Bump mold version to 2.39.1

https://github.com/rui314/mold/issues/1456

.github/workflows/install-extras.sh

@@ -5,7 +5,7 @@ apt-get install -y wget xz-utils
 
 # Install a 32-bit RISC-V toolchain
 mkdir /rv32
-wget -O- -q https://github.com/riscv-collab/riscv-gnu-toolchain/releases/download/2025.01.20/riscv32-glibc-ubuntu-22.04-gcc-nightly-2025.01.20-nightly.tar.xz | tar -C /rv32 --strip-components=1 --xz -xf -
+wget -O- --progress=dot:mega https://github.com/riscv-collab/riscv-gnu-toolchain/releases/download/2025.01.20/riscv32-glibc-ubuntu-22.04-gcc-nightly-2025.01.20-nightly.tar.xz | tar -C /rv32 --strip-components=1 --xz -xf -
 
 ln -sf /rv32/sysroot /usr/riscv32-linux-gnu
 echo '/rv32/bin/riscv32-unknown-linux-gnu-gcc -L/usr/riscv32-linux-gnu "$@"' > /usr/bin/riscv32-linux-gnu-gcc
@@ -18,7 +18,7 @@ done
 
 # Install a LoongArch toolchain
 mkdir /larch
-wget -O- -q https://github.com/loongson/build-tools/releases/download/2025.02.21/x86_64-cross-tools-loongarch64-binutils_2.44-gcc_14.2.0-glibc_2.41.tar.xz | tar -C /larch --strip-components=1 --xz -xf -
+wget -O- --progress=dot:mega https://github.com/loongson/build-tools/releases/download/2025.02.21/x86_64-cross-tools-loongarch64-binutils_2.44-gcc_14.2.0-glibc_2.41.tar.xz | tar -C /larch --strip-components=1 --xz -xf -
 
 cp -r /larch/loongarch64-unknown-linux-gnu/lib/* /larch/target/lib64
 ln -sf /larch/target /usr/loongarch64-linux-gnu
@@ -27,25 +27,39 @@ for i in gcc g++ objdump objcopy strip; do
   ln -sf /larch/bin/loongarch64-unknown-linux-gnu-$i /usr/bin/loongarch64-linux-gnu-$i
 done
 
-wget -O /usr/local/bin/qemu-loongarch64 -q https://github.com/loongson/build-tools/releases/download/2024.11.01/qemu-loongarch64
+wget -O /usr/local/bin/qemu-loongarch64 --progress=dot:mega https://github.com/loongson/build-tools/releases/download/2024.11.01/qemu-loongarch64
 chmod 755 /usr/local/bin/qemu-loongarch64
 
+# Install ARM32 big-endian toolchain
+mkdir /armeb
+wget -O- --progress=dot:mega https://toolchains.bootlin.com/downloads/releases/toolchains/armebv7-eabihf/tarballs/armebv7-eabihf--glibc--stable-2024.05-1.tar.xz | tar -C /armeb --strip-components=1 --xz -xf -
+
+ln -sf /armeb/armeb-buildroot-linux-gnueabihf/sysroot /usr/armeb-linux-gnueabihf
+echo '/armeb/bin/armeb-linux-gcc -L/usr/armeb-linux-gnu "$@"' > /usr/bin/armeb-linux-gnueabihf-gcc
+echo '/armeb/bin/armeb-linux-g++ -L/usr/armeb-linux-gnu "$@"' > /usr/bin/armeb-linux-gnueabihf-g++
+chmod 755 /usr/bin/armeb-linux-gnueabihf-{gcc,g++}
+
+for i in objdump objcopy strip; do
+  ln -sf /armeb/bin/armeb-linux-$i /usr/bin/armeb-linux-gnueabihf-$i
+done
+
 # Install ARM64 big-endian toolchain
 mkdir /aarch64be
-wget -O- -q https://sources.buildroot.net/toolchain-external-arm-aarch64-be/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64_be-none-linux-gnu.tar.xz | tar -C /aarch64be --strip-components=1 --xz -xf -
+wget -O- --progress=dot:mega https://toolchains.bootlin.com/downloads/releases/toolchains/aarch64be/tarballs/aarch64be--glibc--stable-2024.05-1.tar.xz | tar -C /aarch64be --strip-components=1 --xz -xf -
 
-ln -sf /aarch64be/aarch64_be-none-linux-gnu/libc /usr/aarch64_be-linux-gnu
-echo '/aarch64be/bin/aarch64_be-none-linux-gnu-gcc -L/usr/aarch64_be-linux-gnu "$@"' > /usr/bin/aarch64_be-linux-gnu-gcc
-echo '/aarch64be/bin/aarch64_be-none-linux-gnu-g++ -L/usr/aarch64_be-linux-gnu "$@"' > /usr/bin/aarch64_be-linux-gnu-g++
+ln -sf /aarch64be/aarch64_be-buildroot-linux-gnu/sysroot /usr/aarch64_be-linux-gnu
+echo '/aarch64be/bin/aarch64_be-linux-gcc -L/usr/aarch64_be-linux-gnu "$@"' > /usr/bin/aarch64_be-linux-gnu-gcc
+echo '/aarch64be/bin/aarch64_be-linux-g++ -L/usr/aarch64_be-linux-gnu "$@"' > /usr/bin/aarch64_be-linux-gnu-g++
 chmod 755 /usr/bin/aarch64_be-linux-gnu-{gcc,g++}
 
 for i in objdump objcopy strip; do
-  ln -sf /aarch64be/bin/aarch64_be-none-linux-gnu-$i /usr/bin/aarch64_be-linux-gnu-$i
+  ln -sf /aarch64be/bin/aarch64_be-linux-$i /usr/bin/aarch64_be-linux-gnu-$i
 done
 
 # Install SH4 big-endian toolchain
 mkdir /sh4aeb
-wget -O- -q https://toolchains.bootlin.com/downloads/releases/toolchains/sh-sh4aeb/tarballs/sh-sh4aeb--glibc--stable-2024.05-1.tar.xz | tar -C /sh4aeb --strip-components=1 --xz -xf -
+wget -O- --progress=dot:mega https://toolchains.bootlin.com/downloads/releases/toolchains/sh-sh4aeb/tarballs/sh-sh4aeb--glibc--stable-2024.05-1.tar.xz | tar -C /sh4aeb --strip-components=1 --xz -xf -
+
 ln -sf /sh4aeb/sh4aeb-buildroot-linux-gnu/sysroot /usr/sh4aeb-linux-gnu
 echo '/sh4aeb/bin/sh4aeb-linux-gcc -L/usr/sh4aeb-linux-gnu "$@"' > /usr/bin/sh4aeb-linux-gnu-gcc
 echo '/sh4aeb/bin/sh4aeb-linux-g++ -L/usr/sh4aeb-linux-gnu "$@"' > /usr/bin/sh4aeb-linux-gnu-g++
@@ -57,5 +71,5 @@ done
 
 # Install Intel SDE CPU emulator for CET-related tests
 mkdir /sde
-wget -O- -q https://downloadmirror.intel.com/843185/sde-external-9.48.0-2024-11-25-lin.tar.xz | tar -C /sde --strip-components=1 --xz -xf -
+wget -O- --progress=dot:mega https://downloadmirror.intel.com/843185/sde-external-9.48.0-2024-11-25-lin.tar.xz | tar -C /sde --strip-components=1 --xz -xf -
 ln -s /sde/sde64 /usr/bin

CMakeLists.txt

@@ -44,7 +44,7 @@
 # features and behave exactly the same.
 
 cmake_minimum_required(VERSION 3.14)
-project(mold VERSION 2.37.1)
+project(mold VERSION 2.39.1)
 
 include(CMakeDependentOption)
 include(CheckSymbolExists)
@@ -62,7 +62,11 @@ endif()
 # Build mold itself using mold if -DMOLD_USE_MOLD=ON
 option(MOLD_USE_MOLD "Use mold to build mold" OFF)
 if(MOLD_USE_MOLD)
-  target_link_options(mold PRIVATE -fuse-ld=mold -Wl,--gdb-index)
+  target_link_options(mold PRIVATE -fuse-ld=mold)
+
+  if(CMAKE_BUILD_TYPE MATCHES "Deb")
+    target_link_options(mold PRIVATE -Wl,--gdb-index)
+  endif()
 
   if(CMAKE_BUILD_TYPE MATCHES "^Rel")
     target_link_options(mold PRIVATE -Wl,--gc-sections,--icf=safe)
@@ -310,10 +314,12 @@ 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 ARM64LE ARM64BE ARM32 RV32LE RV32BE RV64LE RV64BE PPC32
-  PPC64V1 PPC64V2 S390X SPARC64 M68K SH4LE SH4BE LOONGARCH32 LOONGARCH64)
+  X86_64 I386 ARM64LE ARM64BE ARM32LE ARM32BE RV32LE RV32BE RV64LE
+  RV64BE PPC32 PPC64V1 PPC64V2 S390X SPARC64 M68K SH4LE SH4BE
+  LOONGARCH32 LOONGARCH64)
 
 list(APPEND MOLD_ELF_TEMPLATE_FILES
+  src/arch-arm32.cc
   src/arch-arm64.cc
   src/arch-loongarch.cc
   src/arch-riscv.cc
@@ -385,7 +391,6 @@ target_sources(mold PRIVATE
   lib/perf.cc
   lib/random.cc
   lib/tar.cc
-  src/arch-arm32.cc
   src/arch-i386.cc
   src/arch-m68k.cc
   src/arch-ppc32.cc

README.md

@@ -120,7 +120,7 @@ accept an absolute path as an argument for `-fuse-ld`.
 Create `.cargo/config.toml` in your project directory with the following:
 
 ```toml
-[target.x86_64-unknown-linux-gnu]
+[target.'cfg(target_os = "linux")']
 linker = "clang"
 rustflags = ["-C", "link-arg=-fuse-ld=/path/to/mold"]
 ```
@@ -131,7 +131,7 @@ example above, we use `clang` as a linker driver since it always accepts the
 may be able to remove the `linker = "clang"` line.
 
 ```toml
-[target.x86_64-unknown-linux-gnu]
+[target.'cfg(target_os = "linux")']
 rustflags = ["-C", "link-arg=-fuse-ld=mold"]
 ```
 

dist.sh

@@ -126,10 +126,10 @@ RUN mkdir /build && \
   make install && \
   rm -rf /build
 
-# Build LLVM 19
+# Build LLVM 20
 RUN mkdir /build && \
   cd /build && \
-  wget -O- --no-check-certificate --progress=dot:mega https://github.com/llvm/llvm-project/archive/refs/tags/llvmorg-19.1.7.tar.gz | tar xzf - --strip-components=1 && \
+  wget -O- --no-check-certificate --progress=dot:mega https://github.com/llvm/llvm-project/archive/refs/tags/llvmorg-20.1.3.tar.gz | tar xzf - --strip-components=1 && \
   mkdir b && \
   cd b && \
   cmake -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_PROJECTS=clang ../llvm && \
@@ -195,14 +195,22 @@ dest=mold-$version-$arch-linux
 
 # We use the timestamp of the last Git commit as the file timestamp
 # for build artifacts.
-timestamp="$(git log -1 --format=%ci)"
+timestamp=$(git log -1 --format=%ct)
 
 # Build mold in a container.
 mkdir -p dist
 
+# If the host is ARM64, `uname -m` in an ARM32 container running on it
+# reports it not as ARM32 but as ARM64, which confuses BLAKE3's cmake script
+# and erroneously enables NEON SIMD instructions. `setarch` can be used
+# to change the output of `uname -m`.
+[ $arch = arm ] && setarch='setarch linux32'
+
 podman run --arch $arch -it --rm --userns=host --pids-limit=-1 --network=none \
-  -v "$(pwd):/mold:ro" -v "$(pwd)/dist:/dist" $image bash -c "
+  --pull=never -v "$(pwd):/mold:ro" -v "$(pwd)/dist:/dist" $image \
+   $setarch bash -c "
 set -e
+export SOURCE_DATE_EPOCH=$timestamp
 mkdir /build
 cd /build
 cmake -DCMAKE_BUILD_TYPE=Release -DMOLD_MOSTLY_STATIC=1 -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ /mold
@@ -212,7 +220,7 @@ cmake -DMOLD_USE_MOLD=1 .
 cmake --build . -j\$(nproc)
 ctest --output-on-failure -j\$(nproc)
 cmake --install . --prefix $dest --strip
-find $dest -print | xargs touch --no-dereference --date='$timestamp'
+find $dest -print | xargs touch --no-dereference --date=@$timestamp
 find $dest -print | sort | tar -cf - --no-recursion --files-from=- | gzip -9nc > /dist/$dest.tar.gz
 cp mold /dist
 sha256sum /dist/$dest.tar.gz

lib/gentoo-test.sh

@@ -54,14 +54,14 @@ dir=gentoo/$git_hash
 
 mkdir -p "$dir"/success "$dir"/failure
 
-$podman nice -n 19 bash -c "$cmd1 && $cmd2 && $cmd3" >& "$dir"/"$filename".mold
+$podman chrt --idle 0 nice -n 19 bash -c "$cmd1 && $cmd2 && $cmd3" >& "$dir"/"$filename".mold
 if [ $? = 0 ]; then
   mv "$dir"/"$filename".mold "$dir"/success
 else
   mv "$dir"/"$filename".mold "$dir"/failure
 fi
 
