JMP_SLOT;
   static constexpr u32 R_ABS = R_PPC_ADDR32;
+  static constexpr u32 R_DYNAMIC = R_PPC_ADDR32;
   static constexpr u32 R_RELATIVE = R_PPC_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_PPC_IRELATIVE;
   static constexpr u32 R_DTPOFF = R_PPC_DTPREL32;
@@ -1888,23 +2004,17 @@ struct PPC32 {
   static constexpr u32 R_DTPMOD = R_PPC_DTPMOD32;
 };
 
-struct PPC64V1 {
-  static constexpr std::string_view target_name = "ppc64v1";
+struct PPC64 {
   static constexpr bool is_64 = true;
-  static constexpr bool is_le = false;
   static constexpr bool is_rela = true;
   static constexpr u32 page_size = 65536;
   static constexpr u32 e_machine = EM_PPC64;
-  static constexpr u32 plt_hdr_size = 52;
-  static constexpr u32 plt_size = 8;
-  static constexpr u32 pltgot_size = 0;
-  static constexpr u32 thunk_hdr_size = 0;
-  static constexpr u32 thunk_size = 28;
 
   static constexpr u32 R_COPY = R_PPC64_COPY;
   static constexpr u32 R_GLOB_DAT = R_PPC64_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_PPC64_JMP_SLOT;
   static constexpr u32 R_ABS = R_PPC64_ADDR64;
+  static constexpr u32 R_DYNAMIC = R_PPC64_ADDR64;
   static constexpr u32 R_RELATIVE = R_PPC64_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_PPC64_IRELATIVE;
   static constexpr u32 R_DTPOFF = R_PPC64_DTPREL64;
@@ -1912,28 +2022,23 @@ struct PPC64V1 {
   static constexpr u32 R_DTPMOD = R_PPC64_DTPMOD64;
 };
 
-struct PPC64V2 {
+struct PPC64V1 : PPC64 {
+  static constexpr std::string_view target_name = "ppc64v1";
+  static constexpr bool is_le = false;
+  static constexpr u32 plt_hdr_size = 44;
+  static constexpr u32 pltgot_size = 0;
+  static constexpr u32 thunk_hdr_size = 0;
+  static constexpr u32 thunk_size = 28;
+};
+
+struct PPC64V2 : PPC64 {
   static constexpr std::string_view target_name = "ppc64v2";
-  static constexpr bool is_64 = true;
   static constexpr bool is_le = true;
-  static constexpr bool is_rela = true;
-  static constexpr u32 page_size = 65536;
-  static constexpr u32 e_machine = EM_PPC64;
-  static constexpr u32 plt_hdr_size = 60;
+  static constexpr u32 plt_hdr_size = 52;
   static constexpr u32 plt_size = 4;
   static constexpr u32 pltgot_size = 0;
   static constexpr u32 thunk_hdr_size = 0;
   static constexpr u32 thunk_size = 20;
-
-  static constexpr u32 R_COPY = R_PPC64_COPY;
-  static constexpr u32 R_GLOB_DAT = R_PPC64_GLOB_DAT;
-  static constexpr u32 R_JUMP_SLOT = R_PPC64_JMP_SLOT;
-  static constexpr u32 R_ABS = R_PPC64_ADDR64;
-  static constexpr u32 R_RELATIVE = R_PPC64_RELATIVE;
-  static constexpr u32 R_IRELATIVE = R_PPC64_IRELATIVE;
-  static constexpr u32 R_DTPOFF = R_PPC64_DTPREL64;
-  static constexpr u32 R_TPOFF = R_PPC64_TPREL64;
-  static constexpr u32 R_DTPMOD = R_PPC64_DTPMOD64;
 };
 
 struct S390X {
@@ -1943,14 +2048,15 @@ struct S390X {
   static constexpr bool is_rela = true;
   static constexpr u32 page_size = 4096;
   static constexpr u32 e_machine = EM_S390X;
-  static constexpr u32 plt_hdr_size = 32;
-  static constexpr u32 plt_size = 32;
+  static constexpr u32 plt_hdr_size = 48;
+  static constexpr u32 plt_size = 16;
   static constexpr u32 pltgot_size = 16;
 
   static constexpr u32 R_COPY = R_390_COPY;
   static constexpr u32 R_GLOB_DAT = R_390_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_390_JMP_SLOT;
   static constexpr u32 R_ABS = R_390_64;
+  static constexpr u32 R_DYNAMIC = R_390_64;
   static constexpr u32 R_RELATIVE = R_390_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_390_IRELATIVE;
   static constexpr u32 R_DTPOFF = R_390_TLS_DTPOFF;
@@ -1973,6 +2079,7 @@ struct SPARC64 {
   static constexpr u32 R_GLOB_DAT = R_SPARC_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_SPARC_JMP_SLOT;
   static constexpr u32 R_ABS = R_SPARC_64;
+  static constexpr u32 R_DYNAMIC = R_SPARC_64;
   static constexpr u32 R_RELATIVE = R_SPARC_RELATIVE;
   static constexpr u32 R_IRELATIVE = R_SPARC_IRELATIVE;
   static constexpr u32 R_DTPOFF = R_SPARC_TLS_DTPOFF64;
@@ -1995,8 +2102,8 @@ struct M68K {
   static constexpr u32 R_GLOB_DAT = R_68K_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_68K_JMP_SLOT;
   static constexpr u32 R_ABS = R_68K_32;
+  static constexpr u32 R_DYNAMIC = R_68K_32;
   static constexpr u32 R_RELATIVE = R_68K_RELATIVE;
-  static constexpr u32 R_IRELATIVE = R_NONE;
   static constexpr u32 R_DTPOFF = R_68K_TLS_DTPREL32;
   static constexpr u32 R_TPOFF = R_68K_TLS_TPREL32;
   static constexpr u32 R_DTPMOD = R_68K_TLS_DTPMOD32;
@@ -2017,8 +2124,8 @@ struct SH4 {
   static constexpr u32 R_GLOB_DAT = R_SH_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_SH_JMP_SLOT;
   static constexpr u32 R_ABS = R_SH_DIR32;
+  static constexpr u32 R_DYNAMIC = R_SH_DIR32;
   static constexpr u32 R_RELATIVE = R_SH_RELATIVE;
-  static constexpr u32 R_IRELATIVE = R_NONE;
   static constexpr u32 R_DTPOFF = R_SH_TLS_DTPOFF32;
   static constexpr u32 R_TPOFF = R_SH_TLS_TPOFF32;
   static constexpr u32 R_DTPMOD = R_SH_TLS_DTPMOD32;
@@ -2039,11 +2146,61 @@ struct ALPHA {
   static constexpr u32 R_GLOB_DAT = R_ALPHA_GLOB_DAT;
   static constexpr u32 R_JUMP_SLOT = R_ALPHA_JMP_SLOT;
   static constexpr u32 R_ABS = R_ALPHA_REFQUAD;
+  static constexpr u32 R_DYNAMIC = R_ALPHA_REFQUAD;
   static constexpr u32 R_RELATIVE = R_ALPHA_RELATIVE;
-  static constexpr u32 R_IRELATIVE = R_NONE;
   static constexpr u32 R_DTPOFF = R_ALPHA_DTPREL64;
   static constexpr u32 R_TPOFF = R_ALPHA_TPREL64;
   static constexpr u32 R_DTPMOD = R_ALPHA_DTPMOD64;
 };
 
+struct LOONGARCH64 {
+  static constexpr std::string_view target_name = "loongarch64";
+  static constexpr bool is_64 = true;
+  static constexpr bool is_le = true;
+  static constexpr bool is_rela = true;
+  static constexpr u32 page_size = 16384;
+  static constexpr u32 e_machine = EM_LOONGARCH;
+  static constexpr u32 plt_hdr_size = 32;
+  static constexpr u32 plt_size = 16;
+  static constexpr u32 pltgot_size = 16;
+  static constexpr u32 thunk_hdr_size = 0;
+  static constexpr u32 thunk_size = 16;
+
+  static constexpr u32 R_COPY = R_LARCH_COPY;
+  static constexpr u32 R_GLOB_DAT = R_LARCH_64;
+  static constexpr u32 R_JUMP_SLOT = R_LARCH_JUMP_SLOT;
+  static constexpr u32 R_ABS = R_LARCH_64;
+  static constexpr u32 R_DYNAMIC = R_LARCH_64;
+  static constexpr u32 R_RELATIVE = R_LARCH_RELATIVE;
+  static constexpr u32 R_IRELATIVE = R_LARCH_IRELATIVE;
+  static constexpr u32 R_DTPOFF = R_LARCH_TLS_DTPREL64;
+  static constexpr u32 R_TPOFF = R_LARCH_TLS_TPREL64;
+  static constexpr u32 R_DTPMOD = R_LARCH_TLS_DTPMOD64;
+};
+
+struct LOONGARCH32 {
+  static constexpr std::string_view target_name = "loongarch32";
+  static constexpr bool is_64 = false;
+  static constexpr bool is_le = true;
+  static constexpr bool is_rela = true;
+  static constexpr u32 page_size = 16384;
+  static constexpr u32 e_machine = EM_LOONGARCH;
+  static constexpr u32 plt_hdr_size = 32;
+  static constexpr u32 plt_size = 16;
+  static constexpr u32 pltgot_size = 16;
+  static constexpr u32 thunk_hdr_size = 0;
+  static constexpr u32 thunk_size = 16;
+
+  static constexpr u32 R_COPY = R_LARCH_COPY;
+  static constexpr u32 R_GLOB_DAT = R_LARCH_32;
+  static constexpr u32 R_JUMP_SLOT = R_LARCH_JUMP_SLOT;
+  static constexpr u32 R_ABS = R_LARCH_32;
+  static constexpr u32 R_DYNAMIC = R_LARCH_32;
+  static constexpr u32 R_RELATIVE = R_LARCH_RELATIVE;
+  static constexpr u32 R_IRELATIVE = R_LARCH_IRELATIVE;
+  static constexpr u32 R_DTPOFF = R_LARCH_TLS_DTPREL32;
+  static constexpr u32 R_TPOFF = R_LARCH_TLS_TPREL32;
+  static constexpr u32 R_DTPMOD = R_LARCH_TLS_DTPMOD32;
+};
+
 } // namespace mold::elf

elf/gc-sections.cc

@@ -30,7 +30,7 @@ static bool should_keep(const InputSection<E> &isec) {
 
 template <typename E>
 static bool mark_section(InputSection<E> *isec) {
-  return isec && isec->is_alive && fast_mark(isec->is_visited);
+  return isec && isec->is_alive && !isec->is_visited.test_and_set();
 }
 
 template <typename E>
@@ -53,19 +53,18 @@ static void visit(Context<E> &ctx, InputSection<E> *isec,
     // Symbol can refer either a section fragment or an input section.
     // Mark a fragment as alive.
     if (SectionFragment<E> *frag = sym.get_frag()) {
-      frag->is_alive.store(true, std::memory_order_relaxed);
+      frag->is_alive = true;
       continue;
     }
 
-    if (!mark_section(sym.get_input_section()))
-      continue;
-
     // Mark a section alive. For better performacne, we don't call
     // `feeder.add` too often.
-    if (depth < 3)
-      visit(ctx, sym.get_input_section(), feeder, depth + 1);
-    else
-      feeder.add(sym.get_input_section());
+    if (mark_section(sym.get_input_section())) {
+      if (depth < 3)
+        visit(ctx, sym.get_input_section(), feeder, depth + 1);
+      else
+        feeder.add(sym.get_input_section());
+    }
   }
 }
 
@@ -82,7 +81,7 @@ static void collect_root_set(Context<E> &ctx,
   auto enqueue_symbol = [&](Symbol<E> *sym) {
     if (sym) {
       if (SectionFragment<E> *frag = sym->get_frag())
-        frag->is_alive.store(true, std::memory_order_relaxed);
+        frag->is_alive = true;
       else
         enqueue_section(sym->get_input_section());
     }
@@ -94,10 +93,10 @@ static void collect_root_set(Context<E> &ctx,
       if (!isec || !isec->is_alive)
         continue;
 
-      // -gc-sections discards only SHF_ALLOC sections. If you want to
+      // --gc-sections discards only SHF_ALLOC sections. If you want to
       // reduce the amount of non-memory-mapped segments, you should
       // use `strip` command, compile without debug info or use
-      // -strip-all linker option.
+      // --strip-all linker option.
       u32 flags = isec->shdr().sh_flags;
       if (!(flags & SHF_ALLOC))
         isec->is_visited = true;
@@ -163,27 +162,10 @@ static void sweep(Context<E> &ctx) {
   });
 }
 
-// Non-alloc section fragments are not subject of garbage collection.
-// This function marks such fragments.
-template <typename E>
-static void mark_nonalloc_fragments(Context<E> &ctx) {
-  Timer t(ctx, "mark_nonalloc_fragments");
-
-  tbb::parallel_for_each(ctx.objs, [](ObjectFile<E> *file) {
-    for (std::unique_ptr<MergeableSection<E>> &m : file->mergeable_sections)
-      if (m)
-        for (SectionFragment<E> *frag : m->fragments)
-          if (!(frag->output_section.shdr.sh_flags & SHF_ALLOC))
-            frag->is_alive.store(true, std::memory_order_relaxed);
-  });
-}
-
 template <typename E>
 void gc_sections(Context<E> &ctx) {
   Timer t(ctx, "gc");
 
-  mark_nonalloc_fragments(ctx);
-
   tbb::concurrent_vector<InputSection<E> *> rootset;
   collect_root_set(ctx, rootset);
   mark(ctx, rootset);

elf/icf.cc

@@ -65,7 +65,7 @@
 // conditions.
 
 #include "mold.h"
-#include "../sha.h"
+#include "blake3.h"
 
 #include <array>
 #include <cstdio>
@@ -81,7 +81,7 @@ static constexpr int64_t HASH_SIZE = 16;
 typedef std::array<uint8_t, HASH_SIZE> Digest;
 
 namespace std {
-template<> struct hash<Digest> {
+template <> struct hash<Digest> {
   size_t operator()(const Digest &k) const {
     return *(int64_t *)&k[0];
   }
@@ -90,16 +90,6 @@ template<> struct hash<Digest> {
 
 namespace mold::elf {
 
-class BitVector {
-public:
-  BitVector(u32 size) : vec((size + 7) / 8) {}
-  bool get(u32 idx) const { return vec[idx / 8] & (1 << (idx % 8)); }
-  void set(u32 idx) { vec[idx / 8] |= 1 << (idx % 8); }
-
-private:
-  std::vector<u8> vec;
-};
-
 template <typename E>
 static void uniquify_cies(Context<E> &ctx) {
   Timer t(ctx, "uniquify_cies");
@@ -142,9 +132,11 @@ static bool is_eligible(Context<E> &ctx, InputSection<E> &isec) {
          !is_init && !is_fini && !is_enumerable && !is_addr_taken;
 }
 
-static Digest digest_final(SHA256Hash &sha) {
-  u8 buf[SHA256_SIZE];
-  sha.finish(buf);
+static Digest digest_final(blake3_hasher *hasher) {
+  assert(HASH_SIZE <= BLAKE3_OUT_LEN);
+
+  u8 buf[BLAKE3_OUT_LEN];
+  blake3_hasher_finalize(hasher, buf, BLAKE3_OUT_LEN);
 
   Digest digest;
   memcpy(digest.data(), buf, HASH_SIZE);
@@ -244,15 +236,16 @@ static void merge_leaf_nodes(Context<E> &ctx) {
 
 template <typename E>
 static Digest compute_digest(Context<E> &ctx, InputSection<E> &isec) {
-  SHA256Hash sha;
+  blake3_hasher hasher;
+  blake3_hasher_init(&hasher);
 
   auto hash = [&](auto val) {
-    sha.update((u8 *)&val, sizeof(val));
+    blake3_hasher_update(&hasher, (u8 *)&val, sizeof(val));
   };
 
   auto hash_string = [&](std::string_view str) {
     hash(str.size());
-    sha.update((u8 *)str.data(), str.size());
+    blake3_hasher_update(&hasher, (u8 *)str.data(), str.size());
   };
 
   auto hash_symbol = [&](Symbol<E> &sym) {
@@ -308,7 +301,7 @@ static Digest compute_digest(Context<E> &ctx, InputSection<E> &isec) {
     hash_symbol(*isec.file.symbols[rel.r_sym]);
   }
 
-  return digest_final(sha);
+  return digest_final(&hasher);
 }
 
 template <typename E>
@@ -421,16 +414,17 @@ static i64 propagate(std::span<std::vector<Digest>> digests,
     if (converged.get(i))
       return;
 
-    SHA256Hash sha;
-    sha.update(digests[2][i].data(), HASH_SIZE);
+    blake3_hasher hasher;
+    blake3_hasher_init(&hasher);
+    blake3_hasher_update(&hasher, digests[2][i].data(), HASH_SIZE);
 
     i64 begin = edge_indices[i];
     i64 end = (i + 1 == num_digests) ? edges.size() : edge_indices[i + 1];
 
     for (i64 j : edges.subspan(begin, end - begin))
-      sha.update(digests[slot][j].data(), HASH_SIZE);
+      blake3_hasher_update(&hasher, digests[slot][j].data(), HASH_SIZE);
 
-    digests[!slot][i] = digest_final(sha);
+    digests[!slot][i] = digest_final(&hasher);
 
     if (digests[slot][i] == digests[!slot][i]) {
       // This node has converged. Skip further iterations as it will
@@ -573,7 +567,7 @@ void icf_sections(Context<E> &ctx) {
     }
   }
 
-  // Group sections by SHA digest.
+  // Group sections by BLAKE3 digest.
   {
     Timer t(ctx, "group");
 

elf/input-files.cc

@@ -42,6 +42,12 @@ InputFile<E>::InputFile(Context<E> &ctx, MappedFile<Context<E>> *mf)
   shstrtab = this->get_string(ctx, shstrtab_idx);
 }
 
+template <typename E>
+std::span<ElfPhdr<E>> InputFile<E>::get_phdrs() {
+  ElfEhdr<E> &ehdr = get_ehdr();
+  return {(ElfPhdr<E> *)(mf->data + ehdr.e_phoff), ehdr.e_phnum};
+}
+
 template <typename E>
 ElfShdr<E> *InputFile<E>::find_section(i64 type) {
   for (ElfShdr<E> &sec : elf_sections)
@@ -50,21 +56,20 @@ ElfShdr<E> *InputFile<E>::find_section(i64 type) {
   return nullptr;
 }
 
-// This function is actually racy, in that "tearing" might be observed
-// during the read or write of `sym->file`. However, on most platforms
-// read/write of pointers never tears, and even if it did, it would be
-// harmless, as `sym->file` will not become equal to `sym`, or even if
-// it did by chance, we just wipe a symbol twice.
-//
-// Doing this in a standard-compliant way is too cumbersome, hence
-// here we just pretend a compiler will never optimize this into a
-// form that actually exploits the UB.
 template <typename E>
-__attribute__((no_sanitize("thread")))
 void InputFile<E>::clear_symbols() {
-  for (Symbol<E> *sym : get_global_syms())
-    if (sym->file == this)
-      sym->clear();
+  for (Symbol<E> *sym : get_global_syms()) {
+    if (__atomic_load_n(&sym->file, __ATOMIC_ACQUIRE) == this) {
+      sym->origin = 0;
+      sym->value = -1;
+      sym->sym_idx = -1;
+      sym->ver_idx = 0;
+      sym->is_weak = false;
+      sym->is_imported = false;
+      sym->is_exported = false;
+      __atomic_store_n(&sym->file, nullptr, __ATOMIC_RELEASE);
+    }
+  }
 }
 
 // Find the source filename. It should be listed in symtab as STT_FILE.
@@ -99,7 +104,7 @@ static bool is_debug_section(const ElfShdr<E> &shdr, std::string_view name) {
 
 template <typename E>
 void
-ObjectFile<E>::read_note_gnu_property(Context<E> &ctx, const ElfShdr<E> &shdr) {
+ObjectFile<E>::parse_note_gnu_property(Context<E> &ctx, const ElfShdr<E> &shdr) {
   std::string_view data = this->get_string(ctx, shdr);
 
   while (!data.empty()) {
@@ -135,6 +140,57 @@ ObjectFile<E>::read_note_gnu_property(Context<E> &ctx, const ElfShdr<E> &shdr) {
   }
 }
 
+// <format-version>
+// [ <section-length> "vendor-name" <file-tag> <size> <attribute>*]+ ]*
+template <typename E>
+static void read_riscv_attributes(Context<E> &ctx, ObjectFile<E> &file,
+                                  std::string_view data) {
+  if (data.empty())
+    Fatal(ctx) << file << ": corrupted .riscv.attributes section";
+
+  if (u8 format_version = data[0]; format_version != 'A')
+    return;
+  data = data.substr(1);
+
+  while (!data.empty()) {
+    i64 sz = *(U32<E> *)data.data();
+    if (data.size() < sz)
+      Fatal(ctx) << file << ": corrupted .riscv.attributes section";
+
+    std::string_view p(data.data() + 4, sz - 4);
+    data = data.substr(sz);
+
+    if (!p.starts_with("riscv\0"sv))
+      continue;
+    p = p.substr(6);
+
+    if (!p.starts_with(ELF_TAG_FILE))
+      Fatal(ctx) << file << ": corrupted .riscv.attributes section";
+    p = p.substr(5); // skip the tag and the sub-sub-section size
+
+    while (!p.empty()) {
+      i64 tag = read_uleb(&p);
+
+      switch (tag) {
+      case ELF_TAG_RISCV_STACK_ALIGN:
+        file.extra.stack_align = read_uleb(&p);
+        break;
+      case ELF_TAG_RISCV_ARCH: {
+        i64 pos = p.find_first_of('\0');
+        file.extra.arch = p.substr(0, pos);
+        p = p.substr(pos + 1);
+        break;
+      }
+      case ELF_TAG_RISCV_UNALIGNED_ACCESS:
+        file.extra.unaligned_access = read_uleb(&p);
+        break;
+      default:
+        break;
+      }
+    }
+  }
+}
+
 template <typename E>
 void ObjectFile<E>::initialize_sections(Context<E> &ctx) {
   // Read sections
@@ -145,6 +201,17 @@ void ObjectFile<E>::initialize_sections(Context<E> &ctx) {
         shdr.sh_type != SHT_LLVM_ADDRSIG && !ctx.arg.relocatable)
       continue;
 
