[libc] Simplify integration tests by eliminating the artificial sysroot.

[sivachandra]

Summary

The test binaries are built like any other executable but with two
additional linker options -static and -nostdlib.

[jhuber6]

libc/cmake/modules/LLVMLibCTestRules.cmake

493

Should we use -nostdinc as well? Or do we still expect to pick up some system libraries.

[sivachandra]

libc/cmake/modules/LLVMLibCTestRules.cmake

493

Our libc does not [yet] provide the free standing headers like stddef.h and stdint.h. Using -nostdinc eliminates the paths to the compiler's freestanding headers also. So, we cannot [yet] use -nostdinc.

[jhuber6]

libc/cmake/modules/LLVMLibCTestRules.cmake

493

Alright, we can update this when libc is fully self supported.

[sivachandra]

• Improve the formatting of the code.
• Collect all deps in to a static archive which gets linked to the final binary.

[jhuber6]

libc/cmake/modules/LLVMLibCTestRules.cmake

451

Is it possible to make these dependencies transitive on the start implementation? Since presumably these are only required to implement the corresponding crt1.o and friends. For the GPU we wouldn't use any of these.

[sivachandra]

libc/cmake/modules/LLVMLibCTestRules.cmake

451

Yes, that can be done. I can do that separately.

libc/cmake/modules/LLVMLibCTestRules.cmake

function(add_integration_test test_name)
   # code also has to run init_array callbacks which potentially register
   # their own atexit callbacks. So, link in exit and atexit also with all
   # integration tests.
   list(
       APPEND fq_deps_list
       libc.src.__support.threads.thread
       libc.src.stdlib.atexit
       libc.src.stdlib.exit
       libc.src.unistd.environ
       libc.test.IntegrationTest.test)
   list(REMOVE_DUPLICATES fq_deps_list)
 
   # TODO: Instead of gathering internal object files from entrypoints,
   # collect the object files with public names of entrypoints.
   get_object_files_for_test(
       link_object_files skipped_entrypoints_list ${fq_deps_list})
-  if(skipped_entrypoints_list)
-    message(STATUS "Skipping ${fq_target_name} as it has skipped deps.")
-    return()
-  endif()
-
-  # Create a sysroot structure
-  set(sysroot ${CMAKE_CURRENT_BINARY_DIR}/${test_name}/sysroot)
-  file(MAKE_DIRECTORY ${sysroot})
-  file(MAKE_DIRECTORY ${sysroot}/include)
-  set(sysroot_lib ${sysroot}/lib)
-  file(MAKE_DIRECTORY ${sysroot_lib})
-  set(startup_object_file $<TARGET_OBJECTS:${INTEGRATION_TEST_STARTUP}>)
-  set(crti_object_file $<TARGET_OBJECTS:libc.startup.linux.crti>)
-  set(crtn_object_file $<TARGET_OBJECTS:libc.startup.linux.crtn>)
-  set(dummy_archive $<TARGET_PROPERTY:libc_integration_test_dummy,ARCHIVE_OUTPUT_DIRECTORY>/lib$<TARGET_PROPERTY:libc_integration_test_dummy,ARCHIVE_OUTPUT_NAME>.a)
-  # TODO: Copy the startup files to the correct target-triple directory instead
-  # of to a partly hard-coded directory.
-  set(startup_dst ${sysroot_lib}/${LIBC_TARGET_ARCHITECTURE}-linux-gnu)
-  add_custom_command(
-    OUTPUT ${startup_dst} ${sysroot}/lib/crti.o ${sysroot}/lib/crtn.o ${sysroot}/lib/libm.a ${sysroot}/lib/libc++.a
-    COMMAND cmake -E copy ${startup_object_file} ${startup_dst}/$<TARGET_PROPERTY:${INTEGRATION_TEST_STARTUP},OUTPUT_NAME>
-    COMMAND cmake -E copy ${crti_object_file} ${sysroot_lib}/$<TARGET_PROPERTY:libc.startup.linux.crti,OUTPUT_NAME>
-    COMMAND cmake -E copy ${crtn_object_file} ${sysroot_lib}/$<TARGET_PROPERTY:libc.startup.linux.crtn,OUTPUT_NAME>
-    # We copy the dummy archive as libm.a and libc++.a as the compiler drivers expect them.
-    COMMAND cmake -E copy ${dummy_archive} ${sysroot_lib}/libm.a
-    COMMAND cmake -E copy ${dummy_archive} ${sysroot_lib}/libc++.a
-    DEPENDS ${INTEGRATION_TEST_STARTUP} libc.startup.linux.crti libc.startup.linux.crtn libc_integration_test_dummy
-  )
-  add_custom_target(
-    ${fq_target_name}.__copy_startup__
-    DEPENDS ${startup_dst}
-  )
-
-  add_library(
-    ${fq_libc_target_name}
-    STATIC
-    ${link_object_files}
+  list(APPEND link_object_files
        # We add the memory functions objects explicitly. Note that we
        # are adding objects of the targets which contain the public
        # C symbols. This is because compiler codegen can emit calls to
        # the C memory functions.
        $<TARGET_OBJECTS:libc.src.string.bcmp>
        $<TARGET_OBJECTS:libc.src.string.bzero>
        $<TARGET_OBJECTS:libc.src.string.memcmp>
        $<TARGET_OBJECTS:libc.src.string.memcpy>
        $<TARGET_OBJECTS:libc.src.string.memmove>
-    $<TARGET_OBJECTS:libc.src.string.memset>
-  )
-  set_target_properties(${fq_libc_target_name} PROPERTIES ARCHIVE_OUTPUT_NAME c)
-  set_target_properties(${fq_libc_target_name} PROPERTIES ARCHIVE_OUTPUT_DIRECTORY ${sysroot_lib})
+       $<TARGET_OBJECTS:libc.src.string.memset>)
+
+  if(skipped_entrypoints_list)
+    message(STATUS "Skipping ${fq_target_name} as it has skipped deps.")
+    return()
+  endif()
 
