lld runs LTO even if -fno-lto is passed

[simone-gaia]

Posted a SO question about this: https://stackoverflow.com/questions/72190379/lld-runs-lto-even-if-fno-lto-is-passed. Copying here the content:

I have a CMake project with several subprojects that create static libraries built with -flto=thin.

The project has a lot of tests that are linked against the aforementioned libraries. With LTO it takes a lot of time to build tests, therefore I have disabled LTO for tests using -fno-lto.

What I noticed though, is that lld performs LTO on tests even with -fno-lto. If I run the linker with --time-trace I can see that the majority of the time is spent on LTO.

My questions are:

1. Is this expected? If so I can assume that lld performs LTO whenever it finds the LTO info in the object it links.
2. If not, is there a way to disable this behavior? Adding -fno-lto to the compiler does not seem to work, and lld does not have a param to explicitly disable LTO.
3. If not, is this a bug? If this is a bug I will submit a repro.

This is how I handle lto in CMake:

# Enable Thin LTO only on non-test targets.
if(ENABLE_LTO)
  if (IS_TEST)
    target_compile_options(${TARGET} PRIVATE -fno-lto)
    # Probably pointless.
    target_link_options(${TARGET} PRIVATE -fno-lto)
  else()
    message(STATUS "ENABLE_LTO on target ${TARGET})")
    target_compile_options(${TARGET} PRIVATE -flto=thin)
    target_link_options(${TARGET} PRIVATE -flto=thin -Wl,--thinlto-cache-dir=${CMAKE_BINARY_DIR}/lto.cache)
  endif()
endif()

The command that builds the tests does not contain -flto=thin at all, it only contains -fno-lto

[simone-gaia]

From this SO answer https://stackoverflow.com/a/72249840/1214125 "If you compile the libraries with -flto then, at least for gcc, the object files will only contain the intermediate language and no binary code" . This could be the reason why I'm not able to exclude LTO from a subset of targets.

The questions now are:

1. Is this assumption correct?
2. If so, is there anything like -ffat-lto-objects in clang?

[MaskRay]

The assumption is correct. Clang doesn't implement GCC style -ffat-lto-objects which contains both contain both the intermediate language and the object code. It is a somewhat awkward trade-off which I am not sure any LTO user wants to support.

[MaskRay]

-ffat-lto-objects has been implemented (milestone: LLVM 18).

What I noticed though, is that lld performs LTO on tests even with -fno-lto.

This works as intended. -fno-lto instructs not to compile input files as bitcode files, but is unrelated to whether the linker should perform LTO.