+    if constexpr (is_arm<E>)
+      if (shdr.sh_type == SHT_ARM_ATTRIBUTES)
+        continue;
+
+    if constexpr (is_riscv<E>) {
+      if (shdr.sh_type == SHT_RISCV_ATTRIBUTES) {
+        read_riscv_attributes(ctx, *this, this->get_string(ctx, shdr));
+        continue;
+      }
+    }
+
     switch (shdr.sh_type) {
     case SHT_GROUP: {
       // Get the signature of this section group.
@@ -155,7 +222,7 @@ void ObjectFile<E>::initialize_sections(Context<E> &ctx) {
       std::string_view signature;
       if (esym.st_type == STT_SECTION) {
         signature = this->shstrtab.data() +
-                    this->elf_sections[esym.st_shndx].sh_name;
+                    this->elf_sections[get_shndx(esym)].sh_name;
       } else {
         signature = this->symbol_strtab.data() + esym.st_name;
       }
@@ -181,15 +248,20 @@ void ObjectFile<E>::initialize_sections(Context<E> &ctx) {
       comdat_groups.push_back({group, (u32)i, entries.subspan(1)});
       break;
     }
-    case SHT_SYMTAB_SHNDX:
-      symtab_shndx_sec = this->template get_data<U32<E>>(ctx, shdr);
+    case SHT_REL:
+      if constexpr (E::is_rela)
+        Fatal(ctx) << *this << ": REL-type relocation table is not supported"
+                   << " for this target";
+      break;
+    case SHT_RELA:
+      if constexpr (!E::is_rela)
+        Fatal(ctx) << *this <<": RELA-type relocation table is not supported"
+                   << " for this target";
       break;
     case SHT_SYMTAB:
+    case SHT_SYMTAB_SHNDX:
     case SHT_STRTAB:
-    case SHT_REL:
-    case SHT_RELA:
     case SHT_NULL:
-    case SHT_ARM_ATTRIBUTES:
       break;
     default: {
       std::string_view name = this->shstrtab.data() + shdr.sh_name;
@@ -211,7 +283,7 @@ void ObjectFile<E>::initialize_sections(Context<E> &ctx) {
       }
 
       if (name == ".note.gnu.property") {
-        read_note_gnu_property(ctx, shdr);
+        parse_note_gnu_property(ctx, shdr);
         continue;
       }
 
@@ -235,12 +307,6 @@ void ObjectFile<E>::initialize_sections(Context<E> &ctx) {
           is_debug_section(shdr, name))
         continue;
 
-      // Save .llvm_addrsig for --icf=safe.
-      if (shdr.sh_type == SHT_LLVM_ADDRSIG && !ctx.arg.relocatable) {
-        llvm_addrsig = std::make_unique<InputSection<E>>(ctx, *this, name, i);
-        continue;
-      }
-
       // If an output file doesn't have a section header (i.e.
       // --oformat=binary is given), we discard all non-memory-allocated
       // sections. This is because without a section header, we can't find
@@ -250,9 +316,18 @@ void ObjectFile<E>::initialize_sections(Context<E> &ctx) {
 
       this->sections[i] = std::make_unique<InputSection<E>>(ctx, *this, name, i);
 
+      // Save .llvm_addrsig for --icf=safe.
+      if (shdr.sh_type == SHT_LLVM_ADDRSIG && !ctx.arg.relocatable) {
+        llvm_addrsig = std::move(this->sections[i]);
+        continue;
+      }
+
+      if (name == ".eh_frame")
+        eh_frame_section = this->sections[i].get();
+
       if constexpr (is_ppc32<E>)
         if (name == ".got2")
-          ppc32_got2 = this->sections[i].get();
+          extra.got2 = this->sections[i].get();
 
       // Save debug sections for --gdb-index.
       if (ctx.arg.gdb_index) {
@@ -313,17 +388,6 @@ void ObjectFile<E>::initialize_sections(Context<E> &ctx) {
   }
 }
 
-template <typename E>
-void ObjectFile<E>::initialize_ehframe_sections(Context<E> &ctx) {
-  for (i64 i = 0; i < sections.size(); i++) {
-    std::unique_ptr<InputSection<E>> &isec = sections[i];
-    if (isec && isec->is_alive && isec->name() == ".eh_frame") {
-      read_ehframe(ctx, *isec);
-      isec->is_alive = false;
-    }
-  }
-}
-
 // .eh_frame contains data records explaining how to handle exceptions.
 // When an exception is thrown, the runtime searches a record from
 // .eh_frame with the current program counter as a key. A record that
@@ -354,7 +418,11 @@ void ObjectFile<E>::initialize_ehframe_sections(Context<E> &ctx) {
 //
 // This function parses an input .eh_frame section.
 template <typename E>
-void ObjectFile<E>::read_ehframe(Context<E> &ctx, InputSection<E> &isec) {
+void ObjectFile<E>::parse_ehframe(Context<E> &ctx) {
+  if (!eh_frame_section)
+    return;
+
+  InputSection<E> &isec = *eh_frame_section;
   std::span<ElfRel<E>> rels = isec.get_rels(ctx);
   i64 cies_begin = cies.size();
   i64 fdes_begin = fdes.size();
@@ -490,7 +558,7 @@ void ObjectFile<E>::initialize_symbols(Context<E> &ctx) {
   this->symbols.resize(this->elf_syms.size());
 
   i64 num_globals = this->elf_syms.size() - this->first_global;
-  symvers.resize(num_globals);
+  has_symver.resize(num_globals);
 
   for (i64 i = 0; i < this->first_global; i++)
     this->symbols[i] = &this->local_syms[i];
@@ -505,14 +573,13 @@ void ObjectFile<E>::initialize_symbols(Context<E> &ctx) {
 
     // Parse symbol version after atsign
     if (i64 pos = name.find('@'); pos != name.npos) {
-      std::string_view ver = name.substr(pos + 1);
+      std::string_view ver = name.substr(pos);
       name = name.substr(0, pos);
 
-      if (!ver.empty() && ver != "@") {
-        if (ver.starts_with('@'))
+      if (ver != "@" && ver != "@@") {
+        if (ver.starts_with("@@"))
           key = name;
-        if (!esym.is_undef())
-          symvers[i - this->first_global] = ver.data();
+        has_symver.set(i - this->first_global);
       }
     }
 
@@ -527,7 +594,7 @@ void ObjectFile<E>::initialize_symbols(Context<E> &ctx) {
 // We expect them to be sorted, so sort them if necessary.
 template <typename E>
 void ObjectFile<E>::sort_relocations(Context<E> &ctx) {
-  if constexpr (is_riscv<E>) {
+  if constexpr (is_riscv<E> || is_loongarch<E>) {
     auto less = [&](const ElfRel<E> &a, const ElfRel<E> &b) {
       return a.r_offset < b.r_offset;
     };
@@ -575,7 +642,7 @@ static size_t find_null(std::string_view data, u64 entsize) {
 template <typename E>
 static std::unique_ptr<MergeableSection<E>>
 split_section(Context<E> &ctx, InputSection<E> &sec) {
-  if (!sec.is_alive || sec.sh_size == 0 || sec.relsec_idx != -1)
+  if (!sec.is_alive || sec.relsec_idx != -1)
     return nullptr;
 
   const ElfShdr<E> &shdr = sec.shdr();
@@ -587,6 +654,9 @@ split_section(Context<E> &ctx, InputSection<E> &sec) {
                                                shdr.sh_flags);
   rec->p2align = sec.p2align;
 
+  if (sec.sh_size == 0)
+    return rec;
+
   // If thes section contents are compressed, uncompress them.
   sec.uncompress(ctx);
 
@@ -656,13 +726,14 @@ split_section(Context<E> &ctx, InputSection<E> &sec) {
 // atomic unit of mergeable section "section pieces".
 //
 // This feature is typically used for string literals. String literals
-// are usually put into a mergeable section by a compiler. If the same
-// string literal happen to occur in two different translation units,
-// a linker merges them into a single instance of a string, so that
-// a linker's output doesn't contain duplicate string literals.
+// are usually put into a mergeable section by the compiler. If the same
+// string literal happens to occur in two different translation units,
+// the linker merges them into a single instance of a string, so that
+// the linker's output doesn't contain duplicate string literals.
 //
-// Handling symbols in mergeable sections is a bit tricky. Assume that
-// we have a mergeable section with the following contents and symbols:
+// Handling symbols in the mergeable sections is a bit tricky. Assume
+// that we have a mergeable section with the following contents and
+// symbols:
 //
 //   Hello world\0foo bar\0
 //   ^            ^
@@ -671,23 +742,23 @@ split_section(Context<E> &ctx, InputSection<E> &sec) {
 //
 // '\0' represents a NUL byte. This mergeable section contains two
 // section pieces, "Hello world" and "foo bar". The first string is
-// referred by two symbols, .rodata and .L.str0, and the second by
+// referred to by two symbols, .rodata and .L.str0, and the second by
 // .L.str1. .rodata is a section symbol and therefore a local symbol
-// and refers the begining of the section.
+// and refers to the beginning of the section.
 //
 // In this example, there are actually two different ways to point to
-// string "foo bar", because .rodata+12 and .L.str1+0 refer the same
+// string "foo bar", because .rodata+12 and .L.str1+0 refer to the same
 // place in the section. This kind of "out-of-bound" reference occurs
-// only when a symbol is a section symbol. In other words, compiler
-// may use an offset from the beginning of a section to refer any
+// only when a symbol is a section symbol. In other words, the compiler
+// may use an offset from the beginning of a section to refer to any
 // section piece in a section, but it doesn't do for any other types
 // of symbols.
 //
-// In mold, we attach section pieces symbols. If a relocation refers a
-// section symbol whose section is a mergeable section, we create a
-// new dummy symbol with a section piece and redirect the relocation
-// to the symbol. If a non-section symbol refers a section piece, the
-// section piece is attached to the symbol.
+// In mold, we attach symbols to section pieces. If a relocation refers
+// to a section symbol, and that symbol's section is a mergeable one,
+// we create a new dummy symbol for a section piece and redirect the
+// relocation to this new symbol. If a non-section symbol refers to a
+// section piece, the section piece is attached to the symbol.
 template <typename E>
 void ObjectFile<E>::initialize_mergeable_sections(Context<E> &ctx) {
   mergeable_sections.resize(sections.size());
@@ -708,7 +779,7 @@ void ObjectFile<E>::resolve_section_pieces(Context<E> &ctx) {
     if (m) {
       m->fragments.reserve(m->strings.size());
       for (i64 i = 0; i < m->strings.size(); i++)
-        m->fragments.push_back(m->parent->insert(m->strings[i], m->hashes[i],
+        m->fragments.push_back(m->parent->insert(ctx, m->strings[i], m->hashes[i],
                                                  m->p2align));
 
       // Shrink vectors that we will never use again to reclaim memory.
@@ -726,7 +797,7 @@ void ObjectFile<E>::resolve_section_pieces(Context<E> &ctx) {
       continue;
 
     std::unique_ptr<MergeableSection<E>> &m = mergeable_sections[get_shndx(esym)];
-    if (!m)
+    if (!m || m->fragments.empty())
       continue;
 
     SectionFragment<E> *frag;
@@ -803,7 +874,7 @@ void ObjectFile<E>::mark_addrsig(Context<E> &ctx) {
     u8 *end = cur + llvm_addrsig->contents.size();
 
     while (cur != end) {
-      Symbol<E> &sym = *this->symbols[read_uleb(cur)];
+      Symbol<E> &sym = *this->symbols[read_uleb(&cur)];
       if (sym.file == this)
         if (InputSection<E> *isec = sym.get_input_section())
           isec->address_significant = true;
@@ -833,12 +904,15 @@ void ObjectFile<E>::parse(Context<E> &ctx) {
     this->first_global = symtab_sec->sh_info;
     this->elf_syms = this->template get_data<ElfSym<E>>(ctx, *symtab_sec);
     this->symbol_strtab = this->get_string(ctx, symtab_sec->sh_link);
+
+    if (ElfShdr<E> *shdr = this->find_section(SHT_SYMTAB_SHNDX))
+      symtab_shndx_sec = this->template get_data<U32<E>>(ctx, *shdr);
   }
 
   initialize_sections(ctx);
   initialize_symbols(ctx);
   sort_relocations(ctx);
-  initialize_ehframe_sections(ctx);
+  parse_ehframe(ctx);
 }
 
 // Symbols with higher priorities overwrites symbols with lower priorities.
@@ -854,22 +928,22 @@ void ObjectFile<E>::parse(Context<E> &ctx) {
 //
 // Ties are broken by file priority.
 template <typename E>
-static u64 get_rank(InputFile<E> *file, const ElfSym<E> &esym, bool is_lazy) {
-  if (esym.is_common()) {
-    assert(!file->is_dso);
-    if (is_lazy)
-      return (6 << 24) + file->priority;
-    return (5 << 24) + file->priority;
-  }
+static u64 get_rank(InputFile<E> *file, const ElfSym<E> &esym, bool is_in_archive) {
+  auto get_sym_rank = [&] {
+    if (esym.is_common()) {
+      assert(!file->is_dso);
+      return is_in_archive ? 6 : 5;
+    }
+
+    if (file->is_dso || is_in_archive)
+      return (esym.st_bind == STB_WEAK) ? 4 : 3;
 
-  if (file->is_dso || is_lazy) {
     if (esym.st_bind == STB_WEAK)
-      return (4 << 24) + file->priority;
-    return (3 << 24) + file->priority;
-  }
-  if (esym.st_bind == STB_WEAK)
-    return (2 << 24) + file->priority;
-  return (1 << 24) + file->priority;
+      return 2;
+    return 1;
+  };
+
+  return (get_sym_rank() << 24) + file->priority;
 }
 
 template <typename E>
@@ -943,8 +1017,6 @@ void ObjectFile<E>::resolve_symbols(Context<E> &ctx) {
       sym.sym_idx = i;
       sym.ver_idx = ctx.default_version;
       sym.is_weak = esym.is_weak();
-      sym.is_imported = false;
-      sym.is_exported = false;
     }
   }
 }
@@ -974,7 +1046,7 @@ ObjectFile<E>::mark_live_objects(Context<E> &ctx,
       continue;
 
     bool keep = esym.is_undef() || (esym.is_common() && !sym.esym().is_common());
-    if (keep && fast_mark(sym.file->is_alive)) {
+    if (keep && !sym.file->is_alive.test_and_set()) {
       feeder(sym.file);
 
       if (sym.is_traced)
@@ -984,135 +1056,6 @@ ObjectFile<E>::mark_live_objects(Context<E> &ctx,
   }
 }
 
-// Comdat groups are used to de-duplicate functions and data that may
-// be included into multiple object files. C++ compiler uses comdat
-// groups to de-duplicate instantiated templates.
-//
-// For example, if a compiler decides to instantiate `std::vector<int>`,
-// it generates code and data for `std::vector<int>` and put them into a
-// comdat group whose name is the mangled name of `std::vector<int>`.
-// The instantiation may happen multiple times for different translation
-// units. Then linker de-duplicates them so that the resulting executable
-// contains only a single copy of `std::vector<int>`.
-template <typename E>
-void ObjectFile<E>::resolve_comdat_groups() {
-  for (ComdatGroupRef<E> &ref : comdat_groups)
-    update_minimum(ref.group->owner, this->priority);
-}
-
-template <typename E>
-void ObjectFile<E>::eliminate_duplicate_comdat_groups() {
-  for (ComdatGroupRef<E> &ref : comdat_groups)
-    if (ref.group->owner != this->priority)
-      for (u32 i : ref.members)
-        if (sections[i])
-          sections[i]->kill();
-}
-
-template <typename E>
-void ObjectFile<E>::claim_unresolved_symbols(Context<E> &ctx) {
-  if (!this->is_alive)
-    return;
-
-  auto report_undef = [&](Symbol<E> &sym) {
-    std::stringstream ss;
-    if (std::string_view source = this->get_source_name(); !source.empty())
-      ss << ">>> referenced by " << source << "\n";
-    else
-      ss << ">>> referenced by " << *this << "\n";
-
-    typename decltype(ctx.undef_errors)::accessor acc;
-    ctx.undef_errors.insert(acc, {sym.name(), {}});
-    acc->second.push_back(ss.str());
-  };
-
-  for (i64 i = this->first_global; i < this->elf_syms.size(); i++) {
-    const ElfSym<E> &esym = this->elf_syms[i];
-    Symbol<E> &sym = *this->symbols[i];
-    if (!esym.is_undef())
-      continue;
-
-    std::scoped_lock lock(sym.mu);
-
-    // If a protected/hidden undefined symbol is resolved to an
-    // imported symbol, it's handled as if no symbols were found.
-    if (sym.file && sym.file->is_dso &&
-        (sym.visibility == STV_PROTECTED || sym.visibility == STV_HIDDEN)) {
-      report_undef(sym);
-      continue;
-    }
-
-    if (sym.file &&
-        (!sym.esym().is_undef() || sym.file->priority <= this->priority))
-      continue;
-
-    // If a symbol name is in the form of "foo@version", search for
-    // symbol "foo" and check if the symbol has version "version".
-    std::string_view key = this->symbol_strtab.data() + esym.st_name;
-    if (i64 pos = key.find('@'); pos != key.npos) {
-      Symbol<E> *sym2 = get_symbol(ctx, key.substr(0, pos));
-      if (sym2->file && sym2->file->is_dso &&
-          sym2->get_version() == key.substr(pos + 1)) {
-        this->symbols[i] = sym2;
-        continue;
-      }
-    }
-
-    auto claim = [&](bool is_imported) {
-      if (sym.is_traced)
-        SyncOut(ctx) << "trace-symbol: " << *this << ": unresolved"
-                     << (esym.is_weak() ? " weak" : "")
-                     << " symbol " << sym;
-
-      sym.file = this;
-      sym.origin = 0;
-      sym.value = 0;
-      sym.sym_idx = i;
-      sym.is_weak = false;
-      sym.is_imported = is_imported;
-      sym.is_exported = false;
-      sym.ver_idx = is_imported ? 0 : ctx.default_version;
-    };
-
-    if (esym.is_undef_weak()) {
-      if (ctx.arg.shared && sym.visibility != STV_HIDDEN &&
-          ctx.arg.z_dynamic_undefined_weak) {
-        // Global weak undefined symbols are promoted to dynamic symbols
-        // when when linking a DSO, unless `-z nodynamic_undefined_weak`
-        // was given.
-        claim(true);
-      } else {
-        // Otherwise, weak undefs are converted to absolute symbols with value 0.
-        claim(false);
-      }
-      continue;
-    }
-
-    if (ctx.arg.unresolved_symbols == UNRESOLVED_WARN)
-      report_undef(sym);
-
-    // Traditionally, remaining undefined symbols cause a link failure
-    // only when we are creating an executable. Undefined symbols in
-    // shared objects are promoted to dynamic symbols, so that they'll
-    // get another chance to be resolved at run-time. You can change the
-    // behavior by passing `-z defs` to the linker.
-    //
-    // Even if `-z defs` is given, weak undefined symbols are still
-    // promoted to dynamic symbols for compatibility with other linkers.
-    // Some major programs, notably Firefox, depend on the behavior
-    // (they use this loophole to export symbols from libxul.so).
-    if (ctx.arg.shared && sym.visibility != STV_HIDDEN &&
-        (!ctx.arg.z_defs || ctx.arg.unresolved_symbols != UNRESOLVED_ERROR)) {
-      claim(true);
-      continue;
-    }
-
-    // Convert remaining undefined symbols to absolute symbols with value 0.
-    if (ctx.arg.unresolved_symbols != UNRESOLVED_ERROR || ctx.arg.noinhibit_exec)
-      claim(false);
-  }
-}
-
 template <typename E>
 void ObjectFile<E>::scan_relocations(Context<E> &ctx) {
   // Scan relocations against seciton contents
@@ -1125,6 +1068,11 @@ void ObjectFile<E>::scan_relocations(Context<E> &ctx) {
     for (ElfRel<E> &rel : cie.get_rels()) {
       Symbol<E> &sym = *this->symbols[rel.r_sym];
 
+      if (ctx.arg.pic && rel.r_type == E::R_ABS)
+        Error(ctx) << *this << ": relocation " << rel << " in .eh_frame can"
+                   << " not be used when making a position-independent output;"
+                   << " recompile with -fPIE or -fPIC";
+
       if (sym.is_imported) {
         if (sym.get_type() != STT_FUNC)
           Fatal(ctx) << *this << ": " << sym
@@ -1200,8 +1148,6 @@ void ObjectFile<E>::convert_common_symbols(Context<E> &ctx) {
     sym.sym_idx = i;
     sym.ver_idx = ctx.default_version;
     sym.is_weak = false;
-    sym.is_imported = false;
-    sym.is_exported = false;
 
     sections.push_back(std::move(isec));
   }
@@ -1238,9 +1184,6 @@ void ObjectFile<E>::compute_symtab_size(Context<E> &ctx) {
   this->output_sym_indices.resize(this->elf_syms.size(), -1);
 
   auto is_alive = [&](Symbol<E> &sym) -> bool {
-    if (!ctx.arg.gc_sections)
-      return true;
-
     if (SectionFragment<E> *frag = sym.get_frag())
       return frag->is_alive;
     if (InputSection<E> *isec = sym.get_input_section())
@@ -1286,31 +1229,29 @@ void ObjectFile<E>::populate_symtab(Context<E> &ctx) {
   u8 *strtab_base = ctx.buf + ctx.strtab->shdr.sh_offset;
   i64 strtab_off = this->strtab_offset;
 
-  auto write_sym = [&](Symbol<E> &sym, i64 &symtab_idx) {
+  auto write_sym = [&](Symbol<E> &sym, i64 idx) {
     U32<E> *xindex = nullptr;
     if (ctx.symtab_shndx)
-      xindex = (U32<E> *)(ctx.buf + ctx.symtab_shndx->shdr.sh_offset +
-                          symtab_idx * 4);
+      xindex = (U32<E> *)(ctx.buf + ctx.symtab_shndx->shdr.sh_offset) + idx;
 
-    symtab_base[symtab_idx++] = to_output_esym(ctx, sym, strtab_off, xindex);
+    symtab_base[idx] = to_output_esym(ctx, sym, strtab_off, xindex);
     strtab_off += write_string(strtab_base + strtab_off, sym.name());
   };
 
   i64 local_symtab_idx = this->local_symtab_idx;
   i64 global_symtab_idx = this->global_symtab_idx;
-  for (i64 i = 1; i < this->first_global; i++) {
-    Symbol<E> &sym = *this->symbols[i];
-    if (sym.write_to_symtab)
-      write_sym(sym, local_symtab_idx);
-  }
+
+  for (i64 i = 1; i < this->first_global; i++)
+    if (Symbol<E> &sym = *this->symbols[i]; sym.write_to_symtab)
+      write_sym(sym, local_symtab_idx++);
 
   for (i64 i = this->first_global; i < this->elf_syms.size(); i++) {
     Symbol<E> &sym = *this->symbols[i];
     if (sym.file == this && sym.write_to_symtab) {
       if (sym.is_local(ctx))
-        write_sym(sym, local_symtab_idx);
+        write_sym(sym, local_symtab_idx++);
       else
-        write_sym(sym, global_symtab_idx);
+        write_sym(sym, global_symtab_idx++);
     }
   }
 }
@@ -1341,7 +1282,6 @@ SharedFile<E>::create(Context<E> &ctx, MappedFile<Context<E>> *mf) {
 template <typename E>
 SharedFile<E>::SharedFile(Context<E> &ctx, MappedFile<Context<E>> *mf)
   : InputFile<E>(ctx, mf) {
-  this->is_needed = ctx.as_needed;
   this->is_alive = !ctx.as_needed;
 }
 
@@ -1481,8 +1421,6 @@ void SharedFile<E>::resolve_symbols(Context<E> &ctx) {
       sym.sym_idx = i;
       sym.ver_idx = versyms[i];
       sym.is_weak = false;
-      sym.is_imported = false;
-      sym.is_exported = false;
     }
   }
 }
@@ -1498,8 +1436,8 @@ SharedFile<E>::mark_live_objects(Context<E> &ctx,
     if (sym.is_traced)
       print_trace_symbol(ctx, *this, esym, sym);
 