-$podman nice -n 19 bash -c "$cmd2 && $cmd3" >& "$dir"/"$filename".ld
+$podman chrt --idle 0 nice -n 19 bash -c "$cmd2 && $cmd3" >& "$dir"/"$filename".ld
 if [ $? = 0 ]; then
   mv "$dir"/"$filename".ld "$dir"/success
 else

src/arch-arm32.cc

@@ -31,18 +31,21 @@
 //
 // https://github.com/ARM-software/abi-aa/blob/main/aaelf32/aaelf32.rst
 
+#if MOLD_ARM32LE || MOLD_ARM32BE
+
 #include "mold.h"
 
 #include <tbb/parallel_for.h>
+#include <tbb/parallel_for_each.h>
 
 namespace mold {
 
-using E = ARM32;
+using E = MOLD_TARGET;
 
 template <>
 i64 get_addend(u8 *loc, const ElfRel<E> &rel) {
-  ul32 *arm = (ul32 *)loc;
-  ul16 *thm = (ul16 *)loc;
+  U32<E> *arm = (U32<E> *)loc;
+  U16<E> *thm = (U16<E> *)loc;
 
   switch (rel.r_type) {
   case R_ARM_ABS32:
@@ -59,7 +62,7 @@ i64 get_addend(u8 *loc, const ElfRel<E> &rel) {
   case R_ARM_TLS_GOTDESC:
   case R_ARM_TARGET1:
   case R_ARM_TARGET2:
-    return (il32)*arm;
+    return (I32<E>)*arm;
   case R_ARM_THM_JUMP8:
     return sign_extend(thm[0], 8) << 1;
   case R_ARM_THM_JUMP11:
@@ -125,7 +128,7 @@ i64 get_addend(u8 *loc, const ElfRel<E> &rel) {
 static void write_arm_mov(u8 *loc, u32 val) {
   u32 imm12 = bits(val, 11, 0);
   u32 imm4 = bits(val, 15, 12);
-  *(ul32 *)loc = (*(ul32 *)loc & 0xfff0'f000) | (imm4 << 16) | imm12;
+  *(U32<E> *)loc = (*(U32<E> *)loc & 0xfff0'f000) | (imm4 << 16) | imm12;
 }
 
 static void write_thm_b21(u8 *loc, u32 val) {
@@ -135,7 +138,7 @@ static void write_thm_b21(u8 *loc, u32 val) {
   u32 imm6 = bits(val, 17, 12);
   u32 imm11 = bits(val, 11, 1);
 
-  ul16 *buf = (ul16 *)loc;
+  U16<E> *buf = (U16<E> *)loc;
   buf[0] = (buf[0] & 0b1111'1011'1100'0000) | (S << 10) | imm6;
   buf[1] = (buf[1] & 0b1101'0000'0000'0000) | (J1 << 13) | (J2 << 11) | imm11;
 }
@@ -149,7 +152,7 @@ static void write_thm_b25(u8 *loc, u32 val) {
   u32 imm10 = bits(val, 21, 12);
   u32 imm11 = bits(val, 11, 1);
 
-  ul16 *buf = (ul16 *)loc;
+  U16<E> *buf = (U16<E> *)loc;
   buf[0] = (buf[0] & 0b1111'1000'0000'0000) | (S << 10) | imm10;
   buf[1] = (buf[1] & 0b1101'0000'0000'0000) | (J1 << 13) | (J2 << 11) | imm11;
 }
@@ -160,7 +163,7 @@ static void write_thm_mov(u8 *loc, u32 val) {
   u32 imm3 = bits(val, 10, 8);
   u32 imm8 = bits(val, 7, 0);
 
-  ul16 *buf = (ul16 *)loc;
+  U16<E> *buf = (U16<E> *)loc;
   buf[0] = (buf[0] & 0b1111'1011'1111'0000) | (i << 10) | imm4;
   buf[1] = (buf[1] & 0b1000'1111'0000'0000) | (imm3 << 12) | imm8;
 }
@@ -184,13 +187,13 @@ void write_addend(u8 *loc, i64 val, const ElfRel<E> &rel) {
   case R_ARM_TLS_GOTDESC:
   case R_ARM_TARGET1:
   case R_ARM_TARGET2:
-    *(ul32 *)loc = val;
+    *(U32<E> *)loc = val;
     break;
   case R_ARM_THM_JUMP8:
-    *(ul16 *)loc = (*(ul16 *)loc & 0xff00) | bits(val, 8, 1);
+    *(U16<E> *)loc = (*(U16<E> *)loc & 0xff00) | bits(val, 8, 1);
     break;
   case R_ARM_THM_JUMP11:
-    *(ul16 *)loc = (*(ul16 *)loc & 0xf800) | bits(val, 11, 1);
+    *(U16<E> *)loc = (*(U16<E> *)loc & 0xf800) | bits(val, 11, 1);
     break;
   case R_ARM_THM_CALL:
   case R_ARM_THM_JUMP24:
@@ -200,7 +203,7 @@ void write_addend(u8 *loc, i64 val, const ElfRel<E> &rel) {
   case R_ARM_CALL:
   case R_ARM_JUMP24:
   case R_ARM_PLT32:
-    *(ul32 *)loc = (*(ul32 *)loc & 0xff00'0000) | bits(val, 25, 2);
+    *(U32<E> *)loc = (*(U32<E> *)loc & 0xff00'0000) | bits(val, 25, 2);
     break;
   case R_ARM_MOVW_PREL_NC:
   case R_ARM_MOVW_ABS_NC:
@@ -209,7 +212,7 @@ void write_addend(u8 *loc, i64 val, const ElfRel<E> &rel) {
     write_arm_mov(loc, val);
     break;
   case R_ARM_PREL31:
-    *(ul32 *)loc = (*(ul32 *)loc & 0x8000'0000) | (val & 0x7fff'ffff);
+    *(U32<E> *)loc = (*(U32<E> *)loc & 0x8000'0000) | (val & 0x7fff'ffff);
     break;
   case R_ARM_THM_MOVW_PREL_NC:
   case R_ARM_THM_MOVW_ABS_NC:
@@ -236,7 +239,7 @@ void write_plt_header(Context<E> &ctx, u8 *buf) {
   };
 
   memcpy(buf, insn, sizeof(insn));
-  *(ul32 *)(buf + 16) = ctx.gotplt->shdr.sh_addr - ctx.plt->shdr.sh_addr - 16;
+  *(U32<E> *)(buf + 16) = ctx.gotplt->shdr.sh_addr - ctx.plt->shdr.sh_addr - 16;
 }
 
 static constexpr ul32 plt_entry[] = {
@@ -249,13 +252,13 @@ static constexpr ul32 plt_entry[] = {
 template <>
 void write_plt_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
   memcpy(buf, plt_entry, sizeof(plt_entry));
-  *(ul32 *)(buf + 12) = sym.get_gotplt_addr(ctx) - sym.get_plt_addr(ctx) - 12;
+  *(U32<E> *)(buf + 12) = sym.get_gotplt_addr(ctx) - sym.get_plt_addr(ctx) - 12;
 }
 
 template <>
 void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
   memcpy(buf, plt_entry, sizeof(plt_entry));
-  *(ul32 *)(buf + 12) = sym.get_got_pltgot_addr(ctx) - sym.get_plt_addr(ctx) - 12;
+  *(U32<E> *)(buf + 12) = sym.get_got_pltgot_addr(ctx) - sym.get_plt_addr(ctx) - 12;
 }
 
 template <>
@@ -267,24 +270,16 @@ void EhFrameSection<E>::apply_eh_reloc(Context<E> &ctx, const ElfRel<E> &rel,
   case R_NONE:
     break;
   case R_ARM_ABS32:
-    *(ul32 *)loc = val;
+    *(U32<E> *)loc = val;
     break;
   case R_ARM_REL32:
-    *(ul32 *)loc = val - this->shdr.sh_addr - offset;
+    *(U32<E> *)loc = val - this->shdr.sh_addr - offset;
     break;
   default:
     Fatal(ctx) << "unsupported relocation in .eh_frame: " << rel;
   }
 }
 
-static Thunk<E> &get_reachable_thunk(OutputSection<E> &osec, u64 addr) {
-  auto it = std::upper_bound(osec.thunks.begin(), osec.thunks.end(), addr,
-                             [](u64 addr, std::unique_ptr<Thunk<E>> &thunk) {
-    return addr < thunk->get_addr();
-  });
-  return **it;
-}
-
 template <>
 void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
   std::span<const ElfRel<E>> rels = get_rels(ctx);
@@ -312,7 +307,11 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     auto get_arm_thunk_addr   = [&] { return sym.get_thunk_addr(ctx, P) + 4; };
 
     auto get_tlsdesc_trampoline_addr = [&] {
-      return get_reachable_thunk(*output_section, P).get_addr();
+      auto it = ranges::upper_bound(output_section->thunks, P, {},
+                                    [](std::unique_ptr<Thunk<E>> &thunk) {
+        return thunk->get_addr();
+      });
+      return (*it)->get_addr();
     };
 
     switch (rel.r_type) {
@@ -320,13 +319,13 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     case R_ARM_TARGET1:
       break;
     case R_ARM_REL32:
-      *(ul32 *)loc = S + A - P;
+      *(U32<E> *)loc = S + A - P;
       break;
     case R_ARM_THM_CALL: {
       if (sym.is_remaining_undef_weak()) {
         // On ARM, calling an weak undefined symbol jumps to the
         // next instruction.
-        *(ul32 *)loc = 0x8000'f3af; // NOP.W
+        *(U32<E> *)loc = 0x8000'f3af; // NOP.W
         break;
       }
 
@@ -337,61 +336,61 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       i64 val2 = align_to(S + A - P, 4);
 
       if (T && is_int(val1, 25)) {
-        *(ul16 *)(loc + 2) |= 0x1000;  // BL
+        *(U16<E> *)(loc + 2) |= 0x1000;  // BL
         write_thm_b25(loc, val1);
       } else if (!T && is_int(val2, 25)) {
-        *(ul16 *)(loc + 2) &= ~0x1000; // BLX
+        *(U16<E> *)(loc + 2) &= ~0x1000; // BLX
         write_thm_b25(loc, val2);
       } else {
-        *(ul16 *)(loc + 2) |= 0x1000;  // BL
+        *(U16<E> *)(loc + 2) |= 0x1000;  // BL
         write_thm_b25(loc, get_thumb_thunk_addr() + A - P);
       }
       break;
     }
     case R_ARM_BASE_PREL:
-      *(ul32 *)loc = GOT + A - P;
+      *(U32<E> *)loc = GOT + A - P;
       break;
     case R_ARM_GOTOFF32:
-      *(ul32 *)loc = ((S + A) | T) - GOT;
+      *(U32<E> *)loc = ((S + A) | T) - GOT;
       break;
     case R_ARM_GOT_PREL:
     case R_ARM_TARGET2:
-      *(ul32 *)loc = GOT + G + A - P;
+      *(U32<E> *)loc = GOT + G + A - P;
       break;
     case R_ARM_GOT_BREL:
-      *(ul32 *)loc = G + A;
+      *(U32<E> *)loc = G + A;
       break;
     case R_ARM_CALL: {
       if (sym.is_remaining_undef_weak()) {
-        *(ul32 *)loc = 0xe320'f000; // NOP
+        *(U32<E> *)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);
+      bool is_bl = ((*(U32<E> *)loc & 0xff00'0000) == 0xeb00'0000);
+      bool is_blx = ((*(U32<E> *)loc & 0xfe00'0000) == 0xfa00'0000);
       if (!is_bl && !is_blx)
         Fatal(ctx) << *this << ": R_ARM_CALL refers to neither BL nor BLX";
 
       i64 val = S + A - P;
       if (is_int(val, 26)) {
         if (T) {
-          *(ul32 *)loc = 0xfa00'0000; // BLX
-          *(ul32 *)loc |= (bit(val, 1) << 24) | bits(val, 25, 2);
+          *(U32<E> *)loc = 0xfa00'0000; // BLX
+          *(U32<E> *)loc |= (bit(val, 1) << 24) | bits(val, 25, 2);
         } else {
-          *(ul32 *)loc = 0xeb00'0000; // BL
-          *(ul32 *)loc |= bits(val, 25, 2);
+          *(U32<E> *)loc = 0xeb00'0000; // BL
+          *(U32<E> *)loc |= bits(val, 25, 2);
         }
       } else {
-        *(ul32 *)loc = 0xeb00'0000; // BL
-        *(ul32 *)loc |= bits(get_arm_thunk_addr() + A - P, 25, 2);
+        *(U32<E> *)loc = 0xeb00'0000; // BL
+        *(U32<E> *)loc |= bits(get_arm_thunk_addr() + A - P, 25, 2);
       }
       break;
     }
     case R_ARM_JUMP24: {
       if (sym.is_remaining_undef_weak()) {
-        *(ul32 *)loc = 0xe320'f000; // NOP
+        *(U32<E> *)loc = 0xe320'f000; // NOP
         break;
       }
 
