Build changes for ARM64 #534
Conversation
|
The x86 switch of PartitionAlloc to libc++ breaks x86 CI. It might just need to be installed. |
Should be easy enough to just switch back to libstdc++ on x86. |
|
Should this be changing our submodule to point at https://github.com/immunant/IA2-glibc? Testing this locally I see the glibc configure failing at its gcc version check step: |
fw-immunant
force-pushed
the
am/arm_build
branch
17 times, most recently
from
2ac6a35 to
f4eaf4d
Compare
|
Rebased one more time to incorporate the various AArch64 fixes that just landed in |
fw-immunant
force-pushed
the
am/arm_build
branch
from
f4eaf4d to
ee8e75a
Compare
This commit switches PartitionAlloc from GNU libstdc++ to LLVM libc++ since we need to use dependencies built with our clang fork on ARM. It does not do so on x86_64 because switching C++ stdlib requires both compile- (-stdlib=libc++) and link-time changes, and we don't expect the compile-time flags to be present when building for x86_64.
this was needed to ensure dependents of partition-alloc also link in LLVM libcxx
it's now correctly included in partition-alloc's target_link_libraries
we need to build glibc with our modified clang toolchain for ARM
rpath_link is inherited by static dependencies, so it should be used for libs and does nothing for executables. then pass only -rpath when linking the final executable.
this normalizes paths so that both rebuilt and prebuilt glibc leave binaries at /lib under the glibc sysroot. otherwise, we would need to make our LD_LIBRARY_PATH for qemu differ depending on whether glibc was prebuilt or not. that said, the glibc `make install` invocation does technically fail because it cannot not run the AArch64 ldconfig. this does not fail the build as a whole, so we leave it for now.
… glibc match also remove it from our LD_LIBRARY_PATH when invoking qemu as libc.so.6 and ld-linux-aarch64.so.1 now live in /lib (rather than /usr/lib) under the sysroot
we must link with the one built with our toolchain, whether prebuilt or from building glibc
fw-immunant
force-pushed
the
am/arm_build
branch
from
ee8e75a to
cb3089e
Compare
| @@ -203,11 +208,13 @@ jobs: | |||
| -DCMAKE_OBJCOPY=/usr/bin/llvm-objcopy-15 \ | |||
| -DCMAKE_C_FLAGS="$cross_flags" \ | |||
| -DCMAKE_CXX_FLAGS="$cross_flags" \ | |||
| -DCMAKE_EXE_LINKER_FLAGS="$cross_link_flags -lm -Wl,-rpath-link,${rtlibs_dir}/lib" \ | |||
| -DCMAKE_SHARED_LINKER_FLAGS="$cross_link_flags -Wl,-rpath,${rtlibs_dir}/lib" \ | |||
There was a problem hiding this comment.
Was switching -Wl,-rpath and -Wl,-rpath-link in these two lines intentional?
| @@ -12,7 +12,7 @@ else() | |||
| endif() | |||
| set(QEMU_COMMAND ${CMAKE_CROSSCOMPILING_EMULATOR} | |||
| "-one-insn-per-tb" | |||
| "-L" "${CMAKE_BINARY_DIR}/external/glibc/sysroot/usr/" | |||
There was a problem hiding this comment.
Does the sysroot not include the headers anymore? Why are we removing the usr suffix?
| @@ -55,15 +55,18 @@ target_link_options(libia2 INTERFACE | |||
| set_property(TARGET libia2 PROPERTY INTERFACE_LINK_DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/padding.ld ${CMAKE_CURRENT_BINARY_DIR}/dynsym.syms) | |||
|
|
|||
| if (LIBIA2_AARCH64) | |||
| set(SYSROOT_BIN_DIR ${CMAKE_BINARY_DIR}/external/glibc/sysroot) | |||
| set(LIBC_PATH ${SYSROOT_BIN_DIR}/lib/libc.so.6) | |||
| set(LD_LINUX_PATH ${SYSROOT_BIN_DIR}/lib/ld-linux-aarch64.so.1) | |||
| if (LIBIA2_REBUILD_GLIBC) | |||
| ExternalProject_Add(glibc | |||
| SOURCE_DIR ${CMAKE_SOURCE_DIR}/external/glibc | |||
| BINARY_DIR ${CMAKE_BINARY_DIR}/external/glibc | |||
| CONFIGURE_COMMAND ${CMAKE_COMMAND} -E env | |||
| # ffixed-x18 has to be set here instead of CFLAGS for it to apply to all files | |||
There was a problem hiding this comment.
nit: s/ffixed-x18/compiler flags in the comment
No description provided.