-    if (esym.is_undef() && sym.file && !sym.file->is_dso &&
-        fast_mark(sym.file->is_alive)) {
+    if (esym.is_undef() && !esym.is_weak() && sym.file && !sym.file->is_dso &&
+        !sym.file->is_alive.test_and_set()) {
       feeder(sym.file);
 
       if (sym.is_traced)
@@ -1510,7 +1448,7 @@ SharedFile<E>::mark_live_objects(Context<E> &ctx,
 }
 
 template <typename E>
-std::vector<Symbol<E> *> SharedFile<E>::find_aliases(Symbol<E> *sym) {
+std::span<Symbol<E> *> SharedFile<E>::find_aliases(Symbol<E> *sym) {
   assert(sym->file == this);
 
   std::call_once(init_aliases, [&] {
@@ -1526,16 +1464,12 @@ std::vector<Symbol<E> *> SharedFile<E>::find_aliases(Symbol<E> *sym) {
     });
   });
 
-  struct Cmp {
-    bool operator()(Symbol<E> *sym, u64 val) { return sym->esym().st_value < val; }
-    bool operator()(u64 val, Symbol<E> *sym) { return val < sym->esym().st_value; }
-  };
+  auto [begin, end] = std::equal_range(aliases.begin(), aliases.end(), sym,
+                                       [&](Symbol<E> *x, Symbol<E> *y) {
+    return x->esym().st_value < y->esym().st_value;
+  });
 
-  auto [begin, end] =
-    std::equal_range(aliases.begin(), aliases.end(), sym->esym().st_value, Cmp{});
-  std::vector<Symbol<E> *> vec{begin, end};
-  std::erase(vec, sym);
-  return vec;
+  return {&*begin, &*end};
 }
 
 // Infer an alignment of a DSO symbol. An alignment of a symbol in other
@@ -1555,13 +1489,13 @@ i64 SharedFile<E>::get_alignment(Symbol<E> *sym) {
 }
 
 template <typename E>
-bool SharedFile<E>::is_readonly(Context<E> &ctx, Symbol<E> *sym) {
-  ElfPhdr<E> *phdr = this->get_phdr();
+bool SharedFile<E>::is_readonly(Symbol<E> *sym) {
   u64 val = sym->esym().st_value;
 
-  for (i64 i = 0; i < this->get_ehdr().e_phnum; i++)
-    if (phdr[i].p_type == PT_LOAD && !(phdr[i].p_flags & PF_W) &&
-        phdr[i].p_vaddr <= val && val < phdr[i].p_vaddr + phdr[i].p_memsz)
+  for (ElfPhdr<E> &phdr : this->get_phdrs())
+    if ((phdr.p_type == PT_LOAD || phdr.p_type == PT_GNU_RELRO) &&
+        !(phdr.p_flags & PF_W) &&
+        phdr.p_vaddr <= val && val < phdr.p_vaddr + phdr.p_memsz)
       return true;
   return false;
 }
@@ -1601,8 +1535,8 @@ void SharedFile<E>::populate_symtab(Context<E> &ctx) {
 
     U32<E> *xindex = nullptr;
     if (ctx.symtab_shndx)
-      xindex = (U32<E> *)(ctx.buf + ctx.symtab_shndx->shdr.sh_offset +
-                          (this->global_symtab_idx + i) * 4);
+      xindex = (U32<E> *)(ctx.buf + ctx.symtab_shndx->shdr.sh_offset) +
+               this->global_symtab_idx + i;
 
     *symtab++ = to_output_esym(ctx, sym, strtab_off, xindex);
     strtab_off += write_string(strtab + strtab_off, sym.name());

elf/input-sections.cc

@@ -57,7 +57,10 @@ InputSection<E>::InputSection(Context<E> &ctx, ObjectFile<E> &file,
   // early for REL-type ELF types to read relocation addends from
   // section contents. For RELA-type, we don't need to do this because
   // addends are in relocations.
-  if constexpr (!E::is_rela)
+  //
+  // SH-4 stores addends to sections despite being RELA, which is a
+  // special (and buggy) case.
+  if constexpr (!E::is_rela || is_sh4<E>)
     uncompress(ctx);
 }
 
@@ -104,30 +107,6 @@ void InputSection<E>::uncompress_to(Context<E> &ctx, u8 *buf) {
   }
 }
 
-template <typename E>
-static Action get_rel_action(Context<E> &ctx, Symbol<E> &sym,
-                             const Action table[3][4]) {
-  auto get_output_type = [&] {
-    if (ctx.arg.shared)
-      return 0;
-    if (ctx.arg.pie)
-      return 1;
-    return 2;
-  };
-
-  auto get_sym_type = [&] {
-    if (sym.is_absolute())
-      return 0;
-    if (!sym.is_imported)
-      return 1;
-    if (sym.get_type() != STT_FUNC)
-      return 2;
-    return 3;
-  };
-
-  return table[get_output_type()][get_sym_type()];
-}
-
 template <typename E>
 static void scan_rel(Context<E> &ctx, InputSection<E> &isec, Symbol<E> &sym,
                      const ElfRel<E> &rel, Action action) {
@@ -171,78 +150,125 @@ static void scan_rel(Context<E> &ctx, InputSection<E> &isec, Symbol<E> &sym,
   case NONE:
     break;
   case ERROR:
+    // Print out the "recompile with -fPIC" error message.
     error();
     break;
   case COPYREL:
+    // Create a copy relocation.
     if (!ctx.arg.z_copyreloc)
       error();
     copyrel();
     break;
   case DYN_COPYREL:
+    // Same as COPYREL but try to avoid creating a copy relocation by
+    // creating a dynamic relocation instead if the relocation is in
+    // a writable section.
+    //
+    // GHC (Glasgow Haskell Compiler) places a small amount of data in
+    // .text before each function and access that data with a fixed
+    // offset. The function breaks if we copy-relocate the data. For such
+    // programs, we should avoid copy relocations if possible.
+    //
+    // Besides GHC, copy relocation is a hacky solution, so if we can
+    // represent a relocation either with copyrel or dynrel, we prefer
+    // dynamic relocation.
     if (writable || !ctx.arg.z_copyreloc)
       dynrel();
     else
       copyrel();
     break;
   case PLT:
+    // Create a PLT entry.
     sym.flags |= NEEDS_PLT;
     break;
   case CPLT:
+    // Create a canonical PLT entry.
     sym.flags |= NEEDS_CPLT;
     break;
   case DYN_CPLT:
+    // Same as CPLT but try to avoid creating a canonical PLT creating by
+    // creating a dynamic relocation instead if the relocation is in a
+    // writable section. The motivation behind it is hte same as DYN_COPYREL.
     if (writable)
       dynrel();
     else
       sym.flags |= NEEDS_CPLT;
     break;
   case DYNREL:
+    // Create a dynamic relocation.
     dynrel();
     break;
   case BASEREL:
+    // Create a base relocation.
     check_textrel();
     if (!isec.is_relr_reloc(ctx, rel))
       isec.file.num_dynrel++;
     break;
   case IFUNC:
+    // Create an IRELATIVE relocation for a GNU ifunc symbol.
+    //
+    // We usually create an IRELATIVE relocation in .got for each ifunc.
+    // However, if a statically-initialized pointer is initialized to an
+    // ifunc's address, we have no choice other than emitting an IRELATIVE
+    // relocation for each such pointer.
     dynrel();
-    ctx.num_ifunc_dynrels.fetch_add(1, std::memory_order_relaxed);
+    ctx.num_ifunc_dynrels++;
     break;
   default:
     unreachable();
   }
 }
 
+template <typename E>
+static inline i64 get_output_type(Context<E> &ctx) {
+  if (ctx.arg.shared)
+    return 0;
+  if (ctx.arg.pie)
+    return 1;
+  return 2;
+}
+
+template <typename E>
+static inline i64 get_sym_type(Symbol<E> &sym) {
+  if (sym.is_absolute())
+    return 0;
+  if (!sym.is_imported)
+    return 1;
+  if (sym.get_type() != STT_FUNC)
+    return 2;
+  return 3;
+}
+
 template <typename E>
 static Action get_pcrel_action(Context<E> &ctx, Symbol<E> &sym) {
   // This is for PC-relative relocations (e.g. R_X86_64_PC32).
   // We cannot promote them to dynamic relocations because the dynamic
   // linker generally does not support PC-relative relocations.
-  constexpr static Action table[3][4] = {
+  static Action table[3][4] = {
     // Absolute  Local    Imported data  Imported code
     {  ERROR,    NONE,    ERROR,         PLT    },  // Shared object
     {  ERROR,    NONE,    COPYREL,       PLT    },  // Position-independent exec
     {  NONE,     NONE,    COPYREL,       CPLT   },  // Position-dependent exec
   };
 
-  return get_rel_action(ctx, sym, table);
+  return table[get_output_type(ctx)][get_sym_type(sym)];
 }
 
 template <typename E>
 static Action get_absrel_action(Context<E> &ctx, Symbol<E> &sym) {
   // This is a decision table for absolute relocations that is smaller
-  // than the word size (e.g. R_X86_64_32). Since the dynamic linker
+  // than the pointer size (e.g. R_X86_64_32). Since the dynamic linker
   // generally does not support dynamic relocations smaller than the
-  // word size, we need to report an error if a relocation cannot be
+  // pointer size, we need to report an error if a relocation cannot be
   // resolved at link-time.
-  constexpr static Action table[3][4] = {
+  static Action table[3][4] = {
     // Absolute  Local    Imported data  Imported code
     {  NONE,     ERROR,   ERROR,         ERROR },  // Shared object
     {  NONE,     ERROR,   ERROR,         ERROR },  // Position-independent exec
     {  NONE,     NONE,    COPYREL,       CPLT  },  // Position-dependent exec
   };
 
-  return get_rel_action(ctx, sym, table);
+  return table[get_output_type(ctx)][get_sym_type(sym)];
 }
 
 template <typename E>
@@ -250,17 +276,17 @@ static Action get_dyn_absrel_action(Context<E> &ctx, Symbol<E> &sym) {
   if (sym.is_ifunc())
     return IFUNC;
 
-  // This is a decision table for absolute relocations for the word
+  // This is a decision table for absolute relocations for the pointer
   // size data (e.g. R_X86_64_64). Unlike the absrel_table, we can emit
   // a dynamic relocation if we cannot resolve an address at link-time.
-  constexpr static Action table[3][4] = {
+  static Action table[3][4] = {
     // Absolute  Local    Imported data  Imported code
     {  NONE,     BASEREL, DYNREL,        DYNREL   },  // Shared object
     {  NONE,     BASEREL, DYNREL,        DYNREL   },  // Position-independent exec
     {  NONE,     NONE,    DYN_COPYREL,   DYN_CPLT },  // Position-dependent exec
   };
 
-  return get_rel_action(ctx, sym, table);
+  return table[get_output_type(ctx)][get_sym_type(sym)];
 }
 
 template <typename E>
@@ -272,14 +298,14 @@ static Action get_ppc64_toc_action(Context<E> &ctx, Symbol<E> &sym) {
   // PLTs for .toc sections. PPC64's .toc is a compiler-generated
   // GOT-like section, and no user-generated code directly uses values
   // in it.
-  constexpr static Action table[3][4] = {
+  static Action table[3][4] = {
     // Absolute  Local    Imported data  Imported code
     {  NONE,     BASEREL, DYNREL,        DYNREL },  // Shared object
     {  NONE,     BASEREL, DYNREL,        DYNREL },  // Position-independent exec
     {  NONE,     NONE,    DYNREL,        DYNREL },  // Position-dependent exec
   };
 
-  return get_rel_action(ctx, sym, table);
+  return table[get_output_type(ctx)][get_sym_type(sym)];
 }
 
 template <typename E>
@@ -306,6 +332,38 @@ void InputSection<E>::scan_toc_rel(Context<E> &ctx, Symbol<E> &sym,
   scan_rel(ctx, *this, sym, rel, get_ppc64_toc_action(ctx, sym));
 }
 
+template <typename E>
+void InputSection<E>::scan_tlsdesc(Context<E> &ctx, Symbol<E> &sym) {
+  if (ctx.arg.is_static ||
+      (ctx.arg.relax && sym.is_tprel_linktime_const(ctx))) {
+    // Relax TLSDESC to Local Exec. In this case, we directly materialize
+    // a TP-relative offset, so no dynamic relocation is needed.
+    //
+    // TLSDESC relocs must always be relaxed for statically-linked
+    // executables even if -no-relax is given. It is because a
+    // statically-linked executable doesn't contain a trampoline
+    // function needed for TLSDESC.
+  } else if (ctx.arg.relax && sym.is_tprel_runtime_const(ctx)) {
+    // In this condition, TP-relative offset of a thread-local variable
+    // is known at process startup time, so we can relax TLSDESC to the
+    // code that reads the TP-relative offset from GOT and add TP to it.
+    sym.flags |= NEEDS_GOTTP;
+  } else {
+    // If no relaxation is doable, we simply create a TLSDESC dynamic
+    // relocation.
+    sym.flags |= NEEDS_TLSDESC;
+  }
+}
+
+template <typename E>
+void InputSection<E>::check_tlsle(Context<E> &ctx, Symbol<E> &sym,
+                                  const ElfRel<E> &rel) {
+  if (ctx.arg.shared)
+    Error(ctx) << *this << ": relocation " << rel << " against `" << sym
+               << "` can not be used when making a shared object;"
+               << " recompile with -fPIC";
+}
+
 template <typename E>
 static void apply_absrel(Context<E> &ctx, InputSection<E> &isec,
                          Symbol<E> &sym, const ElfRel<E> &rel, u8 *loc,
@@ -313,8 +371,8 @@ static void apply_absrel(Context<E> &ctx, InputSection<E> &isec,
                          Action action) {
   bool writable = (isec.shdr().sh_flags & SHF_WRITE);
 
-  auto apply_dynrel = [&] {
-    *dynrel++ = ElfRel<E>(P, E::R_ABS, sym.get_dynsym_idx(ctx), A);
+  auto emit_abs_dynrel = [&] {
+    *dynrel++ = ElfRel<E>(P, E::R_DYNAMIC, sym.get_dynsym_idx(ctx), A);
     if (ctx.arg.apply_dynamic_relocs)
       *(Word<E> *)loc = A;
   };
@@ -336,26 +394,29 @@ static void apply_absrel(Context<E> &ctx, InputSection<E> &isec,
     break;
   case DYN_COPYREL:
     if (writable || !ctx.arg.z_copyreloc)
-      apply_dynrel();
+      emit_abs_dynrel();
     else
       *(Word<E> *)loc = S + A;
     break;
   case DYN_CPLT:
     if (writable)
-      apply_dynrel();
+      emit_abs_dynrel();
     else
       *(Word<E> *)loc = S + A;
     break;
   case DYNREL:
-    apply_dynrel();
+    emit_abs_dynrel();
     break;
-  case IFUNC: {
-    u64 addr = sym.get_addr(ctx, NO_PLT) + A;
-    *dynrel++ = ElfRel<E>(P, E::R_IRELATIVE, 0, addr);
-    if (ctx.arg.apply_dynamic_relocs)
-      *(Word<E> *)loc = addr;
+  case IFUNC:
+    if constexpr (supports_ifunc<E>) {
+      u64 addr = sym.get_addr(ctx, NO_PLT) + A;
+      *dynrel++ = ElfRel<E>(P, E::R_IRELATIVE, 0, addr);
+      if (ctx.arg.apply_dynamic_relocs)
+        *(Word<E> *)loc = addr;
+    } else {
+      unreachable();
+    }
     break;
-  }
   default:
     unreachable();
   }
@@ -365,8 +426,8 @@ template <typename E>
 void InputSection<E>::apply_dyn_absrel(Context<E> &ctx, Symbol<E> &sym,
                                        const ElfRel<E> &rel, u8 *loc,
                                        u64 S, i64 A, u64 P,
-                                       ElfRel<E> *&dynrel) {
-  apply_absrel(ctx, *this, sym, rel, loc, S, A, P, dynrel,
+                                       ElfRel<E> **dynrel) {
+  apply_absrel(ctx, *this, sym, rel, loc, S, A, P, *dynrel,
                get_dyn_absrel_action(ctx, sym));
 }
 
@@ -374,8 +435,8 @@ template <typename E>
 void InputSection<E>::apply_toc_rel(Context<E> &ctx, Symbol<E> &sym,
                                     const ElfRel<E> &rel, u8 *loc,
                                     u64 S, i64 A, u64 P,
-                                    ElfRel<E> *&dynrel) {
-  apply_absrel(ctx, *this, sym, rel, loc, S, A, P, dynrel,
+                                    ElfRel<E> **dynrel) {
+  apply_absrel(ctx, *this, sym, rel, loc, S, A, P, *dynrel,
                get_ppc64_toc_action(ctx, sym));
 }
 
@@ -385,11 +446,10 @@ void InputSection<E>::write_to(Context<E> &ctx, u8 *buf) {
     return;
 
   // Copy data
-  if constexpr (is_riscv<E>) {
+  if constexpr (is_riscv<E>)
     copy_contents_riscv(ctx, buf);
-  } else {
+  else
     uncompress_to(ctx, buf);
-  }
 
   // Apply relocations
   if (!ctx.arg.relocatable) {
@@ -402,7 +462,7 @@ void InputSection<E>::write_to(Context<E> &ctx, u8 *buf) {
 
 // Get the name of a function containin a given offset.
 template <typename E>
-std::string_view InputSection<E>::get_func_name(Context<E> &ctx, i64 offset) {
+std::string_view InputSection<E>::get_func_name(Context<E> &ctx, i64 offset) const {
   for (const ElfSym<E> &esym : file.elf_syms) {
     if (esym.st_shndx == shndx && esym.st_type == STT_FUNC &&
         esym.st_value <= offset && offset < esym.st_value + esym.st_size) {
@@ -415,61 +475,69 @@ std::string_view InputSection<E>::get_func_name(Context<E> &ctx, i64 offset) {
   return "";
 }
 
-// Record an undefined symbol error which will be displayed all at
-// once by report_undef_errors().
+// Test if the symbol a given relocation refers to has already been resolved.
+// If not, record that error and returns true.
 template <typename E>
-void InputSection<E>::record_undef_error(Context<E> &ctx, const ElfRel<E> &rel) {
-  std::stringstream ss;
-  if (std::string_view source = file.get_source_name(); !source.empty())
-    ss << ">>> referenced by " << source << "\n";
-  else
-    ss << ">>> referenced by " << *this << "\n";
-
-  ss << ">>>               " << file;
-  if (std::string_view func = get_func_name(ctx, rel.r_offset); !func.empty())
-    ss << ":(" << func << ")";
+bool InputSection<E>::record_undef_error(Context<E> &ctx, const ElfRel<E> &rel) {
+  // If a relocation refers to a linker-synthesized symbol for a
+  // section fragment, it's always been resolved.
+  if (file.elf_syms.size() <= rel.r_sym)
+    return false;
 
   Symbol<E> &sym = *file.symbols[rel.r_sym];
+  const ElfSym<E> &esym = file.elf_syms[rel.r_sym];
 
-  typename decltype(ctx.undef_errors)::accessor acc;
-  ctx.undef_errors.insert(acc, {sym.name(), {}});
-  acc->second.push_back(ss.str());
-}
-
-// Report all undefined symbols, grouped by symbol.
-template <typename E>
-void report_undef_errors(Context<E> &ctx) {
-  constexpr i64 max_errors = 3;
-
-  for (auto &pair : ctx.undef_errors) {
-    std::string_view sym_name = pair.first;
-    std::span<std::string> errors = pair.second;
-
-    if (ctx.arg.demangle)
-      sym_name = demangle(sym_name);
-
-    std::stringstream ss;
-    ss << "undefined symbol: " << sym_name << "\n";
-
-    for (i64 i = 0; i < errors.size() && i < max_errors; i++)
-      ss << errors[i];
-
-    if (errors.size() > max_errors)
-      ss << ">>> referenced " << (errors.size() - max_errors) << " more times\n";
-
-    if (ctx.arg.unresolved_symbols == UNRESOLVED_ERROR)
-      Error(ctx) << ss.str();
-    else if (ctx.arg.unresolved_symbols == UNRESOLVED_WARN)
-      Warn(ctx) << ss.str();
+  // If a symbol is defined in a comdat group, and the comdat group is
+  // discarded, the symbol may not have an owner. It is technically an
+  // violation of the One Definition Rule, so it is a programmer's fault.
+  if (!sym.file) {
+    Error(ctx) << *this << ": " << sym << " refers to a discarded COMDAT section"
+               << " probably due to an ODR violation";
+    return true;
   }
 
-  ctx.checkpoint();
+  auto record = [&] {
+    std::stringstream ss;
+    if (std::string_view source = file.get_source_name(); !source.empty())
+      ss << ">>> referenced by " << source << "\n";
+    else
+      ss << ">>> referenced by " << *this << "\n";
+
+    ss << ">>>               " << file;
+    if (std::string_view func = get_func_name(ctx, rel.r_offset); !func.empty())
+      ss << ":(" << func << ")";
+
+    typename decltype(ctx.undef_errors)::accessor acc;
+    ctx.undef_errors.insert(acc, {sym.name(), {}});
+    acc->second.push_back(ss.str());
+  };
+
+  // A non-weak undefined symbol must be promoted to an imported
+  // symbol or resolved to an defined symbol. Otherwise, it's an
+  // undefined symbol error.
+  //
+  // Every ELF file has an absolute local symbol as its first symbol.
+  // Referring to that symbol is always valid.
+  bool is_undef = esym.is_undef() && !esym.is_weak() && sym.sym_idx;
+  if (!sym.is_imported && is_undef && sym.esym().is_undef()) {
+    record();
+    return true;
+  }
+
+  // If a protected/hidden undefined symbol is resolved to other .so,
+  // it's handled as if no symbols were found.
+  if (sym.file->is_dso &&
+      (sym.visibility == STV_PROTECTED || sym.visibility == STV_HIDDEN)) {
+    record();
+    return true;
+  }
+
+  return false;
 }
 
 using E = MOLD_TARGET;
 
 template struct CieRecord<E>;
 template class InputSection<E>;
-template void report_undef_errors(Context<E> &);
 