@@ -404,26 +403,26 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       i64 val = S + A - P;
       if (T || !is_int(val, 26))
         val = get_arm_thunk_addr() + A - P;
-      *(ul32 *)loc = (*(ul32 *)loc & 0xff00'0000) | bits(val, 25, 2);
+      *(U32<E> *)loc = (*(U32<E> *)loc & 0xff00'0000) | bits(val, 25, 2);
       break;
     }
     case R_ARM_PLT32:
       if (sym.is_remaining_undef_weak()) {
-        *(ul32 *)loc = 0xe320'f000; // NOP
+        *(U32<E> *)loc = 0xe320'f000; // NOP
       } else {
         u64 val = (T ? get_arm_thunk_addr() : S) + A - P;
-        *(ul32 *)loc = (*(ul32 *)loc & 0xff00'0000) | bits(val, 25, 2);
+        *(U32<E> *)loc = (*(U32<E> *)loc & 0xff00'0000) | bits(val, 25, 2);
       }
       break;
     case R_ARM_THM_JUMP8:
       check(S + A - P, -(1 << 8), 1 << 8);
-      *(ul16 *)loc &= 0xff00;
-      *(ul16 *)loc |= bits(S + A - P, 8, 1);
+      *(U16<E> *)loc &= 0xff00;
+      *(U16<E> *)loc |= bits(S + A - P, 8, 1);
       break;
     case R_ARM_THM_JUMP11:
       check(S + A - P, -(1 << 11), 1 << 11);
-      *(ul16 *)loc &= 0xf800;
-      *(ul16 *)loc |= bits(S + A - P, 11, 1);
+      *(U16<E> *)loc &= 0xf800;
+      *(U16<E> *)loc |= bits(S + A - P, 11, 1);
       break;
     case R_ARM_THM_JUMP19:
       check(S + A - P, -(1 << 20), 1 << 20);
@@ -431,7 +430,7 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       break;
     case R_ARM_THM_JUMP24: {
       if (sym.is_remaining_undef_weak()) {
-        *(ul32 *)loc = 0x8000'f3af; // NOP
+        *(U32<E> *)loc = 0x8000'f3af; // NOP
         break;
       }
 
@@ -454,8 +453,8 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       break;
     case R_ARM_PREL31:
       check(S + A - P, -(1LL << 30), 1LL << 30);
-      *(ul32 *)loc &= 0x8000'0000;
-      *(ul32 *)loc |= (S + A - P) & 0x7fff'ffff;
+      *(U32<E> *)loc &= 0x8000'0000;
+      *(U32<E> *)loc |= (S + A - P) & 0x7fff'ffff;
       break;
     case R_ARM_THM_MOVW_ABS_NC:
       write_thm_mov(loc, (S + A) | T);
@@ -473,19 +472,19 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       write_thm_mov(loc, (S + A) >> 16);
       break;
     case R_ARM_TLS_GD32:
-      *(ul32 *)loc = sym.get_tlsgd_addr(ctx) + A - P;
+      *(U32<E> *)loc = sym.get_tlsgd_addr(ctx) + A - P;
       break;
     case R_ARM_TLS_LDM32:
-      *(ul32 *)loc = ctx.got->get_tlsld_addr(ctx) + A - P;
+      *(U32<E> *)loc = ctx.got->get_tlsld_addr(ctx) + A - P;
       break;
     case R_ARM_TLS_LDO32:
-      *(ul32 *)loc = S + A - ctx.dtp_addr;
+      *(U32<E> *)loc = S + A - ctx.dtp_addr;
       break;
     case R_ARM_TLS_IE32:
-      *(ul32 *)loc = sym.get_gottp_addr(ctx) + A - P;
+      *(U32<E> *)loc = sym.get_gottp_addr(ctx) + A - P;
       break;
     case R_ARM_TLS_LE32:
-      *(ul32 *)loc = S + A - ctx.tp_addr;
+      *(U32<E> *)loc = S + A - ctx.tp_addr;
       break;
     case R_ARM_TLS_GOTDESC:
       // ARM32 TLSDESC uses the following code sequence to materialize
@@ -514,35 +513,35 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       //  .L2: .word   foo(gottpoff) + . - .L1
       if (sym.has_tlsdesc(ctx)) {
         // A is odd if the corresponding TLS_CALL is Thumb.
-        *(ul32 *)loc = sym.get_tlsdesc_addr(ctx) - P + A - ((A & 1) ? 6 : 4);
+        *(U32<E> *)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);
+        *(U32<E> *)loc = sym.get_gottp_addr(ctx) - P + A - ((A & 1) ? 5 : 8);
       } else {
-        *(ul32 *)loc = S - ctx.tp_addr;
+        *(U32<E> *)loc = S - ctx.tp_addr;
       }
       break;
     case R_ARM_TLS_CALL:
       if (sym.has_tlsdesc(ctx)) {
-        *(ul32 *)loc = 0xeb00'0000; // bl 0
-        *(ul32 *)loc |= bits(get_tlsdesc_trampoline_addr() - P - 8, 25, 2);
+        *(U32<E> *)loc = 0xeb00'0000; // bl 0
+        *(U32<E> *)loc |= bits(get_tlsdesc_trampoline_addr() - P - 8, 25, 2);
       } else if (sym.has_gottp(ctx)) {
-        *(ul32 *)loc = 0xe79f'0000; // ldr r0, [pc, r0]
+        *(U32<E> *)loc = 0xe79f'0000; // ldr r0, [pc, r0]
       } else {
-        *(ul32 *)loc = 0xe320'f000; // nop
+        *(U32<E> *)loc = 0xe320'f000; // nop
       }
       break;
     case R_ARM_THM_TLS_CALL:
       if (sym.has_tlsdesc(ctx)) {
         u64 val = align_to(get_tlsdesc_trampoline_addr() - P - 4, 4);
         write_thm_b25(loc, val);
-        *(ul16 *)(loc + 2) &= ~0x1000; // rewrite BL with BLX
+        *(U16<E> *)(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]
+        *(U16<E> *)loc = 0x4478;         // add r0, pc
+        *(U16<E> *)(loc + 2) = 0x6800;   // ldr r0, [r0]
       } else {
-        *(ul32 *)loc = 0x8000'f3af;    // nop.w
+        *(U32<E> *)loc = 0x8000'f3af;    // nop.w
       }
       break;
     default:
@@ -573,15 +572,15 @@ void InputSection<E>::apply_reloc_nonalloc(Context<E> &ctx, u8 *base) {
     switch (rel.r_type) {
     case R_ARM_ABS32:
       if (std::optional<u64> val = get_tombstone(sym, frag))
-        *(ul32 *)loc = *val;
+        *(U32<E> *)loc = *val;
       else
-        *(ul32 *)loc = S + A;
+        *(U32<E> *)loc = S + A;
       break;
     case R_ARM_TLS_LDO32:
       if (std::optional<u64> val = get_tombstone(sym, frag))
-        *(ul32 *)loc = *val;
+        *(U32<E> *)loc = *val;
       else
-        *(ul32 *)loc = S + A - ctx.dtp_addr;
+        *(U32<E> *)loc = S + A - ctx.dtp_addr;
       break;
     default:
       Fatal(ctx) << *this << ": invalid relocation for non-allocated sections: "
@@ -703,7 +702,7 @@ void Thunk<E>::copy_buf(Context<E> &ctx) {
 
   for (Symbol<E> *sym : symbols) {
     memcpy(buf, entry, sizeof(entry));
-    *(ul32 *)(buf + 12) = sym->get_addr(ctx) - P - 16;
+    *(U32<E> *)(buf + 12) = sym->get_addr(ctx) - P - 16;
 
     buf += sizeof(entry);
     P += sizeof(entry);
@@ -712,10 +711,14 @@ void Thunk<E>::copy_buf(Context<E> &ctx) {
 
 template <>
 u64 get_eflags(Context<E> &ctx) {
-  return EF_ARM_EABI_VER5;
+  if constexpr (E::is_le)
+    return EF_ARM_EABI_VER5;
+  else
+    return EF_ARM_EABI_VER5 | EF_ARM_BE8;
 }
 
-void create_arm_exidx_section(Context<E> &ctx) {
+template <>
+void create_arm_exidx_section<E>(Context<E> &ctx) {
   for (i64 i = 0; i < ctx.chunks.size(); i++) {
     OutputSection<E> *osec = ctx.chunks[i]->to_osec();
 
@@ -732,28 +735,20 @@ void create_arm_exidx_section(Context<E> &ctx) {
   }
 }
 
-void Arm32ExidxSection::compute_section_size(Context<E> &ctx) {
+template <>
+void Arm32ExidxSection<E>::compute_section_size(Context<E> &ctx) {
   output_section.compute_section_size(ctx);
   this->shdr.sh_size = output_section.shdr.sh_size + 8; // +8 for sentinel
 }
 
-void Arm32ExidxSection::update_shdr(Context<E> &ctx) {
+template <>
+void Arm32ExidxSection<E>::update_shdr(Context<E> &ctx) {
   // .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 (Chunk<E> *chunk = find_chunk(ctx, ".text"))
     this->shdr.sh_link = chunk->shndx;
 }
 
-void Arm32ExidxSection::remove_duplicate_entries(Context<E> &ctx) {
-  this->shdr.sh_size = get_contents(ctx).size();
-}
-
-void Arm32ExidxSection::copy_buf(Context<E> &ctx) {
-  std::vector<u8> contents = get_contents(ctx);
-  assert(this->shdr.sh_size == contents.size());
-  write_vector(ctx.buf + this->shdr.sh_offset, contents);
-}
-
 // Returns the end of the text segment
 static u64 get_text_end(Context<E> &ctx) {
   u64 ret = 0;
@@ -772,7 +767,8 @@ static u64 get_text_end(Context<E> &ctx) {
 // likely that it's due to some historical reason.
 //
 // This function returns contents of .ARM.exidx.
-std::vector<u8> Arm32ExidxSection::get_contents(Context<E> &ctx) {
+template <>
+std::vector<u8> Arm32ExidxSection<E>::get_contents(Context<E> &ctx) {
   // .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
@@ -788,8 +784,8 @@ std::vector<u8> Arm32ExidxSection::get_contents(Context<E> &ctx) {
   const u32 CANTUNWIND = 1;
 
   struct Entry {
-    ul32 addr;
-    ul32 val;
+    U32<E> addr;
+    U32<E> val;
   };
 
   // We reserve one extra slot for the sentinel
@@ -820,19 +816,13 @@ std::vector<u8> Arm32ExidxSection::get_contents(Context<E> &ctx) {
       ent[i].val = 0x7fff'ffff & (ent[i].val + offset);
   });
 
-  std::sort(ent, ent + num_entries, [](const Entry &a, const Entry &b) {
-    return a.addr < b.addr;
-  });
+  ranges::sort(ent, ent + num_entries, {}, &Entry::addr);
 
   // Remove duplicate adjacent entries. That is, if two adjacent functions
   // have the same compact unwind info or are both CANTUNWIND, we can
   // merge them into a single address range.
-  auto it = std::unique(ent, ent + num_entries,
-                        [](const Entry &a, const Entry &b) {
-    return a.val == b.val;
-  });
-
-  num_entries = it - ent;
+  auto tail = ranges::unique(ent, ent + num_entries, {}, &Entry::val);
+  num_entries -= tail.size();
   buf.resize(num_entries * sizeof(Entry));
 
   // Make addresses relative to themselves.
@@ -846,4 +836,79 @@ std::vector<u8> Arm32ExidxSection::get_contents(Context<E> &ctx) {
   return buf;
 }
 