   set(fq_build_target_name ${fq_target_name}.__build__)
   add_executable(
     ${fq_build_target_name}
     EXCLUDE_FROM_ALL
     ${INTEGRATION_TEST_SRCS}
     ${INTEGRATION_TEST_HDRS}
   )
   set_target_properties(${fq_build_target_name}
       PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
   target_include_directories(
     ${fq_build_target_name}
     PRIVATE
       ${LIBC_SOURCE_DIR}
       ${LIBC_BUILD_DIR}
       ${LIBC_BUILD_DIR}/include
   )
   target_compile_options(${fq_build_target_name} PRIVATE -ffreestanding ${INTEGRATION_TEST_COMPILE_OPTIONS})
-  # We set a number of link options to prevent picking up system libc binaries.
-  # Also, we restrict the integration tests to fully static executables. The
-  # rtlib is set to compiler-rt to make the compiler drivers pick up the compiler
-  # runtime binaries using full paths. Otherwise, files like crtbegin.o are passed
-  # as is (and not as paths like /usr/lib/.../crtbegin.o).
-  target_link_options(${fq_build_target_name} PRIVATE --sysroot=${sysroot} -static -stdlib=libc++ --rtlib=compiler-rt)
+  target_link_options(${fq_build_target_name} PRIVATE -nostdlib -static)
+  target_link_libraries(${fq_build_target_name} ${INTEGRATION_TEST_STARTUP} ${link_object_files})
   add_dependencies(${fq_build_target_name}
-                   ${fq_target_name}.__copy_startup__
-                   ${fq_libc_target_name}
                    libc.test.IntegrationTest.test
                    ${INTEGRATION_TEST_DEPENDS})
 
   add_custom_target(
     ${fq_target_name}
     COMMAND ${INTEGRATION_TEST_ENV} $<TARGET_FILE:${fq_build_target_name}> ${INTEGRATION_TEST_ARGS}
     COMMENT "Running integration test ${fq_target_name}"
   )
-
   add_dependencies(${INTEGRATION_TEST_SUITE} ${fq_target_name})
 endfunction(add_integration_test)

libc/startup/linux/aarch64/start.cpp

 #include <linux/auxvec.h>
 #include <linux/elf.h>
 #include <stdint.h>
 #include <sys/mman.h>
 #include <sys/syscall.h>
 #include <unistd.h>
 
 extern "C" int main(int, char **, char **);
+// The BFD linker requires a reference to __dso_handle to trigger creating
+// a symbol for it when -nostdlib is used..
+extern "C" void *__dso_handle = nullptr;
 
 // Source documentation:
 // https://github.com/ARM-software/abi-aa/tree/main/sysvabi64
 
 namespace __llvm_libc {
 
 #ifdef SYS_mmap2
 static constexpr long MMAP_SYSCALL_NUMBER = SYS_mmap2;

libc/startup/linux/x86_64/start.cpp

 #include <linux/auxvec.h>
 #include <linux/elf.h>
 #include <stdint.h>
 #include <sys/mman.h>
 #include <sys/syscall.h>
 #include <unistd.h>
 
 extern "C" int main(int, char **, char **);
+// The BFD linker requires a reference to __dso_handle to trigger creating
+// a symbol for it when -nostdlib is used..
+extern "C" void *__dso_handle = nullptr;
 
 namespace __llvm_libc {
 
 #ifdef SYS_mmap2
 static constexpr long mmapSyscallNumber = SYS_mmap2;
 #elif SYS_mmap
 static constexpr long mmapSyscallNumber = SYS_mmap;
 #else