 } // namespace mold::elf

elf/jobs.cc

@@ -0,0 +1,49 @@
+#include "mold.h"
+
+#ifndef _WIN32
+# include <fcntl.h>
+# include <pwd.h>
+# include <sys/stat.h>
+# include <sys/types.h>
+# include <unistd.h>
+#endif
+
+namespace mold::elf {
+
+template <typename E>
+void acquire_global_lock(Context<E> &ctx) {
+#ifndef _WIN32
+  char *jobs = getenv("MOLD_JOBS");
+  if (!jobs || std::string(jobs) != "1")
+    return;
+
+  char *home = getenv("HOME");
+  if (!home)
+    home = getpwuid(getuid())->pw_dir;
+
+  std::string path = std::string(home) + "/.mold-lock";
+  int fd = open(path.c_str(), O_WRONLY | O_CREAT | O_CLOEXEC, 0600);
+  if (fd == -1)
+    return;
+
+  if (lockf(fd, F_LOCK, 0) == -1)
+    return;
+
+  ctx.global_lock_fd = fd;
+#endif
+}
+
+template <typename E>
+void release_global_lock(Context<E> &ctx) {
+#ifndef _WIN32
+  if (ctx.global_lock_fd)
+    close(*ctx.global_lock_fd);
+#endif
+}
+
+using E = MOLD_TARGET;
+
+template void acquire_global_lock(Context<E> &);
+template void release_global_lock(Context<E> &);
+
+} // namespace mold::elf

elf/linker-script.cc

@@ -279,9 +279,8 @@ static bool read_label(std::span<std::string_view> &tok,
 template <typename E>
 static void
 read_version_script_commands(Context<E> &ctx, std::span<std::string_view> &tok,
-                             std::string_view ver_str, u16 ver_idx, bool is_cpp) {
-  bool is_global = true;
-
+                             std::string_view ver_str, u16 ver_idx,
+                             bool is_global, bool is_cpp) {
   while (!tok.empty() && tok[0] != "}") {
     if (read_label(tok, "global")) {
       is_global = true;
@@ -299,11 +298,11 @@ read_version_script_commands(Context<E> &ctx, std::span<std::string_view> &tok,
       if (!tok.empty() && tok[0] == "\"C\"") {
         tok = tok.subspan(1);
         tok = skip(ctx, tok, "{");
-        read_version_script_commands( ctx, tok, ver_str, ver_idx, false);
+        read_version_script_commands( ctx, tok, ver_str, ver_idx, is_global, false);
       } else {
         tok = skip(ctx, tok, "\"C++\"");
         tok = skip(ctx, tok, "{");
-        read_version_script_commands(ctx, tok, ver_str, ver_idx, true);
+        read_version_script_commands(ctx, tok, ver_str, ver_idx, is_global, true);
       }
 
       tok = skip(ctx, tok, "}");
@@ -349,7 +348,7 @@ void read_version_script(Context<E> &ctx, std::span<std::string_view> &tok) {
     }
 
     tok = skip(ctx, tok, "{");
-    read_version_script_commands(ctx, tok, ver_str, ver_idx, false);
+    read_version_script_commands(ctx, tok, ver_str, ver_idx, true, false);
     tok = skip(ctx, tok, "}");
     if (!tok.empty() && tok[0] != ";")
       tok = tok.subspan(1);

elf/lto-unix.cc

@@ -641,14 +641,34 @@ ObjectFile<E> *read_lto_object(Context<E> &ctx, MappedFile<Context<E>> *mf) {
 
   // Initialize object symbols
   std::vector<ElfSym<E>> *esyms = new std::vector<ElfSym<E>>(1);
+  obj->has_symver.resize(plugin_symbols.size());
+  obj->lto_symbol_versions.resize(plugin_symbols.size());
 
-  for (PluginSymbol &psym : plugin_symbols) {
+  for (i64 i = 0; i < plugin_symbols.size(); i++) {
+    PluginSymbol &psym = plugin_symbols[i];
     esyms->push_back(to_elf_sym<E>(psym));
-    obj->symbols.push_back(get_symbol(ctx, save_string(ctx, psym.name)));
+
+    std::string_view key = save_string(ctx, psym.name);
+    std::string_view name = key;
+
+    // Parse symbol version after atsign
+    if (i64 pos = name.find('@'); pos != name.npos) {
+      std::string_view ver = name.substr(pos);
+      name = name.substr(0, pos);
+
+      if (ver != "@" && ver != "@@") {
+        if (ver.starts_with("@@"))
+          key = name;
+        obj->has_symver.set(i);
+        obj->lto_symbol_versions[i] = ver.substr(1);
+      }
+    }
+
+    obj->symbols.push_back(get_symbol(ctx, key, name));
   }
 
   obj->elf_syms = *esyms;
-  obj->symvers.resize(esyms->size());
+  obj->has_symver.resize(esyms->size());
   plugin_symbols.clear();
   return obj;
 }
@@ -666,16 +686,13 @@ std::vector<ObjectFile<E> *> do_lto(Context<E> &ctx) {
 
   // Set `referenced_by_regular_obj` bit.
   tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
-    if (file->is_lto_obj)
-      return;
-
-    for (i64 i = file->first_global; i < file->symbols.size(); i++) {
-      Symbol<E> &sym = *file->symbols[i];
-
-      if (sym.file && !sym.file->is_dso &&
-          ((ObjectFile<E> *)sym.file)->is_lto_obj) {
-        std::scoped_lock lock(sym.mu);
-        sym.referenced_by_regular_obj = true;
+    if (!file->is_lto_obj) {
+      for (Symbol<E> *sym : file->get_global_syms()) {
+        if (sym->file && !sym->file->is_dso &&
+            ((ObjectFile<E> *)sym->file)->is_lto_obj) {
+          std::scoped_lock lock(sym->mu);
+          sym->referenced_by_regular_obj = true;
+        }
       }
     }
   });
@@ -692,6 +709,9 @@ std::vector<ObjectFile<E> *> do_lto(Context<E> &ctx) {
     get_symbol(ctx, y)->referenced_by_regular_obj = true;
   }
 
+  // Handle --entry
+  get_symbol(ctx, ctx.arg.entry)->referenced_by_regular_obj = true;
+
   // all_symbols_read_hook() calls add_input_file() and add_input_library()
   LOG << "all symbols read\n";
   if (PluginStatus st = all_symbols_read_hook(); st != LDPS_OK)

elf/lto.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "../integers.h"
+#include "../common/integers.h"
 
 namespace mold {
 
@@ -55,9 +55,7 @@ enum PluginApiVersion {
 
 struct PluginTagValue {
   PluginTagValue(PluginTag tag, int val) : tag(tag), val(val) {}
-
-  template <typename T>
-  PluginTagValue(PluginTag tag, T *ptr) : tag(tag), ptr((void *)ptr) {}
+  PluginTagValue(PluginTag tag, auto *ptr) : tag(tag), ptr((void *)ptr) {}
 
   PluginTag tag;
   union {

elf/main.cc

@@ -1,7 +1,7 @@
 #include "mold.h"
-#include "../archive-file.h"
-#include "../cmdline.h"
-#include "../output-file.h"
+#include "../common/archive-file.h"
+#include "../common/cmdline.h"
+#include "../common/output-file.h"
 
 #include <cstring>
 #include <functional>
@@ -70,26 +70,26 @@ std::string_view get_machine_type(Context<E> &ctx, MappedFile<Context<E>> *mf) {
       return SH4::target_name;
     case EM_ALPHA:
       return ALPHA::target_name;
+    case EM_LOONGARCH:
+      return is_64 ? LOONGARCH64::target_name : LOONGARCH32::target_name;
     default:
       return "";
     }
   };
 
-  bool opt_plugin = !ctx.arg.plugin.empty();
-
-  switch (get_file_type(mf, opt_plugin)) {
+  switch (get_file_type(ctx, mf)) {
   case FileType::ELF_OBJ:
   case FileType::ELF_DSO:
   case FileType::GCC_LTO_OBJ:
     return get_elf_type(mf->data);
   case FileType::AR:
     for (MappedFile<Context<E>> *child : read_fat_archive_members(ctx, mf))
-      if (get_file_type(child, opt_plugin) == FileType::ELF_OBJ)
+      if (get_file_type(ctx, child) == FileType::ELF_OBJ)
         return get_elf_type(child->data);
     return "";
   case FileType::THIN_AR:
     for (MappedFile<Context<E>> *child : read_thin_archive_members(ctx, mf))
-      if (get_file_type(child, opt_plugin) == FileType::ELF_OBJ)
+      if (get_file_type(ctx, child) == FileType::ELF_OBJ)
         return get_elf_type(child->data);
     return "";
   case FileType::TEXT:
@@ -163,9 +163,7 @@ void read_file(Context<E> &ctx, MappedFile<Context<E>> *mf) {
   if (ctx.visited.contains(mf->name))
     return;
 
-  bool opt_plugin = !ctx.arg.plugin.empty();
-  FileType type = get_file_type(mf, opt_plugin);
-  switch (type) {
+  switch (get_file_type(ctx, mf)) {
   case FileType::ELF_OBJ:
     ctx.objs.push_back(new_object_file(ctx, mf, ""));
     return;
@@ -176,7 +174,7 @@ void read_file(Context<E> &ctx, MappedFile<Context<E>> *mf) {
   case FileType::AR:
   case FileType::THIN_AR:
     for (MappedFile<Context<E>> *child : read_archive_members(ctx, mf)) {
-      switch (get_file_type(child, opt_plugin)) {
+      switch (get_file_type(ctx, child)) {
       case FileType::ELF_OBJ:
         ctx.objs.push_back(new_object_file(ctx, child, mf->name));
         break;
@@ -185,6 +183,10 @@ void read_file(Context<E> &ctx, MappedFile<Context<E>> *mf) {
         if (ObjectFile<E> *file = new_lto_obj(ctx, child, mf->name))
           ctx.objs.push_back(file);
         break;
+      case FileType::ELF_DSO:
+        Warn(ctx) << mf->name << "(" << child->name
+                  << "): shared object file in an archive is ignored";
+        break;
       default:
         break;
       }
@@ -339,7 +341,6 @@ static void read_input_files(Context<E> &ctx, std::span<std::string> args) {
 // Since elf_main is a template, we can't run it without a type parameter.
 // We speculatively run elf_main with X86_64, and if the speculation was
 // wrong, re-run it with an actual machine type.
-template <typename E>
 static int redo_main(int argc, char **argv, std::string_view target) {
   if (target == I386::target_name)
     return elf_main<I386>(argc, argv);
@@ -371,6 +372,10 @@ static int redo_main(int argc, char **argv, std::string_view target) {
     return elf_main<SH4>(argc, argv);
   if (target == ALPHA::target_name)
     return elf_main<ALPHA>(argc, argv);
+  if (target == LOONGARCH32::target_name)
+    return elf_main<LOONGARCH32>(argc, argv);
+  if (target == LOONGARCH64::target_name)
+    return elf_main<LOONGARCH64>(argc, argv);
   unreachable();
 }
 
@@ -397,7 +402,7 @@ int elf_main(int argc, char **argv) {
   // Redo if -m is not x86-64.
   if constexpr (is_x86_64<E>)
     if (ctx.arg.emulation != X86_64::target_name)
-      return redo_main<E>(argc, argv, ctx.arg.emulation);
+      return redo_main(argc, argv, ctx.arg.emulation);
 
   Timer t_all(ctx, "all");
 
@@ -416,6 +421,8 @@ int elf_main(int argc, char **argv) {
     on_complete = fork_child();
 #endif
 
+  acquire_global_lock(ctx);
+
   tbb::global_control tbb_cont(tbb::global_control::max_allowed_parallelism,
                                ctx.arg.thread_count);
 
@@ -442,7 +449,6 @@ int elf_main(int argc, char **argv) {
     });
   }
 
-  Timer t_total(ctx, "total");
   Timer t_before_copy(ctx, "before_copy");
 
   // Apply -exclude-libs
@@ -464,6 +470,9 @@ int elf_main(int argc, char **argv) {
   // put together in a single phase.
   resolve_symbols(ctx);
 
+  // "Kill" .eh_frame input sections after symbol resolution.
+  kill_eh_frame_sections(ctx);
+
   // Resolve mergeable section pieces to merge them.
   resolve_section_pieces(ctx);
 
@@ -546,10 +555,8 @@ int elf_main(int argc, char **argv) {
   // Beyond this point, no new symbols will be added to the result.
 
   // Handle --print-dependencies
-  if (ctx.arg.print_dependencies == 1)
+  if (ctx.arg.print_dependencies)
     print_dependencies(ctx);
-  else if (ctx.arg.print_dependencies == 2)
-    print_dependencies_full(ctx);
 
   // Handle -repro
   if (ctx.arg.repro)
@@ -711,7 +718,6 @@ int elf_main(int argc, char **argv) {
   if (ctx.has_lto_object)
     lto_cleanup(ctx);
 
-  t_total.stop();
   t_all.stop();
 
   if (ctx.arg.print_map)
@@ -729,6 +735,8 @@ int elf_main(int argc, char **argv) {
   if (on_complete)
     on_complete();
 
+  release_global_lock(ctx);
+
   if (ctx.arg.quick_exit)
     _exit(0);
 
@@ -738,10 +746,6 @@ int elf_main(int argc, char **argv) {
   return 0;
 }
 
-using E = MOLD_TARGET;
-
-template void read_file(Context<E> &, MappedFile<Context<E>> *);
-
 #ifdef MOLD_X86_64
 
 extern template int elf_main<I386>(int, char **);
@@ -759,6 +763,8 @@ extern template int elf_main<SPARC64>(int, char **);
 extern template int elf_main<M68K>(int, char **);
 extern template int elf_main<SH4>(int, char **);
 extern template int elf_main<ALPHA>(int, char **);
+extern template int elf_main<LOONGARCH32>(int, char **);
+extern template int elf_main<LOONGARCH64>(int, char **);
 
 int main(int argc, char **argv) {
   return elf_main<X86_64>(argc, argv);
@@ -766,6 +772,8 @@ int main(int argc, char **argv) {
 
 #else
 
+using E = MOLD_TARGET;
+
 template int elf_main<E>(int, char **);
 
 #endif

elf/mold-wrapper.c

@@ -61,7 +61,8 @@ static bool is_ld(const char *path) {
     ptr--;
 
   return !strcmp(ptr, "ld") || !strcmp(ptr, "ld.lld") ||
-         !strcmp(ptr, "ld.gold") || !strcmp(ptr, "ld.bfd");
+         !strcmp(ptr, "ld.gold") || !strcmp(ptr, "ld.bfd") ||
+         !strcmp(ptr, "ld.mold");
 }
 
 int execvpe(const char *file, char *const *argv, char *const *envp) {

elf/passes.cc

@@ -4,7 +4,6 @@
 #include <functional>
 #include <map>
 #include <optional>
-#include <random>
 #include <regex>
 #include <shared_mutex>
 #include <tbb/parallel_for_each.h>
@@ -34,7 +33,7 @@ void apply_exclude_libs(Context<E> &ctx) {
 
 template <typename E>
 void create_synthetic_sections(Context<E> &ctx) {
-  auto push = [&]<typename T>(T *x) {
+  auto push = [&](auto *x) {
     ctx.chunks.push_back(x);
     ctx.chunk_pool.emplace_back(x);
     return x;
@@ -58,13 +57,14 @@ void create_synthetic_sections(Context<E> &ctx) {
     else
       ctx.phdr = push(new OutputPhdr<E>(0));
 
-    ctx.shdr = push(new OutputShdr<E>);
+    if (ctx.arg.z_sectionheader)
+      ctx.shdr = push(new OutputShdr<E>);
   }
 
   ctx.got = push(new GotSection<E>);
 
   if constexpr (!is_sparc<E>)
-    ctx.gotplt = push(new GotPltSection<E>);
+    ctx.gotplt = push(new GotPltSection<E>(ctx));
 
   ctx.reldyn = push(new RelDynSection<E>);
   ctx.relplt = push(new RelPltSection<E>);
@@ -82,7 +82,7 @@ void create_synthetic_sections(Context<E> &ctx) {
   ctx.copyrel = push(new CopyrelSection<E>(false));
   ctx.copyrel_relro = push(new CopyrelSection<E>(true));
 
-  if (!ctx.arg.oformat_binary)
+  if (ctx.shdr)
     ctx.shstrtab = push(new ShstrtabSection<E>);
 
   if (!ctx.arg.dynamic_linker.empty())
@@ -105,37 +105,34 @@ void create_synthetic_sections(Context<E> &ctx) {
   if (ctx.arg.emit_relocs)
     ctx.eh_frame_reloc = push(new EhFrameRelocSection<E>);
 
-  if (ctx.arg.shared || !ctx.dsos.empty() || ctx.arg.pie)
+  if (ctx.arg.shared || !ctx.dsos.empty() || ctx.arg.pie) {
     ctx.dynamic = push(new DynamicSection<E>);
 
+    // If .dynamic exists, .dynsym and .dynstr must exist as well
+    // since .dynamic refers to them.
+    ctx.dynstr->shdr.sh_size = 1;
+    ctx.dynsym->symbols.resize(1);
+  }
+
   ctx.versym = push(new VersymSection<E>);
   ctx.verneed = push(new VerneedSection<E>);
   ctx.note_package = push(new NotePackageSection<E>);
   ctx.note_property = push(new NotePropertySection<E>);
 
-  if (ctx.arg.is_static) {
-    if constexpr (is_s390x<E>)
-      ctx.extra.tls_get_offset = push(new S390XTlsGetOffsetSection);
-
-    if constexpr (is_sparc<E>)
-      ctx.extra.tls_get_addr = push(new SparcTlsGetAddrSection);
-  }
+  if constexpr (is_riscv<E>)
+    ctx.extra.riscv_attributes = push(new RiscvAttributesSection<E>);
 
   if constexpr (is_ppc64v1<E>)
     ctx.extra.opd = push(new PPC64OpdSection);
 
-  if constexpr (is_alpha<E>)
-    ctx.extra.got = push(new AlphaGotSection);
-
-  // If .dynamic exists, .dynsym and .dynstr must exist as well
-  // since .dynamic refers them.
-  if (ctx.dynamic) {
-    ctx.dynstr->keep();
-    ctx.dynsym->keep();
+  if constexpr (is_sparc<E>) {
+    if (ctx.arg.is_static)
+      ctx.extra.tls_get_addr_sec = push(new SparcTlsGetAddrSection);
+    ctx.extra.tls_get_addr_sym = get_symbol(ctx, "__tls_get_addr");
   }
 
-  ctx.tls_get_addr = get_symbol(ctx, "__tls_get_addr");
-  ctx.tls_get_offset = get_symbol(ctx, "__tls_get_offset");
+  if constexpr (is_alpha<E>)
+    ctx.extra.got = push(new AlphaGotSection);
 }
 
 template <typename E>
@@ -225,11 +222,16 @@ void do_resolve_symbols(Context<E> &ctx) {
     Timer t(ctx, "eliminate_comdats");
 
     tbb::parallel_for_each(ctx.objs, [](ObjectFile<E> *file) {
-      file->resolve_comdat_groups();
+      for (ComdatGroupRef<E> &ref : file->comdat_groups)
+        update_minimum(ref.group->owner, file->priority);
     });
 
     tbb::parallel_for_each(ctx.objs, [](ObjectFile<E> *file) {
-      file->eliminate_duplicate_comdat_groups();
+      for (ComdatGroupRef<E> &ref : file->comdat_groups)
+        if (ref.group->owner != file->priority)
+          for (u32 i : ref.members)
+            if (file->sections[i])
+              file->sections[i]->kill();
     });
   }
 
@@ -273,12 +275,10 @@ void resolve_symbols(Context<E> &ctx) {
     // Redo name resolution from scratch.
     tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
       file->clear_symbols();
-      file->is_alive = !file->is_in_lib;
     });
 
     tbb::parallel_for_each(ctx.dsos, [&](SharedFile<E> *file) {
       file->clear_symbols();
-      file->is_alive = !file->is_needed;
     });
 
     // Remove IR object files.
@@ -292,6 +292,24 @@ void resolve_symbols(Context<E> &ctx) {
   }
 }
 
+// .eh_frame sections are parsed and regenerated by the linker for the purpose
+// of deduplication and garbage collection. As such, the input sections should
+// not be copied over.
+//
+// However, in very rare cases (e.g. GCC CRT compiled with LTO) we might need
+// to resolve cross-object .eh_frame section references (they only point to
+// begin or end and don't depend on the actual section contents).
+// Therefore, the sections are "killed" after symbol resolution as a separate
+// pass.
+template <typename E>
+void kill_eh_frame_sections(Context<E> &ctx) {
+  Timer t(ctx, "kill_eh_frame_sections");
+
+  for (ObjectFile<E> *file : ctx.objs)
+    if (file->eh_frame_section)
+      file->eh_frame_section->is_alive = false;
+}
+
 template <typename E>
 void resolve_section_pieces(Context<E> &ctx) {
   Timer t(ctx, "resolve_section_pieces");
@@ -331,24 +349,13 @@ void add_comment_string(Context<E> &ctx, std::string str) {
 
   std::string_view buf = save_string(ctx, str);
   std::string_view data(buf.data(), buf.size() + 1);
-  SectionFragment<E> *frag = sec->insert(data, hash_string(data), 0);
-  frag->is_alive = true;
+  sec->insert(ctx, data, hash_string(data), 0);
 }
 
 template <typename E>
 void compute_merged_section_sizes(Context<E> &ctx) {
   Timer t(ctx, "compute_merged_section_sizes");
 
-  // Mark section fragments referenced by live objects.
-  if (!ctx.arg.gc_sections) {
-    tbb::parallel_for_each(ctx.objs, [](ObjectFile<E> *file) {
-      for (std::unique_ptr<MergeableSection<E>> &m : file->mergeable_sections)
-        if (m)
-          for (SectionFragment<E> *frag : m->fragments)
-            frag->is_alive.store(true, std::memory_order_relaxed);
-    });
-  }
-
   // Add an identification string to .comment.
   if (!ctx.arg.oformat_binary)
     add_comment_string(ctx, mold_version);
@@ -357,7 +364,6 @@ void compute_merged_section_sizes(Context<E> &ctx) {
   if (char *env = getenv("MOLD_DEBUG"); env && env[0])
     add_comment_string(ctx, "mold command line: " + get_cmdline_args(ctx));
 
-  Timer t2(ctx, "MergedSection assign_offsets");
   tbb::parallel_for_each(ctx.merged_sections,
                          [&](std::unique_ptr<MergedSection<E>> &sec) {
     sec->assign_offsets(ctx);
@@ -394,13 +400,10 @@ static u64 canonicalize_type(std::string_view name, u64 type) {
 }
 
 struct OutputSectionKey {
+  bool operator==(const OutputSectionKey &) const = default;
   std::string_view name;
   u64 type;
   u64 flags;
-
-  bool operator==(const OutputSectionKey &other) const {
-    return name == other.name && type == other.type && flags == other.flags;
-  }
 };
 
 template <typename E>
@@ -443,7 +446,7 @@ get_output_name(Context<E> &ctx, std::string_view name, u64 flags) {
   static std::string_view prefixes[] = {
     ".text.", ".data.rel.ro.", ".data.", ".rodata.", ".bss.rel.ro.", ".bss.",
     ".init_array.", ".fini_array.", ".tbss.", ".tdata.", ".gcc_except_table.",
-    ".ctors.", ".dtors.", ".gnu.warning.",
+    ".ctors.", ".dtors.", ".gnu.warning.", ".openbsd.randomdata.",
   };
 
   for (std::string_view prefix : prefixes) {
@@ -461,10 +464,7 @@ get_output_section_key(Context<E> &ctx, InputSection<E> &isec) {
   const ElfShdr<E> &shdr = isec.shdr();
   std::string_view name = get_output_name(ctx, isec.name(), shdr.sh_flags);
   u64 type = canonicalize_type<E>(name, shdr.sh_type);
-  u64 flags = shdr.sh_flags & ~(u64)SHF_COMPRESSED;
-
-  if (!ctx.arg.relocatable)
-    flags &= ~(u64)SHF_GROUP & ~(u64)SHF_GNU_RETAIN;
+  u64 flags = shdr.sh_flags & ~(u64)(SHF_COMPRESSED | SHF_GROUP | SHF_GNU_RETAIN);
 
   // .init_array is usually writable. We don't want to create multiple
   // .init_array output sections, so make it always writable.
@@ -479,7 +479,7 @@ template <typename E>
 void create_output_sections(Context<E> &ctx) {
   Timer t(ctx, "create_output_sections");
 