+template <>
+void Arm32ExidxSection<E>::remove_duplicate_entries(Context<E> &ctx) {
+  this->shdr.sh_size = get_contents(ctx).size();
+}
+
+template <>
+void Arm32ExidxSection<E>::copy_buf(Context<E> &ctx) {
+  std::vector<u8> contents = get_contents(ctx);
+  assert(this->shdr.sh_size == contents.size());
+  write_vector(ctx.buf + this->shdr.sh_offset, contents);
+}
+
+#if MOLD_ARM32BE
+// Even though using ARM32 in big-endian mode is very rare, the processor
+// technically supports both little- and big-endian modes. There are two
+// variants of big-endian mode: BE32 and BE8. In BE32, instructions and
+// data are encoded in big-endian. In BE8, instructions are encoded in
+// little-endian, and only data is in big-endian. BE8 is the de facto
+// standard for ARMv6 or later. We support only BE8.
+//
+// A tricky thing is that instructions in an object file are always
+// big-endian if the file is compiled for big-endian mode. It is the
+// linker's responsibility to rewrite instructions from big-endian to
+// little-endian for an BE8 output. This function does that.
+//
+// The text section may contain a mix of 32-bit ARM instructions, 16-bit
+// Thumb instructions, and data. We need to distinguish them to swap 4
+// bytes, 2 bytes, or not swap bytes, respectively. The beginning of ARM
+// code, Thumb code, and data is labeled with a mapping symbol of $a, $t,
+// and $d, respectively. We use mapping symbols to determine what to do
+// with the text section.
+void arm32be_swap_bytes(Context<E> &ctx) {
+  tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
+    // Collect mapping symbols
+    std::vector<Symbol<E> *> syms;
+
+    for (Symbol<E> *sym : file->get_local_syms())
+      if (InputSection<E> *isec = sym->get_input_section())
+        if (isec->is_alive && (isec->shdr().sh_flags & SHF_EXECINSTR))
+          if (std::string_view x = sym->name();
+              x == "$a" || x.starts_with("$a.") ||
+              x == "$t" || x.starts_with("$t.") ||
+              x == "$d" || x.starts_with("$d."))
+            syms.push_back(sym);
+
+    // Group mapping symbols by input section and sort by address
+    ranges::stable_sort(syms, {}, [](const Symbol<E> *sym) {
+      return std::tuple{(uintptr_t)sym->get_input_section(), sym->value};
+    });
+
+    // Swap bytes
+    for (i64 i = 0; i < syms.size(); i++) {
+      Symbol<E> &sym = *syms[i];
+      if (sym.name().starts_with("$d"))
+        continue;
+
+      InputSection<E> &isec = *sym.get_input_section();
+      u8 *base = ctx.buf + isec.output_section->shdr.sh_offset + isec.offset;
+
+      u8 *end;
+      if (i + 1 < syms.size() && syms[i + 1]->get_input_section() == &isec)
+        end = base + syms[i + 1]->value;
+      else
+        end = base + isec.sh_size;
+
+      i64 sz = sym.name().starts_with("$a") ? 4 : 2;
+      for (u8 *p = base + sym.value; p < end; p += sz)
+        std::reverse(p, p + sz);
+    }
+  });
+}
+#endif
+
 } // namespace mold
+
+#endif

src/arch-ppc64v1.cc

@@ -484,17 +484,12 @@ static InputSection<E> *get_opd_section(ObjectFile<E> &file) {
 }
 
 static ElfRel<E> *
-get_relocation_at(Context<E> &ctx, InputSection<E> &isec, i64 offset) {
+get_relocation_at(Context<E> &ctx, InputSection<E> &isec, u64 offset) {
   std::span<ElfRel<E>> rels = isec.get_rels(ctx);
-
-  auto it = std::lower_bound(rels.begin(), rels.end(), offset,
-                             [](const ElfRel<E> &r, i64 offset) {
-    return r.r_offset < offset;
+  auto it = ranges::lower_bound(rels, offset, {}, [](const ElfRel<E> &r) {
+    return (u64)r.r_offset;
   });
-
-  if (it == rels.end())
-    return nullptr;
-  if (it->r_offset != offset)
+  if (it == rels.end() || it->r_offset != offset)
     return nullptr;
   return &*it;
 }
@@ -509,9 +504,7 @@ struct OpdSymbol {
 static Symbol<E> *
 get_opd_sym_at(std::span<OpdSymbol> syms, u64 offset) {
   auto it = ranges::lower_bound(syms, offset, {}, &OpdSymbol::r_offset);
-  if (it == syms.end())
-    return nullptr;
-  if (it->r_offset != offset)
+  if (it == syms.end() || it->r_offset != offset)
     return nullptr;
   return it->sym;
 }

src/arch-riscv.cc

@@ -303,8 +303,8 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
 
     switch (rel.r_type) {
     case R_RISCV_32:
-      assert(E::is_64);
-      *(U32<E> *)loc = S + A;
+      if (E::is_64)
+        *(U32<E> *)loc = S + A;
       break;
     case R_RISCV_64:
       break;
@@ -321,12 +321,6 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       i64 val = S + A - P;
       i64 rd = get_rd(buf + rel.r_offset + 4);
 
-      // Calling an undefined weak symbol does not make sense.
-      // We make such call into an infinite loop. This should
-      // help debugging of a faulty program.
-      if (sym.esym().is_undef_weak())
-        val = 0;
-
       if (removed_bytes == 4) {
         // auipc + jalr -> jal
         *(ul32 *)loc = (rd << 7) | 0b1101111;
@@ -473,7 +467,7 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
     }
     case R_RISCV_TLSDESC_HI20:
       // RISC-V TLSDESC uses the following code sequence to materialize
-      // a TP-relative address in x0.
+      // a TP-relative address in a0.
       //
       //   .L0:
       //   auipc  tX, 0

src/cmdline.cc

@@ -587,6 +587,7 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
   bool warn_shared_textrel = false;
   bool error_unresolved_symbols = true;
   std::optional<SeparateCodeKind> z_separate_code;
+  std::optional<bool> allow_shlib_undefined;
   std::optional<bool> report_undefined;
   std::optional<bool> z_relro;
   std::optional<bool> z_dynamic_undefined_weak;
@@ -732,7 +733,9 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       } else if (arg == "aarch64elfb" || arg == "aarch64linuxb") {
         ctx.arg.emulation = ARM64BE::name;
       } else if (arg == "armelf_linux_eabi") {
-        ctx.arg.emulation = ARM32::name;
+        ctx.arg.emulation = ARM32LE::name;
+      } else if (arg == "armelfb_linux_eabi") {
+        ctx.arg.emulation = ARM32BE::name;
       } else if (arg == "elf64lriscv") {
         ctx.arg.emulation = RV64LE::name;
       } else if (arg == "elf64briscv") {
@@ -1291,6 +1294,10 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
       }
     } else if (read_flag("no-build-id")) {
       ctx.arg.build_id.kind = BuildId::NONE;
+    } else if (read_flag("be8")) {
+      ctx.arg.be8 = true;
+    } else if (read_flag("be32")) {
+      ctx.arg.be8 = false;
     } else if (read_arg("format") || read_arg("b")) {
       if (arg == "binary")
         Fatal(ctx)
@@ -1303,9 +1310,9 @@ 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_flag("allow-shlib-undefined")) {
-      ctx.arg.allow_shlib_undefined = true;
+      allow_shlib_undefined = true;
     } else if (read_flag("no-allow-shlib-undefined")) {
-      ctx.arg.allow_shlib_undefined = false;
+      allow_shlib_undefined = false;
     } else if (read_arg("O")) {
     } else if (read_flag("EB")) {
     } else if (read_flag("EL")) {
@@ -1412,6 +1419,11 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
   if (ctx.arg.pic)
     ctx.arg.image_base = 0;
 
+  if (allow_shlib_undefined)
+    ctx.arg.allow_shlib_undefined = *allow_shlib_undefined;
+  else
+    ctx.arg.allow_shlib_undefined = ctx.arg.shared;
+
   if (!report_undefined)
     report_undefined = !ctx.arg.shared;
 
@@ -1485,6 +1497,9 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
   if (ctx.arg.image_base % ctx.page_size)
     Fatal(ctx) << "-image-base must be a multiple of -max-page-size";
 
+  if (ctx.arg.emulation == ARM32BE::name && !ctx.arg.be8)
+    Fatal(ctx) << "--be32 is not supported";
+
   if (ctx.arg.thread_count == 0)
     ctx.arg.thread_count = get_default_thread_count();
 
@@ -1495,9 +1510,7 @@ std::vector<std::string> parse_nonpositional_args(Context<E> &ctx) {
     std::string ver = ctx.arg.soname;
     if (ver.empty())
       ver = path_filename(ctx.arg.output);
-
     ctx.arg.version_definitions.push_back(ver);
-    ctx.default_version = VER_NDX_LAST_RESERVED + 1;
   }
 
   for (std::string_view path : version_scripts) {

src/elf.cc

@@ -232,7 +232,7 @@ std::string rel_to_string<ARM64BE>(u32 r_type) {
 }
 
 template <>
-std::string rel_to_string<ARM32>(u32 r_type) {
+std::string rel_to_string<ARM32LE>(u32 r_type) {
   switch (r_type) {
   CASE(R_ARM_NONE);
   CASE(R_ARM_PC24);
@@ -373,6 +373,11 @@ std::string rel_to_string<ARM32>(u32 r_type) {
   return unknown_type(r_type);
 }
 
+template <>
+std::string rel_to_string<ARM32BE>(u32 r_type) {
+  return rel_to_string<ARM32LE>(r_type);
+}
+
 template <>
 std::string rel_to_string<RV64LE>(u32 r_type) {
   switch (r_type) {

src/elf.h

@@ -13,7 +13,8 @@ struct X86_64;
 struct I386;
 struct ARM64LE;
 struct ARM64BE;
-struct ARM32;
+struct ARM32LE;
+struct ARM32BE;
 struct RV64LE;
 struct RV64BE;
 struct RV32LE;
@@ -357,9 +358,10 @@ enum : u32 {
 };
 
 enum : u32 {
-  EF_ARM_ABI_FLOAT_SOFT = 0x00000200,
-  EF_ARM_ABI_FLOAT_HARD = 0x00000400,
-  EF_ARM_EABI_VER5 = 0x05000000,
+  EF_ARM_ABI_FLOAT_SOFT = 0x200,
+  EF_ARM_ABI_FLOAT_HARD = 0x400,
+  EF_ARM_BE8 = 0x800000,
+  EF_ARM_EABI_VER5 = 0x5000000,
 };
 
 enum : u32 {
@@ -1810,7 +1812,8 @@ 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_arm64le = std::same_as<E, ARM64LE>;
 template <typename E> concept is_arm64be = std::same_as<E, ARM64BE>;
-template <typename E> concept is_arm32 = std::same_as<E, ARM32>;
+template <typename E> concept is_arm32le = std::same_as<E, ARM32LE>;
+template <typename E> concept is_arm32be = std::same_as<E, ARM32BE>;
 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>;
@@ -1827,6 +1830,7 @@ 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> concept is_x86 = is_x86_64<E> || is_i386<E>;
+template <typename E> concept is_arm32 = is_arm32le<E> || is_arm32be<E>;
 template <typename E> concept is_arm64 = is_arm64le<E> || is_arm64be<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>;
@@ -1920,7 +1924,7 @@ struct ARM64BE : ARM64LE {
   static constexpr bool is_le = false;
 };
 
-struct ARM32 {
+struct ARM32LE {
   static constexpr std::string_view name = "arm32";
   static constexpr bool is_64 = false;
   static constexpr bool is_le = true;
@@ -1950,6 +1954,11 @@ struct ARM32 {
   };
 };
 
+struct ARM32BE : ARM32LE {
+  static constexpr std::string_view name = "arm32be";
+  static constexpr bool is_le = false;
+};
+
 struct RV64LE {
   static constexpr std::string_view name = "riscv64";
   static constexpr bool is_64 = true;

src/filetype.cc

@@ -139,7 +139,7 @@ static std::string_view get_elf_type(u8 *buf) {
   case EM_X86_64:
     return X86_64::name;
   case EM_ARM:
-    return ARM32::name;
+    return is_le ? ARM32LE::name : ARM32BE::name;
   case EM_AARCH64:
     return is_le ? ARM64LE::name : ARM64BE::name;
   case EM_RISCV:

src/input-files.cc

@@ -999,6 +999,7 @@ void ObjectFile<E>::resolve_symbols(Context<E> &ctx) {
     }
 
     std::scoped_lock lock(sym.mu);
+
     if (get_rank(this, esym, !this->is_reachable) < get_rank(sym)) {
       sym.file = this;
       sym.set_input_section(isec);
@@ -1006,6 +1007,7 @@ 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_versioned_default = false;
     }
   }
 }
@@ -1273,23 +1275,48 @@ void SharedFile<E>::parse(Context<E> &ctx) {
     if (vers.empty() || esyms[i].is_undef())
       ver = VER_NDX_GLOBAL;
     else
-      ver = (vers[i] & ~VERSYM_HIDDEN);
+      ver = vers[i] & ~VERSYM_HIDDEN;
 
     if (ver == VER_NDX_LOCAL)
       continue;
 
-    std::string_view name = this->symbol_strtab.data() + esyms[i].st_name;
-    bool is_default = vers.empty() || !(vers[i] & VERSYM_HIDDEN);
-
     this->elf_syms2.push_back(esyms[i]);
     this->versyms.push_back(ver);
 
-    if (is_default) {
-      this->symbols.push_back(get_symbol(ctx, name));
-    } else {
-      std::string_view mangled_name = save_string(
+    std::string_view name = this->symbol_strtab.data() + esyms[i].st_name;
+
+    auto get_versioned_sym = [&] {
+      std::string_view key = save_string(
         ctx, std::string(name) + "@" + std::string(version_strings[ver]));
-      this->symbols.push_back(get_symbol(ctx, mangled_name, name));
+      return get_symbol(ctx, key, name);
+    };
+
+    // Symbol resolution involving symbol versioning is tricky because one
+    // symbol can be resolved with two different identifiers. Among
+    // symbols with the same name but different versions, one of them is
+    // always marked as the "default" one. This symbol is often denoted
+    // with two atsigns as `foo@@VERSION` and can be referred to either
+    // as `foo` or `foo@VERSION`. No other symbols have two names like that.
+    //
+    // On contrary, a versioned non-default symbol can be referred only
+    // with an explicit version suffix, e.g., `foo@VERSION`.
+    //
+    // Here is how we resolve versioned default symbols. We resolve `foo`
+    // and `foo@VERSION` as usual, but with information to forward
+    // references to `foo@VERSION` to `foo`. After name resolution, we
+    // visit all symbol references to redirect `foo@VERSION` to `foo`.
+    if (vers.empty() || ver == VER_NDX_GLOBAL) {
+      // Unversioned symbol
+      this->symbols.push_back(get_symbol(ctx, name));
+      this->symbols2.push_back(nullptr);
+    } else if (vers[i] & VERSYM_HIDDEN) {
+      // Versioned non-default symbol
+      this->symbols.push_back(get_versioned_sym());
+      this->symbols2.push_back(nullptr);
+    } else {
+      // Versioned default symbol
+      this->symbols.push_back(get_symbol(ctx, name));
+      this->symbols2.push_back(get_versioned_sym());
     }
   }
 