-  struct Cmp {
+  struct Hash {
     size_t operator()(const OutputSectionKey &k) const {
       u64 h = hash_string(k.name);
       h = combine_hash(h, std::hash<u64>{}(k.type));
@@ -488,61 +488,90 @@ void create_output_sections(Context<E> &ctx) {
     }
   };
 
-  std::unordered_map<OutputSectionKey, OutputSection<E> *, Cmp> map;
+  std::unordered_map<OutputSectionKey, OutputSection<E> *, Hash> map;
   std::shared_mutex mu;
 
+  i64 size = ctx.osec_pool.size();
+
   // Instantiate output sections
   tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
+    // Make a per-thread cache of the main map to avoid lock contention.
+    // It makes a noticeable difference if we have millions of input sections.
+    decltype(map) cache;
+    {
+      std::shared_lock lock(mu);
+      cache = map;
+    }
+
     for (std::unique_ptr<InputSection<E>> &isec : file->sections) {
       if (!isec || !isec->is_alive)
         continue;
 
-      OutputSectionKey key = get_output_section_key(ctx, *isec);
-
-      {
-        std::shared_lock lock(mu);
-        auto it = map.find(key);
-        if (it != map.end()) {
-          isec->output_section = it->second;
-          continue;
-        }
+      const ElfShdr<E> &shdr = isec->shdr();
+      if (ctx.arg.relocatable && (shdr.sh_flags & SHF_GROUP)) {
+        OutputSection<E> *osec =
+          new OutputSection<E>(ctx, isec->name(), shdr.sh_type, shdr.sh_flags);
+        isec->output_section = osec;
+        ctx.osec_pool.emplace_back(osec);
+        continue;
       }
 
-      std::unique_ptr<OutputSection<E>> osec =
-        std::make_unique<OutputSection<E>>(key.name, key.type, key.flags);
-      std::unique_lock lock(mu);
+      OutputSectionKey key = get_output_section_key(ctx, *isec);
 
-      auto [it, inserted] = map.insert({key, osec.get()});
-      isec->output_section = it->second;
-      if (inserted)
-        ctx.osec_pool.emplace_back(std::move(osec));
+      if (auto it = cache.find(key); it != cache.end()) {
+        isec->output_section = it->second;
+        continue;
+      }
+
+      auto get_or_insert = [&] {
+        {
+          std::shared_lock lock(mu);
+          if (auto it = map.find(key); it != map.end())
+            return it->second;
+        }
+
+        std::unique_ptr<OutputSection<E>> osec =
+          std::make_unique<OutputSection<E>>(ctx, key.name, key.type, key.flags);
+
+        std::unique_lock lock(mu);
+        auto [it, inserted] = map.insert({key, osec.get()});
+        OutputSection<E> *ret = it->second;
+
+        if (inserted)
+          ctx.osec_pool.emplace_back(std::move(osec));
+        return ret;
+      };
+
+      OutputSection<E> *osec = get_or_insert();
+      isec->output_section = osec;
+      cache.insert({key, osec});
     }
   });
 
   // Add input sections to output sections
+  std::vector<Chunk<E> *> chunks;
+  for (i64 i = size; i < ctx.osec_pool.size(); i++)
+    chunks.push_back(ctx.osec_pool[i].get());
+
   for (ObjectFile<E> *file : ctx.objs)
     for (std::unique_ptr<InputSection<E>> &isec : file->sections)
       if (isec && isec->is_alive)
         isec->output_section->members.push_back(isec.get());
 
   // Add output sections and mergeable sections to ctx.chunks
-  std::vector<Chunk<E> *> vec;
-  for (std::pair<const OutputSectionKey, OutputSection<E> *> &kv : map)
-    vec.push_back(kv.second);
-
   for (std::unique_ptr<MergedSection<E>> &osec : ctx.merged_sections)
     if (osec->shdr.sh_size)
-      vec.push_back(osec.get());
+      chunks.push_back(osec.get());
 
   // Sections are added to the section lists in an arbitrary order
   // because they are created in parallel. Sort them to to make the
   // output deterministic.
-  tbb::parallel_sort(vec.begin(), vec.end(), [](Chunk<E> *x, Chunk<E> *y) {
+  tbb::parallel_sort(chunks.begin(), chunks.end(), [](Chunk<E> *x, Chunk<E> *y) {
     return std::tuple(x->name, x->shdr.sh_type, x->shdr.sh_flags) <
            std::tuple(y->name, y->shdr.sh_type, y->shdr.sh_flags);
   });
 
-  append(ctx.chunks, vec);
+  append(ctx.chunks, chunks);
 }
 
 // Create a dummy object file containing linker-synthesized
@@ -615,7 +644,7 @@ void create_internal_file(Context<E> &ctx) {
       add(get_symbol(ctx, ord.name));
 
   obj->elf_syms = ctx.internal_esyms;
-  obj->symvers.resize(ctx.internal_esyms.size() - 1);
+  obj->has_symver.resize(ctx.internal_esyms.size() - 1);
 }
 
 template <typename E>
@@ -628,7 +657,7 @@ get_start_stop_name(Context<E> &ctx, Chunk<E> &chunk) {
     if (ctx.arg.start_stop) {
       auto isalnum = [](char c) {
         return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') ||
-          ('0' <= c && c <= '9');
+               ('0' <= c && c <= '9');
       };
 
       std::string s{chunk.name};
@@ -649,10 +678,10 @@ template <typename E>
 void add_synthetic_symbols(Context<E> &ctx) {
   ObjectFile<E> &obj = *ctx.internal_obj;
 
-  auto add = [&](std::string_view name) {
+  auto add = [&](std::string_view name, u32 type = STT_NOTYPE) {
     ElfSym<E> esym;
     memset(&esym, 0, sizeof(esym));
-    esym.st_type = STT_NOTYPE;
+    esym.st_type = type;
     esym.st_shndx = SHN_ABS;
     esym.st_bind = STB_GLOBAL;
     esym.st_visibility = STV_HIDDEN;
@@ -698,7 +727,7 @@ void add_synthetic_symbols(Context<E> &ctx) {
     ctx.__dso_handle = add("__dso_handle");
 
   if constexpr (supports_tlsdesc<E>)
-    ctx._TLS_MODULE_BASE_ = add("_TLS_MODULE_BASE_");
+    ctx._TLS_MODULE_BASE_ = add("_TLS_MODULE_BASE_", STT_TLS);
 
   if constexpr (is_riscv<E>)
     if (!ctx.arg.shared)
@@ -728,7 +757,7 @@ void add_synthetic_symbols(Context<E> &ctx) {
   }
 
   obj.elf_syms = ctx.internal_esyms;
-  obj.symvers.resize(ctx.internal_esyms.size() - 1);
+  obj.has_symver.resize(ctx.internal_esyms.size() - 1);
 
   obj.resolve_symbols(ctx);
 
@@ -765,19 +794,20 @@ void add_synthetic_symbols(Context<E> &ctx) {
 template <typename E>
 void check_cet_errors(Context<E> &ctx) {
   bool warning = (ctx.arg.z_cet_report == CET_REPORT_WARNING);
-  assert(warning || (ctx.arg.z_cet_report == CET_REPORT_ERROR));
+  assert(warning || ctx.arg.z_cet_report == CET_REPORT_ERROR);
 
   auto has_feature = [](ObjectFile<E> *file, u32 feature) {
     return std::any_of(file->gnu_properties.begin(), file->gnu_properties.end(),
-                       [&](auto kv) {
-                         return kv.first == GNU_PROPERTY_X86_FEATURE_1_AND
-                             && (kv.second & feature);
-                       });
+                       [&](std::pair<u32, u32> kv) {
+      return kv.first == GNU_PROPERTY_X86_FEATURE_1_AND &&
+             (kv.second & feature);
+    });
   };
 
   for (ObjectFile<E> *file : ctx.objs) {
     if (file == ctx.internal_obj)
       continue;
+
     if (!has_feature(file, GNU_PROPERTY_X86_FEATURE_1_IBT)) {
       if (warning)
         Warn(ctx) << *file << ": -cet-report=warning: "
@@ -803,37 +833,14 @@ void print_dependencies(Context<E> &ctx) {
   SyncOut(ctx) <<
 R"(# This is an output of the mold linker's --print-dependencies option.
 #
-# Each line consists of three fields, <file1>, <file2> and <symbol>
-# separated by tab characters. It indicates that <file1> depends on
-# <file2> to use <symbol>.)";
-
-  auto print = [&](InputFile<E> *file) {
-    for (i64 i = file->first_global; i < file->elf_syms.size(); i++) {
-      ElfSym<E> &esym = file->elf_syms[i];
-      Symbol<E> &sym = *file->symbols[i];
-      if (esym.is_undef() && sym.file && sym.file != file)
-        SyncOut(ctx) << *file << "\t" << *sym.file << "\t" << sym;
-    }
-  };
-
-  for (InputFile<E> *file : ctx.objs)
-    print(file);
-  for (InputFile<E> *file : ctx.dsos)
-    print(file);
-}
-
-template <typename E>
-void print_dependencies_full(Context<E> &ctx) {
-  SyncOut(ctx) <<
-R"(# This is an output of the mold linker's --print-dependencies=full option.
-#
 # Each line consists of 4 fields, <section1>, <section2>, <symbol-type> and
 # <symbol>, separated by tab characters. It indicates that <section1> depends
 # on <section2> to use <symbol>. <symbol-type> is either "u" or "w" for
 # regular undefined or weak undefined, respectively.
 #
 # If you want to obtain dependency information per function granularity,
-# compile source files with the -ffunction-sections compiler flag.)";
+# compile source files with the -ffunction-sections compiler flag.
+)";
 
   auto println = [&](auto &src, Symbol<E> &sym, ElfSym<E> &esym) {
     if (InputSection<E> *isec = sym.get_input_section())
@@ -962,33 +969,30 @@ template <typename E>
 void check_symbol_types(Context<E> &ctx) {
   Timer t(ctx, "check_symbol_types");
 
-  auto check = [&](InputFile<E> *file) {
-    for (i64 i = file->first_global; i < file->elf_syms.size(); i++) {
-      const ElfSym<E> &esym = file->elf_syms[i];
-      Symbol<E> &sym = *file->symbols[i];
+  std::vector<InputFile<E> *> files;
+  append(files, ctx.objs);
+  append(files, ctx.dsos);
 
-      if (!sym.file)
+  tbb::parallel_for_each(files.begin(), files.end(), [&](InputFile<E> *file) {
+    for (i64 i = file->first_global; i < file->elf_syms.size(); i++) {
+      Symbol<E> &sym = *file->symbols[i];
+      if (!sym.file || sym.file == file)
         continue;
 
-      u32 x = sym.esym().st_type;
-      if (x == STT_GNU_IFUNC)
-        x = STT_FUNC;
+      const ElfSym<E> &esym1 = sym.esym();
+      const ElfSym<E> &esym2 = file->elf_syms[i];
 
-      u32 y = esym.st_type;
-      if (y == STT_GNU_IFUNC)
-        y = STT_FUNC;
+      u32 ty1 = (esym1.st_type == STT_GNU_IFUNC) ? STT_FUNC : esym1.st_type;
+      u32 ty2 = (esym2.st_type == STT_GNU_IFUNC) ? STT_FUNC : esym2.st_type;
 
-      if (x != STT_NOTYPE && y != STT_NOTYPE && x != y)
+      if (ty1 != STT_NOTYPE && ty2 != STT_NOTYPE && ty1 != ty2)
         Warn(ctx) << "symbol type mismatch: " << sym << '\n'
                   << ">>> defined in " << *sym.file << " as "
-                  << stt_to_string<E>(sym.esym().st_type) << '\n'
+                  << stt_to_string<E>(esym1.st_type) << '\n'
                   << ">>> defined in " << *file << " as "
-                  << stt_to_string<E>(esym.st_type);
+                  << stt_to_string<E>(esym2.st_type);
     }
-  };
-
-  tbb::parallel_for_each(ctx.objs, check);
-  tbb::parallel_for_each(ctx.dsos, check);
+  });
 }
 
 template <typename E>
@@ -1011,8 +1015,12 @@ void sort_init_fini(Context<E> &ctx) {
         if (ctx.arg.shuffle_sections == SHUFFLE_SECTIONS_REVERSE)
           std::reverse(osec->members.begin(), osec->members.end());
 
+        std::unordered_map<InputSection<E> *, i64> map;
+        for (InputSection<E> *isec : osec->members)
+          map.insert({isec, get_priority(isec)});
+
         sort(osec->members, [&](InputSection<E> *a, InputSection<E> *b) {
-          return get_priority(a) < get_priority(b);
+          return map[a] < map[b];
         });
       }
     }
@@ -1050,8 +1058,12 @@ void sort_ctor_dtor(Context<E> &ctx) {
         if (ctx.arg.shuffle_sections != SHUFFLE_SECTIONS_REVERSE)
           std::reverse(osec->members.begin(), osec->members.end());
 
+        std::unordered_map<InputSection<E> *, i64> map;
+        for (InputSection<E> *isec : osec->members)
+          map.insert({isec, get_priority(isec)});
+
         sort(osec->members, [&](InputSection<E> *a, InputSection<E> *b) {
-          return get_priority(a) < get_priority(b);
+          return map[a] < map[b];
         });
       }
     }
@@ -1088,37 +1100,35 @@ template <typename E>
 void shuffle_sections(Context<E> &ctx) {
   Timer t(ctx, "shuffle_sections");
 
-  auto is_eligible = [](OutputSection<E> &osec) {
-    return osec.name != ".init" && osec.name != ".fini" &&
-           osec.name != ".ctors" && osec.name != ".dtors" &&
-           osec.name != ".init_array" && osec.name != ".preinit_array" &&
-           osec.name != ".fini_array";
+  auto is_eligible = [](OutputSection<E> *osec) {
+    if (osec) {
+      std::string_view name = osec->name;
+      return name != ".init" && name != ".fini" &&
+             name != ".ctors" && name != ".dtors" &&
+             name != ".init_array" && name != ".preinit_array" &&
+             name != ".fini_array";
+    }
+    return false;
   };
 
   switch (ctx.arg.shuffle_sections) {
-  case SHUFFLE_SECTIONS_NONE:
-    unreachable();
   case SHUFFLE_SECTIONS_SHUFFLE: {
-    u64 seed;
-    if (ctx.arg.shuffle_sections_seed)
-      seed = *ctx.arg.shuffle_sections_seed;
-    else
-      seed = ((u64)std::random_device()() << 32) | std::random_device()();
-
     tbb::parallel_for_each(ctx.chunks, [&](Chunk<E> *chunk) {
-      if (OutputSection<E> *osec = chunk->to_osec())
-        if (is_eligible(*osec))
-          shuffle(osec->members, seed + hash_string(osec->name));
+      if (OutputSection<E> *osec = chunk->to_osec(); is_eligible(osec)) {
+        u64 seed = ctx.arg.shuffle_sections_seed + hash_string(osec->name);
+        shuffle(osec->members, seed);
+      }
     });
     break;
   }
   case SHUFFLE_SECTIONS_REVERSE:
     tbb::parallel_for_each(ctx.chunks, [&](Chunk<E> *chunk) {
-      if (OutputSection<E> *osec = chunk->to_osec())
-        if (is_eligible(*osec))
-          std::reverse(osec->members.begin(), osec->members.end());
+      if (OutputSection<E> *osec = chunk->to_osec(); is_eligible(osec))
+        std::reverse(osec->members.begin(), osec->members.end());
     });
     break;
+  default:
+    unreachable();
   }
 }
 
@@ -1159,19 +1169,16 @@ void compute_section_sizes(Context<E> &ctx) {
       }
     });
 
-    i64 offset = 0;
-    i64 p2align = 0;
+    ElfShdr<E> &shdr = osec->shdr;
+    shdr.sh_size = 0;
 
     for (i64 i = 0; i < groups.size(); i++) {
-      offset = align_to(offset, 1 << groups[i].p2align);
-      groups[i].offset = offset;
-      offset += groups[i].size;
-      p2align = std::max(p2align, groups[i].p2align);
+      shdr.sh_size = align_to(shdr.sh_size, 1 << groups[i].p2align);
+      groups[i].offset = shdr.sh_size;
+      shdr.sh_size += groups[i].size;
+      shdr.sh_addralign = std::max<u32>(shdr.sh_addralign, 1 << groups[i].p2align);
     }
 
-    osec->shdr.sh_size = offset;
-    osec->shdr.sh_addralign = 1 << p2align;
-
     // Assign offsets to input sections.
     tbb::parallel_for_each(groups, [](Group &group) {
       i64 offset = group.offset;
@@ -1191,30 +1198,103 @@ void compute_section_sizes(Context<E> &ctx) {
   // inserting thunks. This pass cannot be parallelized. That is,
   // create_range_extension_thunks is parallelized internally, but the
   // function itself is not thread-safe.
-  if constexpr (needs_thunk<E>) {
-    for (Chunk<E> *chunk : ctx.chunks) {
-      OutputSection<E> *osec = chunk->to_osec();
-      if (osec && (osec->shdr.sh_flags & SHF_EXECINSTR) && !ctx.arg.relocatable) {
-        create_range_extension_thunks(ctx, *osec);
-
-        for (InputSection<E> *isec : osec->members)
-          osec->shdr.sh_addralign =
-            std::max<u32>(osec->shdr.sh_addralign, 1 << isec->p2align);
-      }
-    }
-  }
-
-  for (Chunk<E> *chunk : ctx.chunks)
-    if (OutputSection<E> *osec = chunk->to_osec())
-      if (u32 align = ctx.arg.section_align[osec->name])
-        osec->shdr.sh_addralign = std::max<u32>(osec->shdr.sh_addralign, align);
+  if constexpr (needs_thunk<E>)
+    if (!ctx.arg.relocatable)
+      for (Chunk<E> *chunk : ctx.chunks)
+        if (OutputSection<E> *osec = chunk->to_osec())
+          if (osec->shdr.sh_flags & SHF_EXECINSTR)
+            osec->create_range_extension_thunks(ctx);
 }
 
+// Find all unresolved symbols and attach them to the most appropriate files.
+// Note that even a symbol that will be reported as an undefined symbol will
+// get an owner file in this function. Such symbol will be reported by
+// ObjectFile<E>::scan_relocations().
 template <typename E>
 void claim_unresolved_symbols(Context<E> &ctx) {
   Timer t(ctx, "claim_unresolved_symbols");
+
   tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
-    file->claim_unresolved_symbols(ctx);
+    if (!file->is_alive)
+      return;
+
+    for (i64 i = file->first_global; i < file->elf_syms.size(); i++) {
+      const ElfSym<E> &esym = file->elf_syms[i];
+      Symbol<E> &sym = *file->symbols[i];
+      if (!esym.is_undef())
+        continue;
+
+      std::scoped_lock lock(sym.mu);
+
+      if (sym.file)
+        if (!sym.esym().is_undef() || sym.file->priority <= file->priority)
+          continue;
+
+      // If a symbol name is in the form of "foo@version", search for
+      // symbol "foo" and check if the symbol has version "version".
+      if (file->has_symver.get(i - file->first_global)) {
+        std::string_view str = file->symbol_strtab.data() + esym.st_name;
+        i64 pos = str.find('@');
+        assert(pos != str.npos);
+
+        std::string_view name = str.substr(0, pos);
+        std::string_view ver = str.substr(pos + 1);
+
+        Symbol<E> *sym2 = get_symbol(ctx, name);
+        if (sym2->file && sym2->file->is_dso && sym2->get_version() == ver) {
+          file->symbols[i] = sym2;
+          sym2->is_imported = true;
+          continue;
+        }
+      }
+
+      auto claim = [&](bool is_imported) {
+        if (sym.is_traced)
+          SyncOut(ctx) << "trace-symbol: " << *file << ": unresolved"
+                       << (esym.is_weak() ? " weak" : "")
+                       << " symbol " << sym;
+
+        sym.file = file;
+        sym.origin = 0;
+        sym.value = 0;
+        sym.sym_idx = i;
+        sym.is_weak = false;
+        sym.is_imported = is_imported;
+        sym.is_exported = false;
+        sym.ver_idx = is_imported ? 0 : ctx.default_version;
+      };
+
+      if (esym.is_undef_weak()) {
+        if (ctx.arg.shared && sym.visibility != STV_HIDDEN &&
+            ctx.arg.z_dynamic_undefined_weak) {
+          // Global weak undefined symbols are promoted to dynamic symbols
+          // when linking a DSO unless `-z nodynamic_undefined_weak` was given.
+          claim(true);
+        } else {
+          // Otherwise, weak undefs are converted to absolute symbols with value 0.
+          claim(false);
+        }
+        continue;
+      }
+
+      // Traditionally, remaining undefined symbols cause a link failure
+      // only when we are creating an executable. Undefined symbols in
+      // shared objects are promoted to dynamic symbols, so that they'll
+      // get another chance to be resolved at run-time. You can change the
+      // behavior by passing `-z defs` to the linker.
+      //
+      // Even if `-z defs` is given, weak undefined symbols are still
+      // promoted to dynamic symbols for compatibility with other linkers.
+      // Some major programs, notably Firefox, depend on the behavior
+      // (they use this loophole to export symbols from libxul.so).
+      if (ctx.arg.shared && sym.visibility != STV_HIDDEN && !ctx.arg.z_defs) {
+        claim(true);
+        continue;
+      }
+
+      // Convert remaining undefined symbols to absolute symbols with value 0.
+      claim(false);
+    }
   });
 }
 
@@ -1247,17 +1327,12 @@ void scan_relocations(Context<E> &ctx) {
   std::vector<Symbol<E> *> syms = flatten(vec);
   ctx.symbol_aux.reserve(syms.size());
 
-  auto add_aux = [&](Symbol<E> *sym) {
-    if (sym->aux_idx == -1) {
-      i64 sz = ctx.symbol_aux.size();
-      sym->aux_idx = sz;
-      ctx.symbol_aux.resize(sz + 1);
-    }
-  };
+  if (ctx.needs_tlsld)
+    ctx.got->add_tlsld(ctx);
 
   // Assign offsets in additional tables for each dynamic symbol.
   for (Symbol<E> *sym : syms) {
-    add_aux(sym);
+    sym->add_aux(ctx);
 
     if (sym->is_imported || sym->is_exported)
       ctx.dynsym->add_symbol(ctx, sym);
@@ -1272,7 +1347,7 @@ void scan_relocations(Context<E> &ctx) {
       sym->is_exported = true;
 