@@ -1396,6 +1423,22 @@ void SharedFile<E>::resolve_symbols(Context<E> &ctx) {
       sym.sym_idx = i;
       sym.ver_idx = versyms[i];
       sym.is_weak = true;
+      sym.is_versioned_default = false;
+    }
+
+    // A symbol with the default version is a special case because, unlike
+    // other symbols, the symbol can be referred by two names, `foo` and
+    // `foo@VERSION`. Here, we resolve `foo@VERSOIN` as a proxy of `foo`.
+    Symbol<E> *sym2 = this->symbols2[i];
+    if (sym2 && sym2 != &sym) {
+      std::scoped_lock lock2(sym2->mu);
+
+      if (get_rank(this, esym, false) < get_rank(*sym2)) {
+        sym2->file = this;
+        sym2->origin = (uintptr_t)&sym;
+        sym2->sym_idx = i;
+        sym2->is_versioned_default = true;
+      }
     }
   }
 }

src/lto-unix.cc

@@ -600,10 +600,10 @@ create_plugin_input_file(Context<E> &ctx, MappedFile *mf) {
 template <typename E>
 ObjectFile<E> *read_lto_object(Context<E> &ctx, MappedFile *mf) {
   if (ctx.arg.plugin.empty())
-    Fatal(ctx) << mf->name << ": don't know how to handle this LTO object file "
-               << "because no -plugin option was given. Please make sure you "
-               << "added -flto not only for creating object files but also for "
-               << "creating the final executable.";
+    Fatal(ctx) << mf->name << ": unable to handle this LTO object file because "
+               << "the -plugin option was not provided. Please make sure you "
+               << "added -flto not only when creating object files but also "
+               << "when linking the final executable.";
 
   load_lto_plugin(ctx);
 
@@ -670,8 +670,10 @@ ObjectFile<E> *read_lto_object(Context<E> &ctx, MappedFile *mf) {
     // section. We handle such symbols differently than comdat symbols in
     // a regular file because, unlike regular object files, IR files don't
     // have input sections.
-    if (psym.comdat_key)
-      obj->lto_comdat_groups[i + 1] = insert_comdat_group(ctx, psym.comdat_key);
+    if (psym.comdat_key) {
+      std::string_view key = save_string(ctx, psym.comdat_key);
+      obj->lto_comdat_groups[i + 1] = insert_comdat_group(ctx, key);
+    }
   }
 
   obj->symbol_strtab = save_string(ctx, strtab);

src/main.cc

@@ -44,8 +44,8 @@ static ObjectFile<E> *new_object_file(Context<E> &ctx, ReaderContext &rctx,
   ObjectFile<E> *file = new ObjectFile<E>(ctx, mf, archive_name);
   ctx.obj_pool.emplace_back(file);
   file->priority = ctx.file_priority++;
-  file->is_reachable =
-    !rctx.in_lib && (archive_name.empty() || rctx.whole_archive);
+  file->as_needed =
+    rctx.in_lib || (!archive_name.empty() && !rctx.whole_archive);
 
   rctx.tg->run([file, &ctx] { file->parse(ctx); });
   if (ctx.arg.trace)
@@ -65,8 +65,8 @@ static ObjectFile<E> *new_lto_obj(Context<E> &ctx, ReaderContext &rctx,
   ObjectFile<E> *file = read_lto_object(ctx, mf);
   file->priority = ctx.file_priority++;
   file->archive_name = archive_name;
-  file->is_reachable =
-    !rctx.in_lib && (archive_name.empty() || rctx.whole_archive);
+  file->as_needed =
+    rctx.in_lib || (!archive_name.empty() && !rctx.whole_archive);
 
   if (ctx.arg.trace)
     Out(ctx) << "trace: " << *file;
@@ -81,7 +81,7 @@ new_shared_file(Context<E> &ctx, ReaderContext &rctx, MappedFile *mf) {
   SharedFile<E> *file = new SharedFile<E>(ctx, mf);
   ctx.dso_pool.emplace_back(file);
   file->priority = ctx.file_priority++;
-  file->is_reachable = !rctx.as_needed;
+  file->as_needed = rctx.as_needed;
 
   rctx.tg->run([file, &ctx] { file->parse(ctx); });
   if (ctx.arg.trace)
@@ -496,7 +496,7 @@ int mold_main(int argc, char **argv) {
   // create as few segments as possible.
   sort_output_sections(ctx);
 
-  if (!ctx.arg.separate_debug_file.empty())
+  if (ctx.gnu_debuglink)
     separate_debug_sections(ctx);
 
   // Compute sizes of output sections while assigning offsets
@@ -536,13 +536,6 @@ int mold_main(int argc, char **argv) {
   if (ctx.arg.emit_relocs)
     create_reloc_sections(ctx);
 
-  // We've created range extension thunks with a pessimistive assumption
-  // that all out-of-section references are out of range. Now that we are
-  // able to assign addresses to all SHF_ALLOC output sections, we can
-  // eliminate excessive thunks.
-  if constexpr (needs_thunk<E>)
-    remove_redundant_thunks(ctx);
-
   // Compute .symtab and .strtab sizes for each file.
   if (!ctx.arg.strip_all)
     create_output_symtab(ctx);
@@ -562,6 +555,14 @@ int mold_main(int argc, char **argv) {
     filesize = set_osec_offsets(ctx);
   }
 
+  // We've created range extension thunks with a pessimistive assumption
+  // that all out-of-section references are out of range. Now that we know
+  // the addresses of all sections,, we can eliminate excessive thunks.
+  if constexpr (needs_thunk<E>) {
+    remove_redundant_thunks(ctx);
+    filesize = set_osec_offsets(ctx);
+  }
+
   if constexpr (is_arm32<E>) {
     if (ctx.extra.exidx) {
       ctx.extra.exidx->remove_duplicate_entries(ctx);
@@ -601,6 +602,9 @@ int mold_main(int argc, char **argv) {
   // Copy input sections to the output file and apply relocations.
   copy_chunks(ctx);
 
+  if constexpr (is_arm32be<E>)
+    arm32be_swap_bytes(ctx);
+
   if constexpr (is_x86_64<E>)
     if (ctx.arg.z_rewrite_endbr)
       rewrite_endbr(ctx);
@@ -612,7 +616,7 @@ int mold_main(int argc, char **argv) {
   // .gdb_index's contents cannot be constructed before applying
   // relocations to other debug sections. We have relocated debug
   // sections now, so write the .gdb_index section.
-  if (ctx.gdb_index && ctx.arg.separate_debug_file.empty())
+  if (ctx.gdb_index && !ctx.gnu_debuglink)
     write_gdb_index(ctx);
 
   // .note.gnu.build-id section contains a cryptographic hash of the
@@ -621,7 +625,7 @@ int mold_main(int argc, char **argv) {
   if (ctx.buildid)
     write_build_id(ctx);
 
-  if (!ctx.arg.separate_debug_file.empty())
+  if (ctx.gnu_debuglink)
     write_gnu_debuglink(ctx);
 
   t_copy.stop();
@@ -642,7 +646,7 @@ int mold_main(int argc, char **argv) {
   if (ctx.arg.print_map)
     print_map(ctx);
 
-  if (!ctx.arg.separate_debug_file.empty())
+  if (ctx.gnu_debuglink)
     write_separate_debug_file(ctx);
 
   // Show stats numbers

src/mold.h

@@ -262,9 +262,9 @@ struct FdeRecord {
 template <typename E>
 struct InputSectionExtras {};
 
-template <>
-struct InputSectionExtras<ARM32> {
-  InputSection<ARM32> *exidx = nullptr;
+template <is_arm32 E>
+struct InputSectionExtras<E> {
+  InputSection<E> *exidx = nullptr;
 };
 
 struct RelocDelta {
@@ -570,7 +570,7 @@ public:
   void populate_symtab(Context<E> &ctx) override;
 
   void scan_abs_relocations(Context<E> &ctx);
-  void create_range_extension_thunks(Context<E> &ctx, bool first_pass);
+  void create_range_extension_thunks(Context<E> &ctx);
 
   std::vector<InputSection<E> *> members;
   std::vector<std::unique_ptr<Thunk<E>>> thunks;
@@ -1337,6 +1337,7 @@ public:
   mark_live_objects(Context<E> &ctx,
                     std::function<void(InputFile<E> *)> feeder) = 0;
 
+  std::span<Symbol<E> *> get_local_syms();
   std::span<Symbol<E> *> get_global_syms();
   std::string_view get_source_name() const;
 
@@ -1353,6 +1354,7 @@ public:
   std::string_view shstrtab;
   std::string_view symbol_strtab;
 
+  bool as_needed = false;
   bool has_init_array = false;
   bool has_ctors = false;
 
@@ -1487,6 +1489,7 @@ public:
 
   std::string soname;
   std::vector<std::string_view> version_strings;
+  std::vector<Symbol<E> *> symbols2;
   std::vector<ElfSym<E>> elf_syms2;
 
 private:
@@ -1717,28 +1720,34 @@ void rewrite_endbr(Context<X86_64> &ctx);
 // arch-arm32.cc
 //
 
-class Arm32ExidxSection : public Chunk<ARM32> {
+template <is_arm32 E>
+class Arm32ExidxSection : public Chunk<E> {
 public:
-  Arm32ExidxSection(OutputSection<ARM32> &osec) : output_section(osec) {
+  Arm32ExidxSection(OutputSection<E> &osec) : output_section(osec) {
     this->name = ".ARM.exidx";
     this->shdr.sh_type = SHT_ARM_EXIDX;
     this->shdr.sh_flags = SHF_ALLOC;
     this->shdr.sh_addralign = 4;
   }
 
-  void compute_section_size(Context<ARM32> &ctx) override;
-  void update_shdr(Context<ARM32> &ctx) override;
-  void remove_duplicate_entries(Context<ARM32> &ctx);
-  void copy_buf(Context<ARM32> &ctx) override;
+  void compute_section_size(Context<E> &ctx) override;
+  void update_shdr(Context<E> &ctx) override;
+  void remove_duplicate_entries(Context<E> &ctx);
+  void copy_buf(Context<E> &ctx) override;
 
 private:
-  std::vector<u8> get_contents(Context<ARM32> &ctx);
+  std::vector<u8> get_contents(Context<E> &ctx);
 
-  OutputSection<ARM32> &output_section;
+  OutputSection<E> &output_section;
 };
 
-template <> u64 get_eflags(Context<ARM32> &ctx);
-void create_arm_exidx_section(Context<ARM32> &ctx);
+template <is_arm32 E>
+u64 get_eflags(Context<E> &ctx);
+
+template <is_arm32 E>
+void create_arm_exidx_section(Context<E> &ctx);
+
+void arm32be_swap_bytes(Context<ARM32BE> &ctx);
 
 //
 // arch-riscv.cc
@@ -1888,9 +1897,9 @@ struct ContextExtras<E> {
   NotePropertySection<E> *note_property = nullptr;
 };
 
-template <>
-struct ContextExtras<ARM32> {
-  Arm32ExidxSection *exidx = nullptr;
+template <is_arm32 E>
+struct ContextExtras<E> {
+  Arm32ExidxSection<E> *exidx = nullptr;
 };
 
 template <is_riscv E>
@@ -1961,6 +1970,7 @@ struct Context {
     bool allow_multiple_definition = false;
     bool allow_shlib_undefined = true;
     bool apply_dynamic_relocs = true;
+    bool be8 = false;
     bool color_diagnostics = false;
     bool default_symver = false;
     bool demangle = true;
@@ -2386,6 +2396,9 @@ public:
   bool is_traced : 1 = false;       // for --trace-symbol
   bool is_wrapped : 1 = false;      // for --wrap
 
+  // For symbols with default symbol version, e.g. foo@@VERSION.
+  bool is_versioned_default : 1 = false;
+
   // If a symbol can be resolved to a symbol in a different ELF file at
   // runtime, `is_imported` is true. If a symbol is a dynamic symbol and
   // can be used by other ELF file at runtime, `is_exported` is true.
@@ -2728,6 +2741,11 @@ inline std::string_view InputFile<E>::get_string(Context<E> &ctx, i64 idx) {
   return this->get_string(ctx, elf_sections[idx]);
 }
 