       // We can't use .plt.got for a canonical PLT because otherwise
-      // .plt.got and .got would refer each other, resulting in an
+      // .plt.got and .got would refer to each other, resulting in an
       // infinite loop at runtime.
       ctx.plt->add_symbol(ctx, sym);
     } else if (sym->flags & NEEDS_PLT) {
@@ -1292,31 +1367,10 @@ void scan_relocations(Context<E> &ctx) {
       ctx.got->add_tlsdesc_symbol(ctx, sym);
 
     if (sym->flags & NEEDS_COPYREL) {
-      assert(sym->file->is_dso);
-      SharedFile<E> *file = (SharedFile<E> *)sym->file;
-      sym->copyrel_readonly = file->is_readonly(ctx, sym);
-
-      if (sym->copyrel_readonly)
+      if (((SharedFile<E> *)sym->file)->is_readonly(sym))
         ctx.copyrel_relro->add_symbol(ctx, sym);
       else
         ctx.copyrel->add_symbol(ctx, sym);
-
-      // If a symbol needs copyrel, it is considered both imported
-      // and exported.
-      assert(sym->is_imported);
-      sym->is_exported = true;
-
-      // Aliases of this symbol are also copied so that they will be
-      // resolved to the same address at runtime.
-      for (Symbol<E> *alias : file->find_aliases(sym)) {
-        add_aux(alias);
-        alias->is_imported = true;
-        alias->is_exported = true;
-        alias->has_copyrel = true;
-        alias->value = sym->value;
-        alias->copyrel_readonly = sym->copyrel_readonly;
-        ctx.dynsym->add_symbol(ctx, alias);
-      }
     }
 
     if constexpr (is_ppc64v1<E>)
@@ -1326,9 +1380,6 @@ void scan_relocations(Context<E> &ctx) {
     sym->flags = 0;
   }
 
-  if (ctx.needs_tlsld)
-    ctx.got->add_tlsld(ctx);
-
   if constexpr (is_alpha<E>)
     ctx.extra.got->finalize();
 
@@ -1336,6 +1387,36 @@ void scan_relocations(Context<E> &ctx) {
     Warn(ctx) << "creating a DT_TEXTREL in an output file";
 }
 
+// Report all undefined symbols, grouped by symbol.
+template <typename E>
+void report_undef_errors(Context<E> &ctx) {
+  constexpr i64 max_errors = 3;
+
+  for (auto &pair : ctx.undef_errors) {
+    std::string_view sym_name = pair.first;
+    std::span<std::string> errors = pair.second;
+
+    if (ctx.arg.demangle)
+      sym_name = demangle(sym_name);
+
+    std::stringstream ss;
+    ss << "undefined symbol: " << sym_name << "\n";
+
+    for (i64 i = 0; i < errors.size() && i < max_errors; i++)
+      ss << errors[i];
+
+    if (errors.size() > max_errors)
+      ss << ">>> referenced " << (errors.size() - max_errors) << " more times\n";
+
+    if (ctx.arg.unresolved_symbols == UNRESOLVED_ERROR)
+      Error(ctx) << ss.str();
+    else if (ctx.arg.unresolved_symbols == UNRESOLVED_WARN)
+      Warn(ctx) << ss.str();
+  }
+
+  ctx.checkpoint();
+}
+
 template <typename E>
 void create_reloc_sections(Context<E> &ctx) {
   Timer t(ctx, "create_reloc_sections");
@@ -1376,6 +1457,10 @@ void copy_chunks(Context<E> &ctx) {
       copy(*chunk);
   });
 
+  // Undefined symbols in SHF_ALLOC sections are found by scan_relocations(),
+  // but those in non-SHF_ALLOC sections cannot be found until we copy section
+  // contents. So we need to call this function again to report possible
+  // undefined errors.
   report_undef_errors(ctx);
 
   if constexpr (is_arm32<E>)
@@ -1387,11 +1472,8 @@ void construct_relr(Context<E> &ctx) {
   Timer t(ctx, "construct_relr");
 
   tbb::parallel_for_each(ctx.chunks, [&](Chunk<E> *chunk) {
-    if (OutputSection<E> *osec = chunk->to_osec())
-      osec->construct_relr(ctx);
+    chunk->construct_relr(ctx);
   });
-
-  ctx.got->construct_relr(ctx);
 }
 
 template <typename E>
@@ -1492,17 +1574,23 @@ void parse_symbol_version(Context<E> &ctx) {
     verdefs[ctx.arg.version_definitions[i]] = i + VER_NDX_LAST_RESERVED + 1;
 
   tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
-    for (i64 i = 0; i < file->elf_syms.size() - file->first_global; i++) {
+    for (i64 i = file->first_global; i < file->elf_syms.size(); i++) {
       // Match VERSION part of symbol foo@VERSION with version definitions.
-      // The symbols' VERSION parts are in file->symvers.
-      if (!file->symvers[i])
+      if (!file->has_symver.get(i - file->first_global))
         continue;
 
-      Symbol<E> *sym = file->symbols[i + file->first_global];
+      Symbol<E> *sym = file->symbols[i];
       if (sym->file != file)
         continue;
 
-      std::string_view ver = file->symvers[i];
+      std::string_view ver;
+
+      if (file->is_lto_obj) {
+        ver = file->lto_symbol_versions[i - file->first_global];
+      } else {
+        const char *name = file->symbol_strtab.data() + file->elf_syms[i].st_name;
+        ver = strchr(name, '@') + 1;
+      }
 
       bool is_default = false;
       if (ver.starts_with('@')) {
@@ -1526,7 +1614,8 @@ void parse_symbol_version(Context<E> &ctx) {
       // versioned symbol. Likewise, if `foo@VERSION` and `foo@@VERSION` are
       // defined, the default one takes precedence.
       Symbol<E> *sym2 = get_symbol(ctx, sym->name());
-      if (sym2->file == file && !file->symvers[sym2->sym_idx - file->first_global])
+      if (sym2->file == file &&
+          !file->has_symver.get(sym2->sym_idx - file->first_global))
         if (sym2->ver_idx == ctx.default_version ||
             (sym2->ver_idx & ~VERSYM_HIDDEN) == (sym->ver_idx & ~VERSYM_HIDDEN))
           sym2->ver_idx = VER_NDX_LOCAL;
@@ -1544,7 +1633,8 @@ void compute_import_export(Context<E> &ctx) {
     tbb::parallel_for_each(ctx.dsos, [&](SharedFile<E> *file) {
       for (Symbol<E> *sym : file->symbols) {
         if (sym->file && !sym->file->is_dso && sym->visibility != STV_HIDDEN) {
-          if (sym->ver_idx != VER_NDX_LOCAL || !ctx.default_version_from_version_script) {
+          if (sym->ver_idx != VER_NDX_LOCAL ||
+              !ctx.default_version_from_version_script) {
             std::scoped_lock lock(sym->mu);
             sym->is_exported = true;
           }
@@ -1701,13 +1791,14 @@ void sort_output_sections_regular(Context<E> &ctx) {
     bool writable = (flags & SHF_WRITE);
     bool exec = (flags & SHF_EXECINSTR);
     bool tls = (flags & SHF_TLS);
-    bool relro = is_relro(ctx, chunk);
+    bool relro = chunk->is_relro;
     bool is_bss = (type == SHT_NOBITS);
 
     return (1 << 10) | (!alloc << 9) | (writable << 8) | (exec << 7) |
            (!tls << 6) | (!relro << 5) | (is_bss << 4);
   };
 
+  // Ties are broken by additional rules
   auto get_rank2 = [&](Chunk<E> *chunk) -> i64 {
     if (chunk->shdr.sh_type == SHT_NOTE)
       return -chunk->shdr.sh_addralign;
@@ -1724,14 +1815,8 @@ void sort_output_sections_regular(Context<E> &ctx) {
   };
 
   sort(ctx.chunks, [&](Chunk<E> *a, Chunk<E> *b) {
-    // Sort sections by segments
-    i64 x = get_rank1(a);
-    i64 y = get_rank1(b);
-    if (x != y)
-      return x < y;
-
-    // Ties are broken by additional rules
-    return get_rank2(a) < get_rank2(b);
+    return std::tuple{get_rank1(a), get_rank2(a)} <
+           std::tuple{get_rank1(b), get_rank2(b)};
   });
 }
 
@@ -1840,7 +1925,7 @@ static void set_virtual_addresses_regular(Context<E> &ctx) {
 
   auto get_flags = [&](Chunk<E> *chunk) {
     i64 flags = to_phdr_flags(ctx, chunk);
-    if (is_relro(ctx, chunk))
+    if (chunk->is_relro)
       return flags | RELRO;
     return flags;
   };
@@ -2147,17 +2232,18 @@ i64 set_osec_offsets(Context<E> &ctx) {
     else
       set_virtual_addresses_by_order(ctx);
 
-    i64 fileoff = set_file_offsets(ctx);
-
     // Assigning new offsets may change the contents and the length
     // of the program header, so repeat it until converge.
-    if (!ctx.phdr)
-      return fileoff;
+    i64 fileoff = set_file_offsets(ctx);
 
-    i64 sz = ctx.phdr->shdr.sh_size;
-    ctx.phdr->update_shdr(ctx);
-    if (sz == ctx.phdr->shdr.sh_size)
-      return fileoff;
+    if (ctx.phdr) {
+      i64 sz = ctx.phdr->shdr.sh_size;
+      ctx.phdr->update_shdr(ctx);
+      if (sz != ctx.phdr->shdr.sh_size)
+        continue;
+    }
+
+    return fileoff;
   }
 }
 
@@ -2296,7 +2382,7 @@ void fix_synthetic_symbols(Context<E> &ctx) {
   // this symbol.
   if (ctx._TLS_MODULE_BASE_) {
     ctx._TLS_MODULE_BASE_->set_output_section(sections[0]);
-    ctx._TLS_MODULE_BASE_->value = ctx.tls_begin;
+    ctx._TLS_MODULE_BASE_->value = ctx.dtp_addr;
   }
 
   // __GNU_EH_FRAME_HDR
@@ -2397,7 +2483,8 @@ i64 compress_debug_sections(Context<E> &ctx) {
     ctx.chunks[i] = comp;
   });
 
-  ctx.shstrtab->update_shdr(ctx);
+  if (ctx.shstrtab)
+    ctx.shstrtab->update_shdr(ctx);
 
   if (ctx.ehdr)
     ctx.ehdr->update_shdr(ctx);
@@ -2509,6 +2596,7 @@ template void create_internal_file(Context<E> &);
 template void apply_exclude_libs(Context<E> &);
 template void create_synthetic_sections(Context<E> &);
 template void resolve_symbols(Context<E> &);
+template void kill_eh_frame_sections(Context<E> &);
 template void resolve_section_pieces(Context<E> &);
 template void convert_common_symbols(Context<E> &);
 template void compute_merged_section_sizes(Context<E> &);
@@ -2516,7 +2604,6 @@ template void create_output_sections(Context<E> &);
 template void add_synthetic_symbols(Context<E> &);
 template void check_cet_errors(Context<E> &);
 template void print_dependencies(Context<E> &);
-template void print_dependencies_full(Context<E> &);
 template void write_repro_file(Context<E> &);
 template void check_duplicate_symbols(Context<E> &);
 template void check_symbol_types(Context<E> &);
@@ -2527,6 +2614,7 @@ template void compute_section_sizes(Context<E> &);
 template void sort_output_sections(Context<E> &);
 template void claim_unresolved_symbols(Context<E> &);
 template void scan_relocations(Context<E> &);
+template void report_undef_errors(Context<E> &);
 template void create_reloc_sections(Context<E> &);
 template void copy_chunks(Context<E> &);
 template void construct_relr(Context<E> &);

elf/relocatable.cc

@@ -40,7 +40,7 @@ namespace mold::elf {
 // Create linker-synthesized sections
 template <typename E>
 static void r_create_synthetic_sections(Context<E> &ctx) {
-  auto push = [&]<typename T>(T *x) {
+  auto push = [&](auto *x) {
     ctx.chunks.push_back(x);
     ctx.chunk_pool.emplace_back(x);
     return x;
@@ -54,6 +54,9 @@ static void r_create_synthetic_sections(Context<E> &ctx) {
   ctx.symtab = push(new SymtabSection<E>);
   ctx.shstrtab = push(new ShstrtabSection<E>);
   ctx.note_property = push(new NotePropertySection<E>);
+
+  if constexpr (is_riscv<E>)
+    ctx.extra.riscv_attributes = push(new RiscvAttributesSection<E>);
 }
 
 // Create SHT_GROUP (i.e. comdat group) sections. We uniquify comdat

elf/thunks.cc

@@ -20,6 +20,9 @@
 // we don't need to try too hard to reduce thunk size to the absolute
 // minimum.
 
+#if MOLD_ARM32 || MOLD_ARM64 || MOLD_PPC32 || MOLD_PPC64V1 || MOLD_PPC64V2 || \
+    MOLD_LOONGARCH64 || MOLD_LOONGARCH32
+
 #include "mold.h"
 
 #include <tbb/parallel_for.h>
@@ -27,37 +30,44 @@
 
 namespace mold::elf {
 
-// Branch reach in bytes.
-//
-// ARM64's branch has 26 bits immediate, and it's scaled by 4 because all
-// instructions are 4 bytes aligned, so it's effectively 28 bits long.
-// That means the range is [-2^27, 2^27).
-//
-// ARM32's Thumb branch has 24 bits immediate, and the instructions are
-// aligned to 2, so it's effectively 25 bits. ARM32's non-Thumb branches
-// have twice longer range than its Thumb counterparts, but we
-// conservatively use the Thumb's limitation.
-//
-// PPC's branch has 24 bits immediate, and the instructions are aligned
-// to 4, therefore 26.
-//
-// Here is the summary of branch instructions reaches:
-//
-//   ARM64: PC ± 128 MiB
-//   ARM32: PC ± 16 MiB
-//   PPC:   PC ± 32 MiB
-template <typename E>
-static constexpr i64 max_distance =
-  1LL << (is_arm64<E> ? 27 : is_arm32<E> ? 24 : 25);
+using E = MOLD_TARGET;
+
+// Returns a branch reach in bytes for a given target.
+static consteval i64 max_distance() {
+  // ARM64's branch has 26 bits immediate. The immediate is padded with
+  // implicit two-bit zeros because all instructions are 4 bytes aligned
+  // and therefore the least two bits are always zero. So the branch
+  // operand is effectively 28 bits long. That means the branch range is
+  // [-2^27, 2^27) or PC ± 128 MiB.
+  //
+  // LoongArch's BR instruction also takes a 26 bit immediate.
+  if (is_arm64<E> || is_loongarch<E>)
+    return 1 << 27;
+
+  // ARM32's Thumb branch has 24 bits immediate, and the instructions are
+  // aligned to 2, so it's effectively 25 bits. It's [-2^24, 2^24) or PC ±
+  // 16 MiB.
+  //
+  // ARM32's non-Thumb branches have twice longer range than its Thumb
+  // counterparts, but we conservatively use the Thumb's limitation.
+  if (is_arm32<E>)
+    return 1 << 24;
+
+  // PPC's branch has 24 bits immediate, and the instructions are aligned
+  // to 4, therefore the reach is [-2^25, 2^25) or PC ± 32 MiB.
+  assert(is_ppc<E>);
+  return 1 << 25;
+}
 
 // We create thunks for each 12.8/1.6/3.2 MiB code block for
 // ARM64/ARM32/PPC, respectively.
-template <typename E>
-static constexpr i64 batch_size = max_distance<E> / 10;
+static constexpr i64 batch_size = max_distance() / 10;
 
 // We assume that a single thunk group is smaller than 100 KiB.
 static constexpr i64 max_thunk_size = 102400;
 
+static_assert(max_thunk_size / E::thunk_size < INT16_MAX);
+
 // Returns true if a given relocation is of type used for function calls.
 template <typename E>
 static bool needs_thunk_rel(const ElfRel<E> &r) {
@@ -67,12 +77,15 @@ static bool needs_thunk_rel(const ElfRel<E> &r) {
     return ty == R_AARCH64_JUMP26 || ty == R_AARCH64_CALL26;
   } else if constexpr (is_arm32<E>) {
     return ty == R_ARM_JUMP24 || ty == R_ARM_THM_JUMP24 ||
-           ty == R_ARM_CALL   || ty == R_ARM_THM_CALL;
+           ty == R_ARM_CALL   || ty == R_ARM_THM_CALL   ||
+           ty == R_ARM_PLT32;
   } else if constexpr (is_ppc32<E>) {
     return ty == R_PPC_REL24  || ty == R_PPC_PLTREL24 || ty == R_PPC_LOCAL24PC;
+  } else if constexpr (is_ppc64<E>) {
+    return ty == R_PPC64_REL24 || ty == R_PPC64_REL24_NOTOC;
   } else {
-    static_assert(is_ppc64<E>);
-    return ty == R_PPC64_REL24;
+    static_assert(is_loongarch<E>);
+    return ty == R_LARCH_B26;
   }
 }
 
@@ -102,7 +115,19 @@ static bool is_reachable(Context<E> &ctx, InputSection<E> &isec,
   if constexpr (is_arm32<E>) {
     bool is_thumb = sym.get_addr(ctx) & 1;
     if ((rel.r_type == R_ARM_THM_JUMP24 && !is_thumb) ||
-        (rel.r_type == R_ARM_JUMP24 && is_thumb))
+        (rel.r_type == R_ARM_JUMP24 && is_thumb) ||
+        (rel.r_type == R_ARM_PLT32 && is_thumb))
+      return false;
+  }
+
+  // PowerPC before Power9 lacks PC-relative load/store instructions.
+  // Functions compiled for Power9 or earlier assume that r2 points to
+  // GOT+0x8000, while those for Power10 uses r2 as a scratch register.
+  // We need to a thunk to recompute r2 for interworking.
+  if constexpr (is_ppc64v2<E>) {
+    if (rel.r_type == R_PPC64_REL24 && !sym.esym().preserves_r2())
+      return false;
+    if (rel.r_type == R_PPC64_REL24_NOTOC && sym.esym().uses_toc())
       return false;
   }
 
@@ -112,10 +137,9 @@ static bool is_reachable(Context<E> &ctx, InputSection<E> &isec,
   i64 A = get_addend(isec, rel);
   i64 P = isec.get_addr() + rel.r_offset;
   i64 val = S + A - P;
-  return -max_distance<E> <= val && val < max_distance<E>;
+  return -max_distance() <= val && val < max_distance();
 }
 
-template <typename E>
 static void reset_thunk(RangeExtensionThunk<E> &thunk) {
   for (Symbol<E> *sym : thunk.symbols) {
     sym->extra.thunk_idx = -1;
@@ -125,9 +149,8 @@ static void reset_thunk(RangeExtensionThunk<E> &thunk) {
 }
 
 // Scan relocations to collect symbols that need thunks.
-template <typename E>
 static void scan_rels(Context<E> &ctx, InputSection<E> &isec,
-                      RangeExtensionThunk<E> &thunk) {
+                      RangeExtensionThunk<E> &thunk, i64 thunk_idx) {
   std::span<const ElfRel<E>> rels = isec.get_rels(ctx);
   std::vector<RangeExtensionRef> &range_extn = isec.extra.range_extn;
   range_extn.resize(rels.size());
@@ -156,7 +179,7 @@ static void scan_rels(Context<E> &ctx, InputSection<E> &isec,
 
     // Otherwise, add the symbol to the current thunk if it's not
     // added already.
-    range_extn[i].thunk_idx = thunk.thunk_idx;
+    range_extn[i].thunk_idx = thunk_idx;
     range_extn[i].sym_idx = -1;
 
     if (sym.flags.exchange(-1) == 0) {
@@ -166,9 +189,9 @@ static void scan_rels(Context<E> &ctx, InputSection<E> &isec,
   }
 }
 
-template <typename E>
-void create_range_extension_thunks(Context<E> &ctx, OutputSection<E> &osec) {
-  std::span<InputSection<E> *> m = osec.members;
+template <>
+void OutputSection<E>::create_range_extension_thunks(Context<E> &ctx) {
+  std::span<InputSection<E> *> m = members;
   if (m.empty())
     return;
 
@@ -184,17 +207,36 @@ void create_range_extension_thunks(Context<E> &ctx, OutputSection<E> &osec) {
   // We create thunks from the beginning of the section to the end.
   // We manage progress using four offsets which increase monotonically.
   // The locations they point to are always A <= B <= C <= D.
+  //
+  // Input sections between B and C are in the current batch.
+  //
+  // A is the input section with the smallest address than can reach
+  // anywhere from the current batch.
+  //
+  // D is the input section with the largest address such that the thunk
+  // is reachable from the current batch if it's inserted right before D.
+  //
+  //  ................................ <input sections> ............
+  //     A    B    C    D
+  //                    ^ We insert a thunk for the current batch just before D
+  //          <--->       The current batch, which is smaller than batch_size
+  //     <-------->       Smaller than max_distance
+  //          <-------->  Smaller than max_distance
+  //     <------------->  Reachable from the current batch
   i64 a = 0;
   i64 b = 0;
   i64 c = 0;
   i64 d = 0;
   i64 offset = 0;
 
+  // The smallest thunk index that is reachable from the current batch.
+  i64 t = 0;
+
   while (b < m.size()) {
-    // Move D foward as far as we can jump from B to anywhere in a thunk after D.
+    // Move D foward as far as we can jump from B to anywhere in a thunk at D.
     while (d < m.size() &&
            align_to(offset, 1 << m[d]->p2align) + m[d]->sh_size + max_thunk_size <
-           m[b]->offset + max_distance<E>) {
+           m[b]->offset + max_distance()) {
       offset = align_to(offset, 1 << m[d]->p2align);
       m[d]->offset = offset;
       offset += m[d]->sh_size;
@@ -206,53 +248,58 @@ void create_range_extension_thunks(Context<E> &ctx, OutputSection<E> &osec) {
     // to ensure progress.
     c = b + 1;
     while (c < m.size() &&
-           m[c]->offset + m[c]->sh_size < m[b]->offset + batch_size<E>)
+           m[c]->offset + m[c]->sh_size < m[b]->offset + batch_size)
       c++;
 
     // Move A forward so that A is reachable from C.
     i64 c_offset = (c == m.size()) ? offset : m[c]->offset;
-    while (a < osec.thunks.size() &&
-           osec.thunks[a]->offset + max_distance<E> < c_offset)
-      reset_thunk(*osec.thunks[a++]);
+    while (a < m.size() && m[a]->offset + max_distance() < c_offset)
+      a++;
 
-    // Create a thunk for input sections between B and C and place it at D.
-    osec.thunks.emplace_back(new RangeExtensionThunk<E>{osec});
+    // Erase references to out-of-range thunks.
+    while (t < thunks.size() && thunks[t]->offset < m[a]->offset)
+      reset_thunk(*thunks[t++]);
 
-    RangeExtensionThunk<E> &thunk = *osec.thunks.back();
-    thunk.thunk_idx = osec.thunks.size() - 1;
-    offset = align_to(offset, thunk.alignment);
-    thunk.offset = offset;
+    // Create a new thunk and place it at D.
+    offset = align_to(offset, RangeExtensionThunk<E>::alignment);
+    i64 thunk_idx = thunks.size();
+    RangeExtensionThunk<E> *thunk = new RangeExtensionThunk<E>(*this, offset);
+    thunks.emplace_back(thunk);
 
-    // Scan relocations between B and C to collect symbols that need thunks.
-    tbb::parallel_for_each(m.begin() + b, m.begin() + c, [&](InputSection<E> *isec) {
-      scan_rels(ctx, *isec, thunk);
+    // Scan relocations between B and C to collect symbols that need
+    // entries in the new thunk.
+    tbb::parallel_for_each(m.begin() + b, m.begin() + c,
+                           [&](InputSection<E> *isec) {
+      scan_rels(ctx, *isec, *thunk, thunk_idx);
     });
 
     // Now that we know the number of symbols in the thunk, we can compute
-    // its size.
-    assert(thunk.size() < max_thunk_size);
-    offset += thunk.size();
+    // the thunk's size.
+    assert(thunk->size() < max_thunk_size);
+    offset += thunk->size();
 