+template <typename E>
+inline std::span<Symbol<E> *> InputFile<E>::get_local_syms() {
+  return std::span<Symbol<E> *>(this->symbols).subspan(0, this->first_global);
+}
+
 template <typename E>
 inline std::span<Symbol<E> *> InputFile<E>::get_global_syms() {
   return std::span<Symbol<E> *>(this->symbols).subspan(this->first_global);
@@ -3015,12 +3033,16 @@ template <typename E>
 u64
 Symbol<E>::get_thunk_addr(Context<E> &ctx, u64 P) const requires needs_thunk<E> {
   std::span<u64> vec = ctx.symbol_aux[aux_idx].thunk_addrs;
+
   u64 lo = (P < branch_distance<E>) ? 0 : P - branch_distance<E>;
-  u64 val = *ranges::lower_bound(vec, lo);
-  i64 disp = val - P;
+  auto it = ranges::lower_bound(vec, lo);
+  if (it == vec.end())
+    Fatal(ctx) << "range extension thunk out of range: " << *this;
+
+  i64 disp = *it - P;
   if (disp < -branch_distance<E> || branch_distance<E> <= disp)
     Fatal(ctx) << "range extension thunk out of range: " << *this;
-  return val;
+  return *it;
 }
 
 template <typename E>
@@ -3030,6 +3052,11 @@ inline bool Symbol<E>::has_plt(Context<E> &ctx) const {
 
 template <typename E>
 inline bool Symbol<E>::is_absolute() const {
+  // An unresolved weak symbol acts as if it were an absolute address
+  // at address 0
+  if (is_remaining_undef_weak())
+    return true;
+
   return !is_imported && !get_frag() && !get_input_section() &&
          !get_output_section();
 }

src/output-chunks.cc

@@ -834,20 +834,6 @@ void DynamicSection<E>::copy_buf(Context<E> &ctx) {
   write_vector(ctx.buf + this->shdr.sh_offset, contents);
 }
 
-template <typename T>
-static std::vector<std::span<T>> split(std::vector<T> &input, i64 unit) {
-  std::span<T> span(input);
-  std::vector<std::span<T>> vec;
-
-  while (span.size() >= unit) {
-    vec.push_back(span.subspan(0, unit));
-    span = span.subspan(unit);
-  }
-  if (!span.empty())
-    vec.push_back(span);
-  return vec;
-}
-
 // Assign offsets to OutputSection members
 template <typename E>
 void OutputSection<E>::compute_section_size(Context<E> &ctx) {
@@ -1176,6 +1162,9 @@ void OutputSection<E>::populate_symtab(Context<E> &ctx) {
     u8 *strtab_base = ctx.buf + ctx.strtab->shdr.sh_offset;
     u8 *strtab = strtab_base + this->strtab_offset;
 
+    memset(esym, 0, this->num_local_symtab * sizeof(ElfSym<E>));
+    memset(strtab, 0, this->strtab_size);
+
     auto write_esym = [&](u64 addr, i64 st_name) {
       memset(esym, 0, sizeof(*esym));
       esym->st_name = st_name;
@@ -1315,18 +1304,16 @@ static std::vector<GotEntry<E>> get_got_entries(Context<E> &ctx) {
       }
     }
 
-    // If a symbol is imported, let the dynamic linker to resolve it.
     if (sym->is_imported) {
+      // If a symbol is imported, let the dynamic linker to resolve it.
       add({idx, 0, E::R_GLOB_DAT, sym});
-      continue;
-    }
-
-    // If we know an address at link-time, fill that GOT entry now.
-    // It may need a base relocation, though.
-    if (ctx.arg.pic && sym->is_relative())
+    } else if (ctx.arg.pic && sym->is_relative()) {
+      // We know the symbol's address, but it needs a base relocation.
       add({idx, sym->get_addr(ctx, NO_PLT), E::R_RELATIVE});
-    else
+    } else {
+      // We know the symbol's exact run-time address at link-time.
       add({idx, sym->get_addr(ctx, NO_PLT)});
+    }
   }
 
   // Create GOT entries for TLVs.
@@ -2634,6 +2621,14 @@ void VerdefSection<E>::construct(Context<E> &ctx) {
   if (ctx.arg.version_definitions.empty())
     return;
 
+  // Handle --default-symver
+  if (ctx.arg.default_symver)
+    for (Symbol<E> *sym : ctx.dynsym->symbols)
+      if (sym && !sym->file->is_dso)
+        if (u16 ver = sym->ver_idx;
+            ver == VER_NDX_GLOBAL || ver == VER_NDX_UNSPECIFIED)
+          sym->ver_idx = VER_NDX_LAST_RESERVED + 1;
+
   // Resize .gnu.version and write to it
   ctx.versym->contents.resize(ctx.dynsym->symbols.size(), VER_NDX_GLOBAL);
   ctx.versym->contents[0] = VER_NDX_LOCAL;

src/passes.cc

@@ -26,8 +26,10 @@ int redo_main(Context<E> &ctx, int argc, char **argv) {
     return mold_main<ARM64LE>(argc, argv);
   if (target == ARM64BE::name)
     return mold_main<ARM64BE>(argc, argv);
-  if (target == ARM32::name)
-    return mold_main<ARM32>(argc, argv);
+  if (target == ARM32LE::name)
+    return mold_main<ARM32LE>(argc, argv);
+  if (target == ARM32BE::name)
+    return mold_main<ARM32BE>(argc, argv);
   if (target == RV64LE::name)
     return mold_main<RV64LE>(argc, argv);
   if (target == RV64BE::name)
@@ -215,14 +217,14 @@ static void mark_live_objects(Context<E> &ctx) {
   }
 
   std::vector<InputFile<E> *> roots;
+  append(roots, ctx.objs);
+  append(roots, ctx.dsos);
 
-  for (InputFile<E> *file : ctx.objs)
-    if (file->is_reachable)
-      roots.push_back(file);
+  for (InputFile<E> *file : roots)
+    if (!file->as_needed)
+      file->is_reachable = true;
 
-  for (InputFile<E> *file : ctx.dsos)
-    if (file->is_reachable)
-      roots.push_back(file);
+  std::erase_if(roots, [](InputFile<E> *file) { return !file->is_reachable; });
 
   tbb::parallel_for_each(roots, [&](InputFile<E> *file,
                                     tbb::feeder<InputFile<E> *> &feeder) {
@@ -247,25 +249,14 @@ static void mark_live_objects(Context<E> &ctx) {
 // `VER1`. If it does, that's the symbol we are looking for.
 template <typename E>
 static void resolve_default_symver(Context<E> &ctx) {
-  tbb::parallel_for_each(ctx.objs, [&](ObjectFile<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 && esym.is_undef() &&
-          file->has_symver[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;
-      }
-    }
+  tbb::parallel_for_each(files, [](InputFile<E> *file) {
+    for (Symbol<E> *&sym : file->get_global_syms())
+      if (sym->is_versioned_default)
+        sym = (Symbol<E> *)sym->origin;
   });
 }
 
@@ -310,69 +301,68 @@ void resolve_symbols(Context<E> &ctx) {
     resolve_default_symver(ctx);
     mark_live_objects(ctx);
 
-    // Now that we know the exact set of input files that are to be
-    // included in the output file, we want to redo symbol resolution.
-    // This is because symbols defined by object files in archive files
-    // may have risen as a result of mark_live_objects().
-    //
-    // To redo symbol resolution, we want to clear the state first.
-    clear_symbols(ctx);
-
-    // COMDAT elimination needs to happen exactly here.
-    //
-    // It needs to be after archive extraction, otherwise we might
-    // assign COMDAT leader to an archive member that is not supposed to
-    // be extracted.
-    //
-    // It needs to happen before the final symbol resolution, otherwise
-    // we could eliminate a symbol that is already resolved to and cause
-    // dangling references.
-    tbb::parallel_for_each(ctx.objs, [](ObjectFile<E> *file) {
-      if (file->is_reachable) {
-        for (ComdatGroupRef<E> &ref : file->comdat_groups)
-          update_minimum(ref.group->owner, file->priority);
-        for (ComdatGroup *g : file->lto_comdat_groups)
-          if (g)
-            update_minimum(g->owner, file->priority);
-      }
-    });
-
-    tbb::parallel_for_each(ctx.objs, [](ObjectFile<E> *file) {
-      if (file->is_reachable)
-        for (ComdatGroupRef<E> &ref : file->comdat_groups)
-          if (ref.group->owner != file->priority)
-            for (u32 i : ref.members)
-              if (InputSection<E> *isec = file->sections[i].get())
-                isec->is_alive = false;
-    });
-
-    // Redo symbol resolution
-    tbb::parallel_for_each(files, [&](InputFile<E> *file) {
-      if (file->is_reachable)
-        file->resolve_symbols(ctx);
-    });
-
     // Symbols with hidden visibility need to be resolved within the
     // output file. If a hidden symbol was resolved to a DSO, we'll redo
     // symbol resolution from scratch with the flag to skip that symbol
     // next time. This should be rare.
-    std::atomic_bool flag = false;
-
+    std::atomic_bool redo = false;
     tbb::parallel_for_each(ctx.dsos, [&](SharedFile<E> *file) {
       if (file->is_reachable) {
         for (Symbol<E> *sym : file->symbols) {
           if (sym->file == file && sym->visibility == STV_HIDDEN) {
             sym->skip_dso = true;
-            flag = true;
+            redo = true;
           }
         }
       }
     });
 
-    if (!flag)
-      return;
+    if (!redo)
+      break;
     clear_symbols(ctx);
   }
+
+  // Now that we know the exact set of input files that are to be
+  // included in the output file, we want to redo symbol resolution.
+  // This is because symbols defined by object files in archive files
+  // may have risen as a result of mark_live_objects().
+  //
+  // To redo symbol resolution, we want to clear the state first.
+  clear_symbols(ctx);
+
+  // COMDAT elimination needs to happen exactly here.
+  //
+  // It needs to be after archive extraction, otherwise we might
+  // assign COMDAT leader to an archive member that is not supposed to
+  // be extracted.
+  //
+  // It needs to happen before the final symbol resolution, otherwise
+  // we could eliminate a symbol that is already resolved to and cause
+  // dangling references.
+  tbb::parallel_for_each(ctx.objs, [](ObjectFile<E> *file) {
+    if (file->is_reachable) {
+      for (ComdatGroupRef<E> &ref : file->comdat_groups)
+        update_minimum(ref.group->owner, file->priority);
+      for (ComdatGroup *g : file->lto_comdat_groups)
+        if (g)
+          update_minimum(g->owner, file->priority);
+    }
+  });
+
+  tbb::parallel_for_each(ctx.objs, [](ObjectFile<E> *file) {
+    if (file->is_reachable)
+      for (ComdatGroupRef<E> &ref : file->comdat_groups)
+        if (ref.group->owner != file->priority)
+          for (u32 i : ref.members)
+            if (InputSection<E> *isec = file->sections[i].get())
+              isec->is_alive = false;
+  });
+
+  // Redo symbol resolution
+  tbb::parallel_for_each(files, [&](InputFile<E> *file) {
+    if (file->is_reachable)
+      file->resolve_symbols(ctx);
+  });
 }
 
 // Do link-time optimization. We pass all IR object files to the compiler
@@ -1125,18 +1115,19 @@ void check_shlib_undefined(Context<E> &ctx) {
     return false;
   };
 
-  // Obtain a list of known shared library names.
+  // Skip test if we don't have a complete set of shared object files
+  // for the program, because if there's a missing .so, an undefined
+  // symbol might be defined by that library.
   std::unordered_set<std::string_view> sonames;
   for (std::unique_ptr<SharedFile<E>> &file : ctx.dso_pool)
     sonames.insert(file->soname);
 
-  tbb::parallel_for_each(ctx.dsos, [&](SharedFile<E> *file) {
-    // Skip the file if it depends on a file that we know nothing about.
-    // This is because missing symbols may be provided by that unknown file.
-    for (std::string_view needed : file->get_dt_needed(ctx))
-      if (!sonames.contains(needed))
+  for (SharedFile<E> *file : ctx.dsos)
+    for (std::string_view soname : file->get_dt_needed(ctx))
+      if (!sonames.contains(soname))
         return;
 
+  tbb::parallel_for_each(ctx.dsos, [&](SharedFile<E> *file) {
     // Check if all undefined symbols have been resolved.
     for (i64 i = 0; i < file->elf_syms.size(); i++) {
       const ElfSym<E> &esym = file->elf_syms[i];
@@ -1153,7 +1144,7 @@ void check_shlib_undefined(Context<E> &ctx) {
   // SharedFile<E>::mark_live_objects just for this pass. Therefore,
   // remove unneeded DSOs from the list now.
   for (SharedFile<E> *file : ctx.dsos)
-    file->is_reachable = false;
+    file->is_reachable = !file->as_needed;
 
   tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
     for (Symbol<E> *sym : file->get_global_syms())
@@ -1169,10 +1160,6 @@ template <typename E>
 void check_symbol_types(Context<E> &ctx) {
   Timer t(ctx, "check_symbol_types");
 
-  std::vector<InputFile<E> *> files;
-  append(files, ctx.objs);
-  append(files, ctx.dsos);
-
   auto canonicalize = [](u32 ty) -> u32 {
     if (ty == STT_GNU_IFUNC)
       return STT_FUNC;
@@ -1181,23 +1168,35 @@ void check_symbol_types(Context<E> &ctx) {
     return ty;
   };
 