     // Sort symbols added to the thunk to make the output deterministic.
-    sort(thunk.symbols, [](Symbol<E> *a, Symbol<E> *b) {
+    sort(thunk->symbols, [](Symbol<E> *a, Symbol<E> *b) {
       return std::tuple{a->file->priority, a->sym_idx} <
              std::tuple{b->file->priority, b->sym_idx};
     });
 
     // Assign offsets within the thunk to the symbols.
-    for (i64 i = 0; Symbol<E> *sym : thunk.symbols) {
-      sym->extra.thunk_idx = thunk.thunk_idx;
-      sym->extra.thunk_sym_idx = i++;
+    for (i64 i = 0; i < thunk->symbols.size(); i++) {
+      Symbol<E> &sym = *thunk->symbols[i];
+      sym.extra.thunk_idx = thunk_idx;
+      sym.extra.thunk_sym_idx = i;
     }
 
     // Scan relocations again to fix symbol offsets in the last thunk.
-    tbb::parallel_for_each(m.begin() + b, m.begin() + c, [&](InputSection<E> *isec) {
+    tbb::parallel_for_each(m.begin() + b, m.begin() + c,
+                           [&](InputSection<E> *isec) {
       std::span<Symbol<E> *> syms = isec->file.symbols;
       std::span<const ElfRel<E>> rels = isec->get_rels(ctx);
       std::span<RangeExtensionRef> range_extn = isec->extra.range_extn;
 
       for (i64 i = 0; i < rels.size(); i++)
-        if (range_extn[i].thunk_idx == thunk.thunk_idx)
+        if (range_extn[i].thunk_idx == thunk_idx)
           range_extn[i].sym_idx = syms[rels[i].r_sym]->extra.thunk_sym_idx;
     });
 
@@ -260,15 +307,16 @@ void create_range_extension_thunks(Context<E> &ctx, OutputSection<E> &osec) {
     b = c;
   }
 
-  while (a < osec.thunks.size())
-    reset_thunk(*osec.thunks[a++]);
+  while (t < thunks.size())
+    reset_thunk(*thunks[t++]);
 
-  osec.shdr.sh_size = offset;
+  this->shdr.sh_size = offset;
+
+  for (InputSection<E> *isec : members)
+    this->shdr.sh_addralign =
+      std::max<u32>(this->shdr.sh_addralign, 1 << isec->p2align);
 }
 
-#if MOLD_ARM32 || MOLD_ARM64 || MOLD_PPC32 || MOLD_PPC64V1 || MOLD_PPC64V2
-using E = MOLD_TARGET;
-template void create_range_extension_thunks(Context<E> &, OutputSection<E> &);
-#endif
-
 } // namespace mold::elf
+
+#endif

elf/tls.cc

@@ -0,0 +1,213 @@
+// This file contains helper functions for thread-local storage (TLS).
+// TLS is probably the most obscure feature the linker has to support,
+// so I'll explain it in detail in this comment.
+//
+// TLS is a per-thread storage. Thread-local variables (TLVs) are in a TLS
+// so that each thread has its own set of thread-local variables. Taking
+// an address of a TLV returns a unique value for each thread. For example,
+// `&foo` for the following code returns different pointer values for
+// different threads.
+//
+//   thread_local int foo;
+//
+// TLV is a relatively new feature. C for example didn't provide the
+// official support for it through the keyword `thread_local` until C11.
+// TLV needs a coordination between the compiler, the linker and the
+// runtime to work correctly.
+//
+// An ELF exectuable or a shared library using TLV contains a "TLS template
+// image" in the PT_TLS segment. For each newly created thread including the
+// initial one, the runtime allocates a contiguous memory for an executable
+// and its depending shared libraries and copies template images there. That
+// per-thread memory is called the "initial TLS block". After allocating and
+// initializing the initial TLS block, the runtime sets a register to refer
+// to the initial TLS block, so that the thread-local variables are
+// accessible relative to the register.
+//
+// The register referring to the per-thread storage is called the Thread
+// Pointer (TP). TP is part of the thread's context. When the kernel
+// scheduler switches threads, TP is saved and restored automatically just
+// like other registers are.
+//
+// The TLS template image is read-only. It contains TLVs' initial values
+// for new threads, and no one writes to it at runtime.
+//
+// Now, let's think about how to access a TLV. We need to know the TLV's
+// address to access it which can be done in several different ways as
+// follows:
+//
+//  1. If we are creating an executable, we know the exact size of the TLS
+//     template image we are creating, and we know where the TP will be set
+//     to after the template is copied to the initial TLS block. Therefore,
+//     the TP-relative address of a TLV in the main executable is known at
+//     link-time. That means, computing a TLV's address can be as easy as
+//     `add %dst, %tp, <link-time constant>`.
+//
+//  2. If we are creating a shared library, we don't exactly know where
+//     its TLS template image will be copied to in terms of the
+//     TP-relative address, because we don't know how large the main
+//     executable's and other libraries' TLS template images are. Only the
+//     runtime knows the exact TP-relative address.
+//
+//     We can solve the problem with an indirection. Specifically, for
+//     each TLV whose TP-relative address is known only at process startup
+//     time, we create a GOT entry to store its TP-relative address. We
+//     then emit a dynamic relocation to let the runtime to fill the GOT
+//     entry with a TP-relative address.
+//
+//     Computing a TLV address in this scheme needs at least two machine
+//     instructions in most ISAs; the first instruction loads a value from
+//     the GOT entry, and the second one adds the loaded value to TP.
+//
+//  3. Now, think about libraries that are dynamically loaded with dlopen.
+//     The TLS block for such library may not be allocated next to the
+//     initial TLS block, so we can have two or more discontiguous TLS
+//     blocks. There's no easy formula to compute an address of a TLV in a
+//     separate TLS block.
+//
+//     The address of a TLV in a separate TLS block can be obtained by
+//     calling the libc-provided function, __tls_get_addr(). The function
+//     takes two arguments; a module ID to identify the ELF file and the
+//     TLV's offset within the ELF file's TLS template image. Accessing a
+//     TLV is sometimes compiled to a function call! The module ID and the
+//     offset are usually stored to GOT as two consecutive words.
+//
+// The last access method is the most generic, so the compiler emits such
+// code by default. But that's the most expensive one, so the linker
+// rewrites instructions if possible so that 3) is relaxed to 2) or even
+// to 1).
+//
+// 1) is called the Local Exec access model. 2) is Initial Exec, and 3) is
+// General Dynamic.
+//
+// There's another little trick that the compiler can use if it knows two
+// TLVs are in the same ELF file (usually in the same file as the code is).
+// In this case, we can call __tls_get_addr() only once with a module ID and
+// the offset 0 to obtain the base address of the ELF file's TLS block. The
+// base address obtained this way is sometimes called Dynamic Thread Pointer
+// or DTP. We can then compute TLVs' addresses by adding their DTP-relative
+// addresses to DTP. This access model is called the Local Dynamic.
+//
+// === TLS Descriptor access model ===
+//
+// As described above, there are arguably too many different TLS access
+// models from the most generic one you can use in any ELF file to the most
+// efficient one you can use only when building a main executable. Compiling
+// source code with an appropriate TLS access model is bothersome. To solve
+// the problem, a new TLS access model was proposed. That is called the TLS
+// Descriptor (TLSDESC) model.
+//
+// For a TLV compiled with TLSDESC, we allocate two consecutive GOT slots
+// and create a TLSDESC dynamic relocation for them. The dynamic linker
+// sets a function pointer to the first GOT slot and its argument to the
+// second slot.
+//
+// To access the TLV, we call the function pointer with the argument we
+// read from the second GOT slot. The function returns the TLV's
+// TP-relative address.
+//
+// The runtime chooses the best access method depending on the situation
+// and sets a pointer to the most efficient code to the first GOT slot.
+// For example, if a TLV's TP-relative address is known at process startup
+// time, the runtime sets that address to the second GOT slot and set a
+// function that just returns its argument to the first GOT slot.
+//
+// With TLSDECS, the compiler can always emit the same code for TLVs
+// without sacrificing runtime performance.
+//
+// TLSDESC is better than the traditional, non-TLSDESC TLS access models.
+// It's the default on ARM64, but on other targets, TLSDESC is
+// unfortunately either optional or even not supported at all. So we still
+// need to support both the traditional TLS models and the TLSDESC model.
+
+#include "mold.h"
+
+namespace mold::elf {
+
+template <typename E>
+static ElfPhdr<E> *get_tls_segment(Context<E> &ctx) {
+  if (ctx.phdr)
+    for (ElfPhdr<E> &phdr : ctx.phdr->phdrs)
+      if (phdr.p_type == PT_TLS)
+        return &phdr;
+  return nullptr;
+}
+
+template <typename E>
+u64 get_tls_begin(Context<E> &ctx) {
+  if (ElfPhdr<E> *phdr = get_tls_segment(ctx))
+    return phdr->p_vaddr;
+  return 0;
+}
+
+// Returns the TP address which can be used for efficient TLV accesses in
+// the main executable. TP at runtime refers to a per-process TLS block
+// whose address is not known at link-time. So the address returned from
+// this function is the TP if the TLS template image were a TLS block.
+template <typename E>
+u64 get_tp_addr(Context<E> &ctx) {
+  ElfPhdr<E> *phdr = get_tls_segment(ctx);
+  if (!phdr)
+    return 0;
+
+  if constexpr (is_x86<E> || is_sparc<E> || is_s390x<E>) {
+    // On x86, SPARC and s390x, TP (%gs on i386, %fs on x86-64, %g7 on SPARC
+    // and %a0/%a1 on s390x) refers to past the end of the TLS block for
+    // historical reasons. TLVs are accessed with negative offsets from TP.
+    return align_to(phdr->p_vaddr + phdr->p_memsz, phdr->p_align);
+  } else if constexpr (is_arm<E> || is_sh4<E> || is_alpha<E>) {
+    // On ARM, SH4 and Alpha, the runtime appends two words at the beginning
+    // of TLV template image when copying TLVs to the TLS block, so we need
+    // to offset it.
+    return align_down(phdr->p_vaddr - sizeof(Word<E>) * 2, phdr->p_align);
+  } else if constexpr (is_ppc<E> || is_m68k<E>) {
+    // On PowerPC and m68k, TP is 0x7000 (28 KiB) past the beginning
+    // of the TLV block to maximize the addressable range of load/store
+    // instructions with 16-bits signed immediates. It's not exactly 0x8000
+    // (32 KiB) off because there's a small implementation-defined piece of
+    // data before the initial TLV block, and the runtime wants to access
+    // them efficiently too.
+    return phdr->p_vaddr + 0x7000;
+  } else {
+    // RISC-V and LoongArch just uses the beginning of the main executable's
+    // TLV block as TP. Their load/store instructions usually take 12-bits
+    // signed immediates, so the beginning of the TLS block ± 2 KiB is
+    // accessible with a single load/store instruction.
+    static_assert(is_riscv<E> || is_loongarch<E>);
+    return phdr->p_vaddr;
+  }
+}
+
+// Returns the address __tls_get_addr() would return if it's called
+// with offset 0.
+template <typename E>
+u64 get_dtp_addr(Context<E> &ctx) {
+  ElfPhdr<E> *phdr = get_tls_segment(ctx);
+  if (!phdr)
+    return 0;
+
+  if constexpr (is_ppc<E> || is_m68k<E>) {
+    // On PowerPC and m68k, R_DTPOFF is resolved to the address 0x8000
+    // (32 KiB) past the start of the TLS block. The bias maximizes the
+    // accessible range of load/store instructions with 16-bits signed
+    // immediates. That is, if the offset were right at the beginning of the
+    // start of the TLS block, the half of addressible space (negative
+    // immediates) would have been wasted.
+    return phdr->p_vaddr + 0x8000;
+  } else if constexpr (is_riscv<E>) {
+    // On RISC-V, the bias is 0x800 as the load/store instructions in the
+    // ISA usually have a 12-bit immediate.
+    return phdr->p_vaddr + 0x800;
+  } else {
+    // On other targets, DTP simply refers to the beginning of the TLS block.
+    return phdr->p_vaddr;
+  }
+}
+
+using E = MOLD_TARGET;
+
+template u64 get_tls_begin<E>(Context<E> &);
+template u64 get_tp_addr<E>(Context<E> &);
+template u64 get_dtp_addr<E>(Context<E> &);
+
+} // namespace mold::elf

install-build-deps.sh

@@ -1,6 +1,6 @@
-#!/bin/bash
+#!/bin/sh
 set -e
-source /etc/os-release
+. /etc/os-release
 
 set -x
 
@@ -12,33 +12,42 @@ set -x
 case "$ID-$VERSION_ID" in
 ubuntu-20.* | pop-20.*)
   apt-get update
-  apt-get install -y cmake libssl-dev zlib1g-dev gcc g++ g++-10
+  apt-get install -y cmake gcc g++ g++-10
   apt-get install -y file
   ;;
 ubuntu-* | pop-* | linuxmint-* | debian-* | raspbian-*)
   apt-get update
-  apt-get install -y cmake libssl-dev zlib1g-dev gcc g++
+  apt-get install -y cmake gcc g++
   apt-get install -y file
   ;;
 fedora-*)
-  dnf install -y gcc-g++ cmake openssl-devel zlib-devel
+  dnf install -y gcc-g++ cmake
   dnf install -y glibc-static file libstdc++-static diffutils util-linux
   ;;
 opensuse-leap-*)
-  zypper install -y make cmake zlib-devel libopenssl-devel gcc-c++ gcc11-c++
+  zypper install -y make cmake gcc-c++ gcc11-c++
   zypper install -y glibc-devel-static tar diffutils util-linux
   ;;
 opensuse-tumbleweed-*)
-  zypper install -y make cmake zlib-devel libopenssl-devel gcc-c++
+  zypper install -y make cmake gcc-c++
   zypper install -y glibc-devel-static tar diffutils util-linux
   ;;
 gentoo-*)
   emerge-webrsync
-  emerge dev-util/cmake sys-libs/zlib
+  emerge dev-util/cmake
   ;;
-arch-*)
+arch-* | archarm-*)
   pacman -Sy
-  pacman -S --needed --noconfirm base-devel zlib openssl cmake util-linux
+  pacman -S --needed --noconfirm base-devel cmake util-linux
+  ;;
+void-*)
+  xbps-install -Sy xbps
+  xbps-install -Sy bash make cmake gcc
+  xbps-install -Sy tar diffutils util-linux
+  ;;
+alpine-*)
+  apk update
+  apk add bash make linux-headers cmake gcc g++
   ;;
 *)
   echo "Error: don't know anything about build dependencies on $ID-$VERSION_ID"

install-cross-tools.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+set -e
+. /etc/os-release
+
+set -x
+
+# This script install packages for -DMOLD_ENABLE_QEMU_TESTS=1
+# to enable cross-target tests.
+#
+# Feel free to send me a PR if your OS is not on this list.
+
+case "$ID-$VERSION_ID" in
+ubuntu-* | pop-* | linuxmint-* | debian-* | raspbian-*)
+  apt-get install -y qemu-user {gcc,g++}-{i686,aarch64,riscv64,powerpc,powerpc64,powerpc64le,s390x,sparc64,m68k,sh4,alpha}-linux-gnu {gcc,g++}-arm-linux-gnueabihf
+  ;;
+*)
+  echo "Error: don't know anything about build dependencies on $ID-$VERSION_ID"
+  exit 1
+esac

sha.h

@@ -1,81 +0,0 @@
-#pragma once
-
-#include <cstdint>
-
-typedef uint8_t u8;
-static constexpr int64_t SHA256_SIZE = 32;
-
-#ifdef _WIN32
-// On Windows, we use Microsoft CNG.
-
-#include <Windows.h>
-#include <bcrypt.h>
-#include <ntstatus.h>
-
-inline static BCRYPT_ALG_HANDLE get_sha256_handle() {
-  static std::once_flag once;
-  static BCRYPT_ALG_HANDLE alg;
-
-  std::call_once(once, [&] {
-    BCryptOpenAlgorithmProvider(&alg, BCRYPT_SHA256_ALGORITHM, nullptr, 0);
-  });
-  return alg;
-}
-
-inline void sha256_hash(u8 *in, size_t len, u8 *out) {
-  BCryptHash(get_sha256_handle(), nullptr, 0, in, len, out, SHA256_SIZE);
-}
-
-class SHA256Hash {
-public:
-  SHA256Hash() {
-    BCryptCreateHash(get_sha256_handle(), &handle, nullptr, 0, nullptr, 0, 0);
-  }
-
-  void update(u8 *data, size_t len) {
-    BCryptHashData(handle, data, len, 0);
-  }
-
-  void finish(u8 *out) {
-    BCryptFinishHash(handle, out, SHA256_SIZE, 0);
-  }
-
-private:
-  BCRYPT_HASH_HANDLE handle;
-};
-
-#else
-// On Unix, we use OpenSSL or the Apple's OpenSSL-compatible API.
-
-#ifdef __APPLE__
-#  define COMMON_DIGEST_FOR_OPENSSL
-#  include <CommonCrypto/CommonDigest.h>
-#  define SHA256(data, len, md) CC_SHA256(data, len, md)
-#else
-#  define OPENSSL_SUPPRESS_DEPRECATED 1
-#  include <openssl/sha.h>
-#endif
-
-inline void sha256_hash(u8 *in, size_t len, u8 *out) {
-  SHA256(in, len, out);
-}
-
-class SHA256Hash {
-public:
-  SHA256Hash() {
-    SHA256_Init(&ctx);
-  }
-
-  void update(u8 *data, size_t len) {
-    SHA256_Update(&ctx, data, len);
-  }
-
-  void finish(u8 *out) {
-    SHA256_Final(out, &ctx);
-  }
-
-private:
-  SHA256_CTX ctx;
-};
-
-#endif

test/elf/CMakeLists.txt

@@ -1,10 +1,16 @@
 option(MOLD_ENABLE_QEMU_TESTS "Enable tests on non-native targets" OFF)
-option(MOLD_ENABLE_QEMU_TESTS_RV32 "Enable tests for RV32" OFF)
 
 function(add_target TRIPLE)
   set(HOST ${CMAKE_HOST_SYSTEM_PROCESSOR})
-  if(${HOST} MATCHES "arm.*")
+
+  if(${HOST} MATCHES "amd64")
+    set(HOST x86_64)
+  elseif(${HOST} MATCHES "arm.*")
     set(HOST arm)
+  elseif(${HOST} STREQUAL "ppc64")
+    set(HOST powerpc64)
+  elseif(${HOST} STREQUAL "ppc64le")
+    set(HOST powerpc64le)
   endif()
 
   if(${TRIPLE} MATCHES "${HOST}-.*")
@@ -54,10 +60,25 @@ add_target(powerpc64-linux-gnu)
 add_target(powerpc64le-linux-gnu)
 add_target(sparc64-linux-gnu)
 add_target(s390x-linux-gnu)
-add_target(m68k-linux-gnu)
 add_target(sh4-linux-gnu)
 add_target(alpha-linux-gnu)
 
+option(MOLD_ENABLE_QEMU_TESTS_RV32 "Enable tests for RV32" OFF)
 if(MOLD_ENABLE_QEMU_TESTS_RV32)
   add_target(riscv32-linux-gnu)
 endif()
+
+option(MOLD_ENABLE_QEMU_TESTS_POWER10 "Enable tests for Power10" OFF)
+if(MOLD_ENABLE_QEMU_TESTS_POWER10)
+  add_target(powerpc64le_power10-linux-gnu)
+endif()
+
+option(MOLD_ENABLE_QEMU_TESTS_M68K "Enable tests for m68k" OFF)
+if(MOLD_ENABLE_QEMU_TESTS_M68K)
+  add_target(m68k-linux-gnu)
+endif()
+
+option(MOLD_ENABLE_QEMU_TESTS_LOONGARCH "Enable tests for LoongArch" OFF)
+if(MOLD_ENABLE_QEMU_TESTS_LOONGARCH)
+  add_target(loongarch64-linux-gnu)
+endif()

test/elf/abs-error.sh

@@ -6,6 +6,7 @@
 [ $MACHINE = ppc64le ] && skip
 [ $MACHINE = s390x ] && skip
 [ $MACHINE = alpha ] && skip
+[[ $MACHINE = loongarch* ]] && skip
 
 cat <<EOF | $CC -fPIC -c -o $t/a.o -xassembler -
 .globl foo

test/elf/arm_range-extension-thunk-disassembly.sh

@@ -22,6 +22,5 @@ $CC -B. -o $t/exe $t/a.o \
 
 $OBJDUMP -dr $t/exe | grep -F -A7 '<fn1$thunk>:' > $t/log
 
-grep -Eq 'mov\s+ip, pc' $t/log
-grep -Eq 'bx\s+ip' $t/log
-grep -Eq 'add\s+ip, ip, pc' $t/log
+grep -Eq 'bx\s+pc' $t/log
+grep -Eq 'add\s+pc, ip, pc' $t/log

test/elf/as-needed-dso.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -o $t/libfoo.so -shared -fPIC -Wl,-soname,libfoo.so -xc -
+int fn1() { return 42; }
+EOF
+
+cat <<EOF | $CC -o $t/libbar.so -shared -fPIC -Wl,-soname,libbar.so -xc -
+int fn1();
+int fn2() { return fn1(); }
+EOF
+
+cat <<EOF | $CC -o $t/a.o -c -xc -
+int fn2();
+int main() { fn2(); }
+EOF
+
+$CC -B. -o $t/exe $t/a.o -L$t -Wl,--as-needed -lbar -lfoo
+readelf -W --dynamic $t/exe > $t/log2
+grep -q libbar $t/log2
+! grep -q libfoo $t/log2 || false

test/elf/as-needed-dso2.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -c -fPIC -o $t/a.o -xc -
+int foo() {
+  return 0;
+}
+EOF
+
+cat <<EOF | $CC -c -fPIC -o $t/b.o -xc -
+__attribute__((weak)) int foo();
+
+int bar() {
+  if (foo) return foo();
+  return 0;
+}
+EOF
+
+cat <<EOF | $CC -xc -c -o $t/c.o -
+int bar();
+
+int main() {
+  return bar();
+}
+EOF
+
+$CC -B. -shared -o $t/libfoo.so $t/a.o
+$CC -B. -shared -o $t/libbar.so $t/b.o
+$CC -B. -o $t/exe $t/c.o -L$t -Wl,--as-needed -lfoo -lbar
+
+readelf --dynamic $t/exe > $t/log
+! grep libfoo.so $t/log || false
+grep -q libbar.so $t/log

test/elf/as-needed-weak.sh

@@ -5,26 +5,27 @@ cat <<EOF | $CC -fPIC -o $t/a.o -c -xc -
 __attribute__((weak)) int fn1();
 
 int main() {
-  fn1();
+  if (fn1)
+    fn1();
 }
 EOF
 
-cat <<EOF | $CC -o $t/b.so -shared -fPIC -Wl,-soname,libfoo.so -xc -
+cat <<EOF | $CC -o $t/libfoo.so -shared -fPIC -Wl,-soname,libfoo.so -xc -
 int fn1() { return 42; }
 EOF
 
-cat <<EOF | $CC -o $t/c.so -shared -fPIC -Wl,-soname,libbar.so -xc -
+cat <<EOF | $CC -o $t/libbar.so -shared -fPIC -Wl,-soname,libbar.so -xc -
 int fn2() { return 42; }
 EOF
 