-  tbb::parallel_for_each(files.begin(), files.end(), [&](InputFile<E> *file) {
+  auto check = [&](InputFile<E> &file, Symbol<E> &sym,
+                   const ElfSym<E> &esym1, const ElfSym<E> &esym2) {
+    if (sym.file && sym.file != &file &&
+        esym1.st_type != STT_NOTYPE && esym2.st_type != STT_NOTYPE &&
+        canonicalize(esym1.st_type) != canonicalize(esym2.st_type)) {
+      Warn(ctx) << "symbol type mismatch: " << sym << '\n'
+                << ">>> defined in " << *sym.file << " as "
+                << stt_to_string<E>(esym1.st_type) << '\n'
+                << ">>> defined in " << file << " as "
+                << stt_to_string<E>(esym2.st_type);
+    }
+  };
+
+  tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
+    for (i64 i = file->first_global; i < file->elf_syms.size(); i++)
+      if (Symbol<E> *sym = file->symbols[i];
+          sym->file && sym->file != file)
+        check(*file, *sym, sym->esym(), file->elf_syms[i]);
+  });
+
+  tbb::parallel_for_each(ctx.dsos, [&](SharedFile<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;
+      if (Symbol<E> *sym = file->symbols[i];
+          sym->file && sym->file != file)
+        check(*file, *sym, sym->esym(), file->elf_syms[i]);
 
-      const ElfSym<E> &esym1 = sym.esym();
-      const ElfSym<E> &esym2 = file->elf_syms[i];
-
-      if (esym1.st_type != STT_NOTYPE && esym2.st_type != STT_NOTYPE &&
-          canonicalize(esym1.st_type) != canonicalize(esym2.st_type)) {
-        Warn(ctx) << "symbol type mismatch: " << sym << '\n'
-                  << ">>> defined in " << *sym.file << " as "
-                  << stt_to_string<E>(esym1.st_type) << '\n'
-                  << ">>> defined in " << *file << " as "
-                  << stt_to_string<E>(esym2.st_type);
-      }
+      if (Symbol<E> *sym = file->symbols2[i])
+        if (sym->file && sym->file != file)
+          check(*file, *sym, sym->esym(), file->elf_syms[i]);
     }
   });
 }
@@ -1441,7 +1440,7 @@ void compute_section_sizes(Context<E> &ctx) {
 
     // create_range_extension_thunks is not thread-safe
     for (Chunk<E> *chunk : std::span(vec.begin(), tail.begin()))
-      chunk->to_osec()->create_range_extension_thunks(ctx, true);
+      chunk->to_osec()->create_range_extension_thunks(ctx);
 
     tbb::parallel_for_each(tail, [&](Chunk<E> *chunk) {
       chunk->compute_section_size(ctx);
@@ -2764,7 +2763,7 @@ i64 set_osec_offsets(Context<E> &ctx) {
     if (ctx.phdr) {
       i64 sz = ctx.phdr->shdr.sh_size;
       ctx.phdr->update_shdr(ctx);
-      if (sz != ctx.phdr->shdr.sh_size)
+      if (sz < ctx.phdr->shdr.sh_size)
         continue;
     }
 

src/shrink-sections.cc

@@ -122,10 +122,6 @@ i64 compute_distance(Context<E> &ctx, Symbol<E> &sym,
   if (sym.is_absolute())
     return INT64_MAX;
 
-  // Likewise, relocations against weak undefined symbols won't be relaxed.
-  if (sym.esym().is_undef_weak())
-    return INT64_MAX;
-
   // Compute a distance between the relocated place and the symbol.
   i64 S = sym.get_addr(ctx);
   i64 A = rel.r_addend;

src/thunks.cc

@@ -20,8 +20,8 @@
 // we don't need to try too hard to reduce thunk size to the absolute
 // minimum.
 
-#if MOLD_ARM32 || MOLD_ARM64LE || MOLD_ARM64BE || MOLD_PPC32 || \
-    MOLD_PPC64V1 || MOLD_PPC64V2
+#if MOLD_ARM32LE || MOLD_ARM32BE || MOLD_ARM64LE || MOLD_ARM64BE || \
+    MOLD_PPC32 || MOLD_PPC64V1 || MOLD_PPC64V2
 
 #include "mold.h"
 
@@ -45,68 +45,67 @@ static constexpr i64 max_thunk_size = 1024 * 1024;
 static constexpr i64 thunk_align = 16;
 
 template <typename E>
-static bool is_reachable(Context<E> &ctx, bool first_pass, InputSection<E> &isec,
-                         Symbol<E> &sym, const ElfRel<E> &rel) {
-  // If the target section is in the same output section but
-  // hasn't got any address yet, that's unreacahble.
-  InputSection<E> *isec2 = sym.get_input_section();
-  if (isec2 && isec.output_section == isec2->output_section &&
-      isec2->offset == -1)
+static bool
+requires_thunk(Context<E> &ctx, InputSection<E> &isec, const ElfRel<E> &rel,
+               bool first_pass) {
+  if (!is_func_call_rel(rel))
     return false;
+  Symbol<E> &sym = *isec.file.symbols[rel.r_sym];
 
-  // We don't know about the final file layout on the first pass, so
-  // we assume pessimistically that all out-of-section relocations are
-  // out-of-range. Excessive thunks will be removed on the second pass.
   if (first_pass) {
+    // On the first pass, we pessimistically assume that all out-of-section
+    // relocations are out of range.
+    InputSection<E> *isec2 = sym.get_input_section();
     if (!isec2 || isec.output_section != isec2->output_section)
-      return false;
+      return true;
+
+    // If the target section is in the same output section but
+    // hasn't got any address yet, that's unreacahble.
+    if (isec2->offset == -1)
+      return true;
 
     // Even if the target is the same section, we branch to its PLT
     // if it has one. So a symbol with a PLT is also considered an
     // out-of-section reference.
     if (sym.has_plt(ctx))
-      return false;
+      return true;
   }
 
+  // Thumb and ARM B instructions cannot be converted to BX, so we
+  // always have to make them jump to a thunk to switch processor mode
+  // even if their destinations are within their ranges.
+  if constexpr (is_arm32<E>)
+    if (bool is_thumb = sym.get_addr(ctx) & 1;
+        (rel.r_type == R_ARM_THM_JUMP24 && !is_thumb) ||
+        (rel.r_type == R_ARM_JUMP24 && is_thumb) ||
+        (rel.r_type == R_ARM_PLT32 && is_thumb))
+      return true;
+
+  // On PowerPC, all PLT calls go through range extension thunks.
+  if constexpr (is_ppc<E>)
+    if (sym.has_plt(ctx))
+      return true;
+
+  // 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 a thunk to recompute r2 for interworking.
+  if constexpr (is_ppc64v2<E>)
+    if ((rel.r_type == R_PPC64_REL24 && !sym.esym().ppc64_preserves_r2()) ||
+        (rel.r_type == R_PPC64_REL24_NOTOC && sym.esym().ppc64_uses_toc()))
+      return true;
+
   // Compute a distance between the relocated place and the symbol
   // and check if they are within reach.
   i64 S = sym.get_addr(ctx, NO_OPD);
   i64 A = get_addend(isec, rel);
   i64 P = isec.get_addr() + rel.r_offset;
   i64 val = S + A - P;
-  return -branch_distance<E> <= val && val < branch_distance<E>;
-}
-
-template <typename E>
-static bool needs_shim(Context<E> &ctx, Symbol<E> &sym, const ElfRel<E> &rel) {
-  // Thumb and ARM B instructions cannot be converted to BX, so we
-  // always have to make them jump to a thunk to switch processor mode
-  // even if their destinations are within their ranges.
-  if constexpr (is_arm32<E>) {
-    bool is_thumb = sym.get_addr(ctx) & 1;
-    return (rel.r_type == R_ARM_THM_JUMP24 && !is_thumb) ||
-           (rel.r_type == R_ARM_JUMP24 && is_thumb) ||
-           (rel.r_type == R_ARM_PLT32 && is_thumb);
-  }
-
-  // On PowerPC, all PLT calls go through range extension thunks.
-  if constexpr (is_ppc32<E> || is_ppc64v1<E>)
-    return sym.has_plt(ctx);
-
-  // 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>)
-    return sym.has_plt(ctx) ||
-           (rel.r_type == R_PPC64_REL24 && !sym.esym().ppc64_preserves_r2()) ||
-           (rel.r_type == R_PPC64_REL24_NOTOC && sym.esym().ppc64_uses_toc());
-  return false;
+  return val < -branch_distance<E> || branch_distance<E> <= val;
 }
 
 template <>
-void OutputSection<E>::create_range_extension_thunks(Context<E> &ctx,
-                                                     bool first_pass) {
+void OutputSection<E>::create_range_extension_thunks(Context<E> &ctx) {
   std::span<InputSection<E> *> m = members;
   if (m.empty())
     return;
@@ -181,33 +180,21 @@ void OutputSection<E>::create_range_extension_thunks(Context<E> &ctx,
     // Create a new thunk and place it at D.
     offset = align_to(offset, thunk_align);
     thunks.emplace_back(std::make_unique<Thunk<E>>(*this, offset));
+
     Thunk<E> &thunk = *thunks.back();
+    std::mutex mu;
 
     // Scan relocations between B and C to collect symbols that need
     // entries in the new thunk.
-    std::mutex mu;
-
     tbb::parallel_for(b, c, [&](i64 i) {
       InputSection<E> &isec = *m[i];
-
       for (const ElfRel<E> &rel : isec.get_rels(ctx)) {
-        if (!is_func_call_rel(rel))
-          continue;
-
-        // Skip if the symbol is undefined. apply_reloc() will report an error.
-        Symbol<E> &sym = *isec.file.symbols[rel.r_sym];
-        if (!sym.file)
-          continue;
-
-        // Skip if we can directly branch to the destination.
-        if (is_reachable(ctx, first_pass, isec, sym, rel) &&
-            !needs_shim(ctx, sym, rel))
-          continue;
-
-        // Add the symbol to the current thunk if it's not added already.
-        if (!sym.flags.test_and_set()) {
-          std::scoped_lock lock(mu);
-          thunk.symbols.push_back(&sym);
+        if (requires_thunk(ctx, isec, rel, true)) {
+          if (Symbol<E> &sym = *isec.file.symbols[rel.r_sym];
+              !sym.flags.test_and_set()) {
+            std::scoped_lock lock(mu);
+            thunk.symbols.push_back(&sym);
+          }
         }
       }
     });
@@ -226,6 +213,7 @@ void OutputSection<E>::create_range_extension_thunks(Context<E> &ctx,
     b = c;
   }
 
+  // Reset flags for future use
   for (; t < thunks.size(); t++)
     for (Symbol<E> *sym : thunks[t]->symbols)
       sym->flags = 0;
@@ -233,15 +221,75 @@ void OutputSection<E>::create_range_extension_thunks(Context<E> &ctx,
   this->shdr.sh_size = offset;
 }
 
+// create_range_extension_thunks() creates thunks with a pessimistic
+// assumption that all out-of-section references are out of range.
+// After computing output section addresses, we revisit all thunks to
+// remove unneeded symbols from them.
+//
+// We create more thunks than necessary and then eliminate some of
+// them later, instead of just creating thunks at this stage. This is
+// because we can safely shrink sections after assigning addresses to
+// them without worrying about making existing references to thunks go
+// out of range. On the other hand, if we insert thunks after
+// assigning addresses to sections, references to thunks could become
+// out of range due to the new extra gaps for thunks. Thus, the
+// creation of thunks is a two-pass process.
 template <>
 void remove_redundant_thunks(Context<E> &ctx) {
   Timer t(ctx, "remove_redundant_thunks");
-  set_osec_offsets(ctx);
 
+  // Gather output executable sections
+  std::vector<OutputSection<E> *> sections;
   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, false);
+        sections.push_back(osec);
+
+  // Mark all symbols that actually need range extension thunks
+  for (OutputSection<E> *osec : sections) {
+    tbb::parallel_for_each(osec->members, [&](InputSection<E> *isec) {
+      for (const ElfRel<E> &rel : isec->get_rels(ctx))
+        if (requires_thunk(ctx, *isec, rel, false))
+          isec->file.symbols[rel.r_sym]->flags.test_and_set();
+    });
+  }
+
+  // Remove symbols from thunks if they don't actually need range
+  // extension thunks
+  std::vector<Symbol<E> *> syms;
+
+  for (OutputSection<E> *osec : sections) {
+    for (std::unique_ptr<Thunk<E>> &thunk : osec->thunks) {
+      append(syms, thunk->symbols);
+      std::erase_if(thunk->symbols, [&](Symbol<E> *sym) { return !sym->flags; });
+    }
+  }
+
+  // Reset flags for future use
+  for (Symbol<E> *sym : syms)
+    sym->flags = 0;
+
+  // Recompute section sizes
+  tbb::parallel_for_each(sections, [&](OutputSection<E> *osec) {
+    std::span<InputSection<E> *> m = osec->members;
+    std::span<std::unique_ptr<Thunk<E>>> t = osec->thunks;
+    i64 offset = 0;
+
+    while (!m.empty() || !t.empty()) {
+      if (!m.empty() && (t.empty() || m[0]->offset < t[0]->offset)) {
+        offset = align_to(offset, 1 << m[0]->p2align);
+        m[0]->offset = offset;
+        offset += m[0]->sh_size;
+        m = m.subspan(1);
+      } else {
+        offset = align_to(offset, thunk_align);
+        t[0]->offset = offset;
+        offset += t[0]->size();
+        t = t.subspan(1);
+      }
+    }
+    osec->shdr.sh_size = offset;
+  });
 }
 
 // When applying relocations, we want to know the address in a reachable
@@ -259,7 +307,8 @@ void gather_thunk_addresses(Context<E> &ctx) {
   std::vector<OutputSection<E> *> sections;
   for (Chunk<E> *chunk : ctx.chunks)
     if (OutputSection<E> *osec = chunk->to_osec())
-      sections.push_back(osec);
+      if (osec->shdr.sh_flags & SHF_EXECINSTR)
+        sections.push_back(osec);
 
   ranges::stable_sort(sections, {}, [](OutputSection<E> *x) {
     return x->shdr.sh_addr;

test/CMakeLists.txt

@@ -30,8 +30,8 @@ endif()
 
 if(MOLD_ENABLE_QEMU_TESTS)
   list(APPEND QEMU_ARCHS
-    x86_64 i386 arm aarch64 aarch64_be ppc ppc64 ppc64le sparc64 sh4 sh4eb
-    s390x riscv64 riscv32 m68k loongarch64)
+    x86_64 i386 arm armeb aarch64 aarch64_be ppc ppc64 ppc64le sparc64
+    sh4 sh4eb s390x riscv64 riscv32 m68k loongarch64)
 
   LIST(APPEND TRIPLES
     x86_64-linux-gnu
@@ -39,6 +39,7 @@ if(MOLD_ENABLE_QEMU_TESTS)
     aarch64-linux-gnu
     aarch64_be-linux-gnu
     arm-linux-gnueabihf
+    armeb-linux-gnueabihf
     riscv64-linux-gnu
     powerpc-linux-gnu
     powerpc64-linux-gnu
@@ -121,6 +122,10 @@ if(${MACHINE} STREQUAL "arm" OR (HAS_qemu-arm AND HAS_arm-linux-gnueabihf-gcc))
   add_target(arm arm-linux-gnueabihf)
 endif()
 
+if(${MACHINE} STREQUAL "armeb" OR (HAS_qemu-armeb AND HAS_armeb-linux-gnueabihf-gcc))
+  add_target(armeb armeb-linux-gnueabihf)
+endif()
+
 if(${MACHINE} STREQUAL "riscv64" OR (HAS_qemu-riscv64 AND HAS_riscv64-linux-gnu-gcc))
   add_target(riscv64 riscv64-linux-gnu)
 endif()

test/arch-armeb-be32.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -o $t/a.o -c -xc -
+int main() {}
+EOF
+
+not $CC -B. -o $t/exe $t/a.o -mbe32 |& grep 'be32 is not supported'

test/common.inc

@@ -7,6 +7,7 @@ canonical_name() {
   case $1 in
   i?86) echo i686 ;;
   amd64) echo x86_64 ;;
+  armeb*) echo armeb ;;
   arm*) echo arm ;;
   powerpc) echo ppc ;;
   powerpc64) echo ppc64 ;;
@@ -64,7 +65,7 @@ t=$TESTDIR/$testname
 mkdir -p $t
 
 case $MACHINE in
-x86_64 | i686 | arm)
+x86_64 | i686 | arm | armeb)
   tlsdesc_opt=-mtls-dialect=gnu2 ;;
 aarch64 | riscv* | loongarch*)
   tlsdesc_opt=-mtls-dialect=desc ;;
@@ -101,7 +102,7 @@ supports_ifunc() {
 
 supports_tlsdesc() {
   # musl's tlsdesc on arm32 seems to be broken
-  [ $MACHINE = arm ] && is_musl && return 1
+  [[ $MACHINE = arm* ]] && is_musl && return 1
 
   # FreeBSD's loader doesn't seem to support TLSDESC relocs in an executable
   [ "$(uname)" = FreeBSD ] && return 1

test/compress-debug-sections-zstd.sh

@@ -2,7 +2,7 @@
 . $(dirname $0)/common.inc
 
 # arm-linux-gnueabihf-objcopy crashes on x86-64
-[ $MACHINE = arm ] && skip
+[[ $MACHINE = arm* ]] && skip
 [ $MACHINE = riscv32 ] && skip
 
 command -v zstdcat >& /dev/null || skip

test/crel.sh

@@ -2,7 +2,7 @@
 . $(dirname $0)/common.inc
 
 # Currently, CREL is not supported on REL-type targets
-[ $MACHINE = arm ] && skip
+[[ $MACHINE = arm* ]] && skip
 [ $MACHINE = i686 ] && skip
 
 [ "$CC" = cc ] || skip

test/default-symver-version-script.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -o $t/a.o -c -xc -
+void foo() {}
+EOF
+
+echo '{ global: foo; local: *; };' > $t/b.ver
+
+$CC -B. -o $t/c.so -shared $t/a.o -Wl,--default-symver -Wl,--version-script=$t/b.ver
+readelf --dyn-syms $t/c.so | grep -F ' foo@@c.so'

test/depaudit2.sh

@@ -22,8 +22,8 @@ void bar();
 int main() { bar(); }
 EOF
 
-$CC -B. -o $t/exe1 $t/c.o $t/b.so
+$CC -B. -o $t/exe1 $t/c.o $t/b.so -Wl,--allow-shlib-undefined
 readelf --dynamic $t/exe1 | grep 'Dependency audit library:..*/a.so'
 
-$CC -B. -o $t/exe2 $t/c.o $t/b.so -Wl,--depaudit=foo
+$CC -B. -o $t/exe2 $t/c.o $t/b.so -Wl,--depaudit=foo -Wl,--allow-shlib-undefined
 readelf --dynamic $t/exe2 | grep 'Dependency audit library:..*foo:.*/a.so'

test/dt-init.sh

@@ -6,7 +6,7 @@
 
 # musl libc does not support init/fini on ARM
 # https://github.com/rui314/mold/issues/951
-[ $MACHINE = arm -o $MACHINE = aarch64 ] && is_musl && skip
+[[ $MACHINE = arm* -o $MACHINE = aarch64 ]] && is_musl && skip
 
 cat <<EOF | $CC -c -fPIC -o $t/a.o -xc -
 void keep();

test/emit-relocs-cpp.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = arm ] && skip
+[[ $MACHINE = arm* ]] && skip
 
 cat <<EOF | $CXX -o $t/a.o -c -fPIC -xc++ -
 #include <cstdio>

test/large-max-page-size-strip.sh

@@ -2,7 +2,7 @@
 . $(dirname $0)/common.inc
 
 # ARM32's strip command crashes on the output of this test for some reason.
-[ $MACHINE = arm ] && skip
+[[ $MACHINE = arm* ]] && skip
 
 strip=$STRIP
 command -v $strip >& /dev/null || skip

test/lto-no-plugin.sh

@@ -9,4 +9,4 @@ echo 'int main() {}' | clang -c -o $t/a.o -xc -
 echo 'void foo() {}' | clang -c -o $t/b.o -xc - -flto
 
 not ./mold -o /dev/null $t/a.o $t/b.o |&
-  grep "b.o: don't know how to handle this LTO object file"
+  grep "b.o: unable to handle this LTO object file"

test/no-allow-shlib-undefined-circular.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -B. -shared -fPIC -o $t/libfoo.so -xc -
+void foo() {}
+EOF
+
+cat <<EOF | $CC -B. -shared -fPIC -o $t/libbar.so -xc - -L$t -lfoo
+void foo();
+void bar() { foo(); }
+EOF
+
+cat <<EOF | $CC -B. -shared -fPIC -o $t/libfoo.so -xc - -L$t -lbar
+void baz();
+void foo() { baz(); }
+EOF
+
+cat <<EOF | $CC -c -o $t/a.o -c -xc -
+int bar();
+int main() { bar(); }
+EOF
+
+not $CC -B. -o $t/exe1 $t/a.o -Wl,--no-allow-shlib-undefined -L$t -lfoo -lbar |&
+  grep 'libfoo.so: --no-allow-shlib-undefined: undefined symbol: baz'

test/no-allow-shlib-undefined3.sh

@@ -0,0 +1,29 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -B. -shared -fPIC -o $t/libfoo.so -xc -
+void foo() {}
+EOF
+
+cat <<EOF | $CC -B. -shared -fPIC -o $t/libbar.so -xc - -L$t -lfoo
+void foo();
+void bar() { foo(); }
+EOF
+
+cat <<EOF | $CC -B. -shared -fPIC -o $t/libbaz.so -xc - -L$t -lbar
+void bar();
+void baz() { bar(); }
+EOF
+
+cat <<EOF | $CC -c -o $t/a.o -c -xc -
+int baz();
+int main() { baz(); }
+EOF
+
+mv $t/libfoo.so $t/libfoo.so.bak
+echo | $CC -B. -shared -fPIC -o $t/libfoo.so -xc -
+
+$CC -B. -o $t/exe1 $t/a.o -Wl,--no-allow-shlib-undefined -L$t -lbar -lbaz
+
+not $CC -B. -o $t/exe2 $t/a.o -Wl,--no-allow-shlib-undefined \
+  -L$t -lfoo -lbar -lbaz |& grep 'no-allow-shlib-undefined: undefined symbol: foo'

test/no-allow-shlib-undefined4.sh

@@ -0,0 +1,29 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -o $t/a.o -c -xc - -fPIC
+void f1() {}
+EOF
+
+cat <<EOF | $CC -o $t/b.o -c -xc - -fPIC
+void f1();
+void f2() { f1(); }
+EOF
+
+cat <<EOF | $CC -o $t/c.o -c -xc - -fPIC
+void f1();
+void f3() { f1(); }
+EOF
+
+$CC -B. -o $t/libfoo.so -shared $t/a.o
+$CC -B. -o $t/libbar.so -shared $t/b.o -L$t -lfoo
+$CC -B. -o $t/libbaz.so -shared $t/c.o
+
+
+cat <<EOF | $CC -o $t/d.o -c -xc - -fPIC
+void f2();
+void f3();
+int main() { f2(); f3(); }
+EOF
+
+$CC -B. -o $t/exe $t/d.o -L$t -lbar -lbaz

test/nocopyreloc.sh

@@ -3,10 +3,10 @@
 
 [ $MACHINE = i686 ] && skip
 [ $MACHINE = m68k ] && skip
-[ $MACHINE = arm ] && skip
 [ $MACHINE = ppc64 ] && skip
 [ $MACHINE = ppc64le ] && skip
 [ $MACHINE = sh4 ] && skip
+[[ $MACHINE = arm* ]] && skip
 [[ $MACHINE = loongarch* ]] && skip
 
 cat <<EOF | $CC -shared -o $t/a.so -xc -

test/rodata-name.sh

@@ -3,7 +3,7 @@
 
 # ARM assembler has a differnet grammar than the others.
 # Concretely speaking, ARM as uses "@" as a start of a comment.
-[ $MACHINE = arm ] && skip
+[[ $MACHINE = arm* ]] && skip
 
 # All data symbols need to be aligned to 2 byte boundaries on s390x,
 # so rodata.str1.1 in this file is invalid on s390x.

test/symbol-version5.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -o $t/a.o -c -xc - -fPIC
+void foo() {}
+EOF
+
+cat <<EOF | $CC -o $t/b.o -c -xc - -fPIC
+__asm__(".symver foo_,foo@@VERSION");
+void foo_() {}
+EOF
+
+cat <<EOF > $t/c.version
+VERSION { global: foo; local: *; };
+EOF
+
+$CC -B. -shared -o $t/libfoo.so $t/a.o
+$CC -B. -shared -o $t/libbar.so $t/b.o -Wl,--version-script=$t/c.version
+
+cat <<EOF | $CC -o $t/d.o -c -xc -
+void foo();
+int main() { foo(); }
+__asm__(".symver foo,foo@VERSION");
+EOF
+
+$CC -B. -o $t/exe $t/d.o -L$t -lfoo -lbar

test/sysroot-linker-script.sh

@@ -1,6 +1,7 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
+[ $MACHINE = armeb ] && skip
 [[ $MACHINE = loongarch* ]] && skip
 
 cat <<EOF | $CC -o $t/a.o -c -xc -

test/sysroot.sh

@@ -1,6 +1,7 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
+[ $MACHINE = armeb ] && skip
 [[ $MACHINE = loongarch* ]] && skip
 
 cat <<EOF | $CC -c -o $t/a.o -xc -

test/sysroot2.sh

@@ -1,6 +1,7 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
+[ $MACHINE = armeb ] && skip
 [[ $MACHINE = loongarch* ]] && skip
 
 mkdir -p $t/bin $t/sysroot/foo

test/warn-symbol-type.sh

@@ -9,5 +9,4 @@ int main() {
 }
 EOF
 
-$CC -B. -shared -o $t/a.so $t/a.o |&
-  grep 'warning: symbol type mismatch: times'
+$CC -B. -shared -o $t/b.so $t/a.o |& grep 'warning: symbol type mismatch: times'