-$CC -B. -o $t/exe $t/a.o -Wl,-no-as-needed $t/b.so $t/c.so
+$CC -o $t/exe1 $t/a.o -Wl,-no-as-needed -L$t -lbar -lfoo
 
-readelf --dynamic $t/exe > $t/readelf
-grep -Fq 'Shared library: [libfoo.so]' $t/readelf
-grep -Fq 'Shared library: [libbar.so]' $t/readelf
+readelf --dynamic $t/exe1 > $t/log1
+grep -Fq 'Shared library: [libfoo.so]' $t/log1
+grep -Fq 'Shared library: [libbar.so]' $t/log1
 
-$CC -B. -o $t/exe $t/a.o -Wl,-as-needed $t/b.so $t/c.so
+$CC -o $t/exe2 $t/a.o -Wl,-as-needed -L$t -lbar -lfoo
 
-readelf --dynamic $t/exe > $t/readelf
-! grep -Fq 'Shared library: [libfoo.so]' $t/readelf || false
-! grep -Fq 'Shared library: [libbar.so]' $t/readelf || false
+readelf --dynamic $t/exe2 > $t/log2
+! grep -Fq 'Shared library: [libfoo.so]' $t/log2 || false
+! grep -Fq 'Shared library: [libbar.so]' $t/log2 || false

test/elf/as-needed2.sh

@@ -1,33 +0,0 @@
-#!/bin/bash
-. $(dirname $0)/common.inc
-
-cat <<EOF | $CC -shared -fPIC -o $t/a.so -xc -
-int foo() { return 3; }
-EOF
-
-cat <<EOF | $CC -shared -fPIC -o $t/b.so -xc -
-int bar() { return 3; }
-EOF
-
-cat <<EOF | $CC -fPIC -c -o $t/c.o -xc -
-int foo();
-int baz() { return foo(); }
-EOF
-
-$CC -shared -o $t/c.so $t/c.o $t/a.so
-
-cat <<EOF | $CC -c -o $t/d.o -xc -
-#include <stdio.h>
-int baz();
-int main() {
-  printf("%d\n", baz());
-}
-EOF
-
-$CC -B. -o $t/exe $t/d.o -Wl,--as-needed $t/c.so $t/b.so $t/a.so
-$QEMU $t/exe | grep -q '^3$'
-
-readelf --dynamic $t/exe > $t/log
-! grep -q /a.so $t/log || false
-grep -q /c.so $t/log || false
-! grep -q /b.so $t/log || false

test/elf/common.inc

@@ -16,23 +16,10 @@ canonical_name() {
 
 HOST=$(canonical_name $(uname -m))
 
-if [ -n "$TRIPLE" ]; then
-  MACHINE=$(canonical_name $(echo $TRIPLE | sed 's/-.*//'))
-else
-  MACHINE=$HOST
-fi
-
 # Set tool names
-if [ -n "$TRIPLE" ]; then
-  CC="${TEST_CC:-$TRIPLE-gcc}"
-  CXX="${TEST_CXX:-$TRIPLE-g++}"
-  GCC="${TEST_GCC:-$TRIPLE-gcc}"
-  GXX="${TEST_GXX:-$TRIPLE-g++}"
-  OBJDUMP="$TRIPLE-objdump"
-  OBJCOPY="$TRIPLE-objcopy"
-  STRIP="$TRIPLE-strip"
-  QEMU="qemu-$MACHINE -L /usr/$TRIPLE"
-else
+if [ "$TRIPLE" = "" ]; then
+  MACHINE=$HOST
+  TESTDIR=out/test/elf/$HOST
   CC="${TEST_CC:-cc}"
   CXX="${TEST_CXX:-c++}"
   GCC="${TEST_GCC:-gcc}"
@@ -41,6 +28,35 @@ else
   OBJCOPY=objcopy
   STRIP=strip
   QEMU=
+elif [ "$TRIPLE" = powerpc64le_power10-linux-gnu ]; then
+  TRIPLE=powerpc64le-linux-gnu
+  MACHINE=ppc64le
+  TESTDIR=out/test/elf/ppc64le-power10
+  CC="${TEST_CC:-$TRIPLE-gcc} -mcpu=power10"
+  CXX="${TEST_CXX:-$TRIPLE-g++} -mcpu=power10"
+  GCC="${TEST_GCC:-$TRIPLE-gcc} -mcpu=power10"
+  GXX="${TEST_GXX:-$TRIPLE-g++} -mcpu=power10"
+  OBJDUMP="$TRIPLE-objdump"
+  OBJCOPY="$TRIPLE-objcopy"
+  STRIP="$TRIPLE-strip"
+  QEMU="qemu-ppc64le -L /usr/$TRIPLE -cpu power10"
+else
+  MACHINE=$(canonical_name $(echo $TRIPLE | sed 's/-.*//'))
+  TESTDIR=out/test/elf/$MACHINE
+  CC="${TEST_CC:-$TRIPLE-gcc}"
+  CXX="${TEST_CXX:-$TRIPLE-g++}"
+  GCC="${TEST_GCC:-$TRIPLE-gcc}"
+  GXX="${TEST_GXX:-$TRIPLE-g++}"
+  OBJDUMP="$TRIPLE-objdump"
+  OBJCOPY="$TRIPLE-objcopy"
+  STRIP="$TRIPLE-strip"
+  QEMU="qemu-$MACHINE -L /usr/$TRIPLE"
+fi
+
+if [ $MACHINE = x86_64 -o $MACHINE = i386 -o $MACHINE = arm ]; then
+  tlsdesc_opt=-mtls-dialect=gnu2
+elif [ $MACHINE = aarch64 ]; then
+  tlsdesc_opt=-mtls-dialect=desc
 fi
 
 # Common functions
@@ -53,6 +69,10 @@ supports_ifunc() {
     $CC -c -o /dev/null -xc - >& /dev/null
 }
 
+supports_tlsdesc() {
+  [ -n "$tlsdesc_opt" ]
+}
+
 skip() {
   echo skipped
   trap - EXIT
@@ -77,5 +97,5 @@ trap on_exit EXIT
 # Print out the startup message
 testname=$(basename "$0" .sh)
 echo -n "Testing $testname ... "
-t=out/test/elf/$MACHINE/$testname
+t=$TESTDIR/$testname
 mkdir -p $t

test/elf/copyrel-alignment.sh

@@ -4,6 +4,7 @@
 [ $MACHINE = ppc64 ] && skip
 [ $MACHINE = ppc64le ] && skip
 [ $MACHINE = alpha ] && skip
+[[ $MACHINE = loongarch* ]] && skip
 
 cat <<EOF | $CC -fPIC -shared -o $t/a.so -xc -
 __attribute__((aligned(32))) int foo = 5;

test/elf/copyrel-protected.sh

@@ -4,6 +4,7 @@
 [ $MACHINE = ppc64 ] && skip
 [ $MACHINE = ppc64le ] && skip
 [ $MACHINE = alpha ] && skip
+[[ $MACHINE = loongarch* ]] && skip
 
 cat <<EOF | $CC -o $t/a.o -c -xc -fno-PIE -
 extern int foo;

test/elf/copyrel-relro2.sh

@@ -0,0 +1,41 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+
+cat <<EOF | $CC -o $t/a.o -c -xc -fno-PIE -
+#include <setjmp.h>
+#include <signal.h>
+#include <stdio.h>
+
+extern char readonly[100];
+extern char readwrite[100];
+
+static int segv = 0;
+static jmp_buf buf;
+
+void handler(int sig) {
+  segv = 1;
+  longjmp(buf, 1);
+}
+
+int main() {
+  signal(SIGSEGV, handler);
+
+  readwrite[0] = 5;
+  int x = segv;
+
+  if (setjmp(buf) == 0)
+    *(char *)readonly = 5;
+  int y = segv;
+
+  printf("sigsegv %d %d\n", x, y);
+}
+EOF
+
+cat <<EOF | $CC -fPIC -shared -o $t/b.so -xc -
+__attribute__((section (".data.rel.ro"))) char readonly[100] = "abc";
+char readwrite[100] = "abc";
+EOF
+
+$CC -B. $t/a.o $t/b.so -o $t/exe -no-pie
+$QEMU $t/exe | grep -q '^sigsegv 0 1$'

test/elf/dt-init.sh

@@ -2,6 +2,11 @@
 . $(dirname $0)/common.inc
 
 [ $MACHINE = riscv64 -o $MACHINE = riscv32 ] && skip
+[[ $MACHINE = loongarch* ]] && skip
+
+# musl libc does not support init/fini on ARM
+# https://github.com/rui314/mold/issues/951
+[ $MACHINE = arm -o $MACHINE = aarch64 ] && ldd --help 2>&1 | grep -q musl && skip
 
 cat <<EOF | $CC -c -fPIC -o $t/a.o -xc -
 void keep();

test/elf/emit-relocs-dead-sections.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+[ $MACHINE = sh4 ] && skip
+
+cat <<EOF | $CXX -o $t/a.o -c -fPIC -xc++ -g -ffunction-sections -
+#include <iostream>
+struct Foo {
+  Foo() { std::cout << "Hello world\n"; }
+};
+
+Foo x;
+EOF
+
+cat <<EOF | $CXX -o $t/b.o -c -fPIC -xc++ -g -ffunction-sections -
+#include <iostream>
+struct Foo {
+  Foo() { std::cout << "Hello world\n"; }
+};
+
+Foo y;
+int main() {}
+EOF
+
+$CXX -B. -o $t/exe $t/a.o $t/b.o -Wl,-emit-relocs
+$QEMU $t/exe | grep -q 'Hello world'

test/elf/exception.sh

@@ -3,8 +3,11 @@
 
 [ $MACHINE = sh4 ] && skip
 
-static=1
-test_cflags -static || static=0
+static=
+test_cflags -static && static=-static
+
+# I don't know why, but we need -pthread on m68k
+static="$static -pthread"
 
 cat <<EOF > $t/a.cc
 int main() {
@@ -20,15 +23,11 @@ EOF
 $CXX -c -o $t/b.o $t/a.cc -fPIC
 $CXX -c -o $t/c.o $t/a.cc -fno-PIC
 
-if [ $static = 1 ]; then
-  $CXX -B. -o $t/exe1 $t/b.o -static
-  $QEMU $t/exe1
-fi
+$CXX -B. -o $t/exe1 $t/b.o $static
+$QEMU $t/exe1
 
-if [ $static = 1 ]; then
-  $CXX -B. -o $t/exe2 $t/c.o -static
-  $QEMU $t/exe2
-fi
+$CXX -B. -o $t/exe2 $t/c.o -no-pie $static
+$QEMU $t/exe2
 
 $CXX -B. -o $t/exe3 $t/b.o -pie
 $QEMU $t/exe3
@@ -39,18 +38,14 @@ $QEMU $t/exe4
 $CXX -B. -o $t/exe5 $t/b.o -pie -Wl,--gc-sections
 $QEMU $t/exe5
 
-if [ $static = 1 ]; then
-  $CXX -B. -o $t/exe6 $t/c.o -static -Wl,--gc-sections
-  $QEMU $t/exe6
-fi
+$CXX -B. -o $t/exe6 $t/c.o -no-pie $static -Wl,--gc-sections
+$QEMU $t/exe6
 
 if [ $MACHINE = x86_64 ]; then
   $CXX -c -o $t/d.o $t/a.cc -mcmodel=large -fPIC
 
-  if [ $static = 1 ]; then
-    $CXX -B. -o $t/exe7 $t/d.o -static
-    $QEMU $t/exe7
-  fi
+  $CXX -B. -o $t/exe7 $t/d.o $static
+  $QEMU $t/exe7
 
   $CXX -B. -o $t/exe8 $t/d.o -pie
   $QEMU $t/exe8
@@ -59,10 +54,8 @@ fi
 if [ $MACHINE = x86_64 -o $MACHINE = aarch64 ]; then
   $CXX -c -o $t/e.o $t/a.cc -mcmodel=large -fno-PIC
 
-  if [ $static = 1 ]; then
-    $CXX -B. -o $t/exe9 $t/e.o -static
-    $QEMU $t/exe9
-  fi
+  $CXX -B. -o $t/exe9 $t/e.o -no-pie $static
+  $QEMU $t/exe9
 
   $CXX -B. -o $t/exe10 $t/e.o -no-pie
   $QEMU $t/exe10

test/elf/execute-only.sh

@@ -4,6 +4,9 @@
 # .text is writable on sparc, which is not compatible with --execute-only
 [ $MACHINE = sparc64 ] && skip
 
+# GCC emits data to .text for PPC64, so PPC64 is not compatible with -execute-only
+[ $MACHINE = ppc64 ] && skip
+
 cat <<EOF | $CC -o $t/a.o -c -xc -
 #include <stdio.h>
 int main() {

test/elf/gdb-index-split-dwarf.sh

@@ -0,0 +1,80 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+[ $MACHINE = $HOST ] || skip
+[ $MACHINE = riscv64 -o $MACHINE = riscv32 -o $MACHINE = sparc64 ] && skip
+
+command -v gdb >& /dev/null || skip
+
+test_cflags -gdwarf-5 -g || skip
+
+cat <<EOF > $t/a.c
+void fn3();
+
+static void fn2() {
+  fn3();
+}
+
+void fn1() {
+  fn2();
+}
+EOF
+
+cat <<EOF > $t/b.c
+void fn5();
+
+static void fn4() {
+  fn5();
+}
+
+void fn3() {
+  fn4();
+}
+EOF
+
+cat <<EOF > $t/c.c
+void fn7();
+
+static void fn6() {
+  fn7();
+}
+
+void fn5() {
+  fn6();
+}
+EOF
+
+cat <<EOF > $t/d.c
+#include <stdio.h>
+void trap() {}
+
+static void fn8() {
+  printf("Hello world\n");
+  trap();
+}
+
+void fn7() {
+  fn8();
+}
+EOF
+
+$CC -c -o $t/a.o $t/a.c -fPIC -g -ggnu-pubnames -gdwarf-5 -gsplit-dwarf
+$CC -c -o $t/b.o $t/b.c -fPIC -g -ggnu-pubnames -gdwarf-4 -gsplit-dwarf
+$CC -c -o $t/c.o $t/c.c -fPIC -g -ggnu-pubnames -gdwarf-5
+$CC -c -o $t/d.o $t/d.c -fPIC -g -ggnu-pubnames -gdwarf-5 -gsplit-dwarf
+
+$CC -B. -shared -o $t/e.so $t/a.o $t/b.o $t/c.o $t/d.o -Wl,--gdb-index
+readelf -WS $t/e.so 2> /dev/null | grep -Fq .gdb_index
+
+cat <<EOF | $CC -c -o $t/f.o -fPIC -g -ggnu-pubnames -gdwarf-5 -gsplit-dwarf -xc - -gz
+void fn1();
+
+int main() {
+  fn1();
+}
+EOF
+
+$CC -B. -o $t/exe $t/e.so $t/f.o -Wl,--gdb-index
+readelf -WS $t/exe 2> /dev/null | grep -Fq .gdb_index
+
+$QEMU $t/exe | grep -q 'Hello world'

test/elf/hash-style.sh

@@ -1,6 +1,8 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
+# MIPS ABIs are not compatible with .gnu.hash
+
 cat <<EOF | $CC -o $t/a.o -c -xc -
 void foo() {}
 EOF
@@ -9,3 +11,9 @@ EOF
 
 readelf -WS $t/b.so | grep -Fq ' .hash'
 readelf -WS $t/b.so | grep -Fq ' .gnu.hash'
+
+./mold -shared -o $t/c.so $t/a.o --hash-style=both --hash-style=none
+
+readelf -WS $t/c.so > $t/log
+! grep -Fq ' .hash' $t/log || false
+! grep -Fq ' .gnu.hash' $t/log || false

test/elf/ifunc-dlopen.sh

@@ -0,0 +1,42 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+supports_ifunc || skip
+
+cat <<EOF | $CC -fno-PIE -o $t/a.o -c -xc -
+#include <dlfcn.h>
+#include <stdio.h>
+
+typedef void Func();
+void foo(void);
+
+int main() {
+  void *handle = dlopen(NULL, RTLD_NOW);
+  Func *p = dlsym(handle, "foo");
+
+  foo();
+  p();
+  printf("%p %p\n", foo, p);
+}
+EOF
+
+cat <<EOF | $CC -fPIC -o $t/b.o -c -xc -
+#include <stdio.h>
+
+__attribute__((ifunc("resolve_foo")))
+void foo(void);
+
+static void real_foo(void) {
+  printf("foo ");
+}
+
+typedef void Func();
+
+static Func *resolve_foo(void) {
+  return real_foo;
+}
+EOF
+
+$CC -B. -o $t/c.so $t/b.o -shared
+$CC -B. -o $t/exe $t/a.o $t/c.so -no-pie -ldl
+$QEMU $t/exe | grep -q 'foo foo'

test/elf/ifunc-dso.sh

@@ -11,7 +11,7 @@ int main() {
 }
 EOF
 
-cat <<EOF | $CC -fPIC -shared -o $t/b.so -xc -
+cat <<EOF | $CC -fPIC -o $t/b.o -c -xc -
 #include <stdio.h>
 
 __attribute__((ifunc("resolve_foobar")))
@@ -28,5 +28,8 @@ static Func *resolve_foobar(void) {
 }
 EOF
 
-$CC -B. -o $t/exe $t/a.o $t/b.so
+$CC -B. -o $t/c.so $t/b.o -shared
+readelf -W --dyn-syms $t/c.so | grep -Eq '(IFUNC|<OS specific>: 10).*foobar'
+
+$CC -B. -o $t/exe $t/a.o $t/c.so
 $QEMU $t/exe | grep -q 'Hello world'

test/elf/ifunc-funcptr.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+supports_ifunc || skip
+
+cat <<EOF | $CC -c -fPIC -o $t/a.o -c -xc -
+typedef int Fn();
+
+int foo() __attribute__((ifunc("resolve_foo")));
+
+int real_foo() { return 3; }
+
+Fn *resolve_foo(void) {
+  return real_foo;
+}
+EOF
+
+cat <<EOF | $CC -c -fPIC -o $t/b.o -xc -
+typedef int Fn();
+int foo();
+Fn *get_foo() { return foo; }
+EOF
+
+cat <<EOF | $CC -c -fPIC -o $t/c.o -xc -
+#include <stdio.h>
+
+typedef int Fn();
+Fn *get_foo();
+
+int main() {
+  Fn *f = get_foo();
+  printf("%d\n", f());
+}
+EOF
+
+$CC -B. -o $t/exe $t/a.o $t/b.o $t/c.o
+$QEMU $t/exe | grep -q '^3$'

test/elf/lto-nostdlib.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -flto -c -o $t/a.o -xc -
+void _start() {}
+EOF
+
+$CC -B. -o $t/exe -flto $t/a.o -nostdlib
+readelf -Ws $t/exe | grep -Eq ' _start$'

test/elf/main-in-dso.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -o $t/a.o -c -xc - -fPIC
+#include <stdio.h>
+int main() {
+  printf("Hello world\n");
+}
+EOF
+
+$CC -B. -shared -o $t/b.so $t/a.o
+
+$CC -o $t/c.o -c -xc /dev/null -fPIC
+$CC -B. -o $t/exe1 $t/c.o $t/b.so -pie
+$QEMU $t/exe1 | grep -q 'Hello world'
+
+$CC -o $t/c.o -c -xc /dev/null -fno-PIC
+$CC -B. -o $t/exe2 $t/c.o $t/b.so -no-pie
+$QEMU $t/exe2 | grep -q 'Hello world'

test/elf/many-sections.sh

@@ -1,8 +1,8 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-seq 1 100000 | sed 's/.*/int x\0 __attribute__((section(".data.\0")));/g' | \
-  $CC -c -xc -o $t/a.o -
+seq 1 100000 | sed 's/.*/.section .data.\0,"aw"\n.word 0\n/g' | \
+  $CC -c -xassembler -o $t/a.o -
 
 cat <<'EOF' | $CC -c -xc -o $t/b.o -
 #include <stdio.h>

test/elf/many-sections2.sh

@@ -4,9 +4,16 @@
 # OneTBB isn't tsan-clean
 nm mold | grep -q '__tsan_init' && skip
 
-seq 1 100000 | sed 's/.*/int x\0 __attribute__((section(".data.\0")));/g' | \
-  $CC -c -xc -o $t/a.o -
+echo 'foo = 0x1000' > $t/a.s
+seq 1 100000 | sed 's/.*/.section .data.\0,"aw"\n.globl x\0\nx\0: .word 0\n/g' >> $t/a.s
+$CC -c -xassembler -o $t/a.o $t/a.s
 
 ./mold --relocatable -o $t/b.o $t/a.o
-readelf -WS $t/b.o | grep -Fq .data.100000
-readelf -Ws $t/b.o | grep -Fq 'OBJECT  GLOBAL DEFAULT 100000'
+
+readelf -WS $t/b.o > $t/log1
+grep -Fq .data.100000 $t/log1
+
+readelf -Ws $t/b.o > $t/log2
+grep -Fq 'GLOBAL DEFAULT 100000' $t/log2
+grep -Fq 'ABS foo' $t/log2
+! grep -Fq 'ABS x68966' $t/log2 || false

test/elf/nocopyreloc.sh

@@ -8,6 +8,7 @@
 [ $MACHINE = ppc64le ] && skip
 [ $MACHINE = sh4 ] && skip
 [ $MACHINE = alpha ] && skip
+[[ $MACHINE = loongarch* ]] && skip
 
 cat <<EOF | $CC -shared -o $t/a.so -xc -
 int foo = 3;

test/elf/physical-image-base.sh

@@ -3,6 +3,9 @@
 
 [ $MACHINE = ppc64 ] && skip
 
+# BusyBox's grep can't handle capture groups (e.g. \1, \2 ...)
+grep --version 2>&1 | grep BusyBox && skip
+
 cat <<EOF | $CC -o $t/a.o -c -xc -
 #include <stdio.h>
 

test/elf/print-dependencies.sh

@@ -10,7 +10,5 @@ void foo();
 int main() { foo(); }
 EOF
 
-! $CC -B. -o $t/exe $t/a.o $t/b.o \
-  -Wl,--print-dependencies=full > $t/log 2> /dev/null
-
+! $CC -B. -o $t/exe $t/a.o $t/b.o -Wl,--print-dependencies > $t/log 2> /dev/null
 grep -q 'b\.o.*a\.o.*foo$' $t/log

test/elf/range-extension-thunk.sh

@@ -3,7 +3,6 @@
 
 # Skip if 32 bits as we use very large addresses in this test.
 [ $MACHINE = i386 ] && skip
-[ $MACHINE = arm32 ] && skip
 [ $MACHINE = riscv32 ] && skip
 
 # It looks like SPARC's runtime can't handle PLT if it's too far from GOT.

test/elf/relax-got-load.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -o $t/a.o -c -xc - -fPIC
+#include <stdio.h>
+extern char *msg;
+void hello() { printf("%s\n", msg); }
+EOF
+
+cat <<EOF | $CC -o $t/b.o -c -xc - -fPIC
+char *msg = "Hello world";
+void hello();
+int main() { hello(); }
+EOF
+
+$CC -B. -o $t/exe $t/a.o $t/b.o
+$QEMU $t/exe | grep -q 'Hello world'