[cmake] Export `LLVM_ENABLE_LTO`.

[carlocab]

Summary

LTO builds differ in substantial ways from a standard build of LLVM. In
particular, if the build happens to include static archives, then those
archives will contain LLVM bitcode, which may require special handling.

I'm not aware of an easy way to detect whether LLVM has been built with
LTO. This would be made much simpler if this were exported by
LLVMConfig.

For an example where this might be helpful, see:

https://github.com/banach-space/llvm-tutor/commit/af9d1935027ebd61287930f16c0b6ca2fe0845f2

[MaskRay]

LGTM.

[carlocab]

Thanks @MaskRay. I don’t have commit access, can you land this patch for me? Please use “Carlo Cabrera [e-mail]” to commit the change.

[nikic]

No objection from my side, but I will caution about interpreting too much into the flag: At least Fedora (and I'd expect other distros as well), will build everything with LTO, but then convert bitcode to ELF in any shipped object files or static archives. So while the build is done with LTO, no LTO artifacts are exposed to the user. (Though I guess distros probably don't use LLVM_ENABLE_LTO either and instead pass global CFLAGS.)

[carlocab]

In D134936#3826244, @nikic wrote:

At least Fedora (and I'd expect other distros as well), will build everything with LTO, but then convert bitcode to ELF in any shipped object files or static archives.

Ah, I didn't realise Fedora does this. Do you mind pointing me to where/how this is done? I had a look at llvm.spec for LLVM 15 but couldn't see it there.

[nikic]

In D134936#3826263, @carlocab wrote:

In D134936#3826244, @nikic wrote:

At least Fedora (and I'd expect other distros as well), will build everything with LTO, but then convert bitcode to ELF in any shipped object files or static archives.

Ah, I didn't realise Fedora does this. Do you mind pointing me to where/how this is done? I had a look at llvm.spec for LLVM 15 but couldn't see it there.

This is done as part of __spec_install_post by https://src.fedoraproject.org/rpms/redhat-rpm-config/blob/rawhide/f/brp-llvm-compile-lto-elf.

[Ericson2314]

Do we care about LLVM itself being built with LTO, or various runtime libs being built with LTO? LLVMConfig.cmake should *just* be about LLVM itself.

[Ericson2314]

If each project used a separate ${project_upper}_ENABLE_LTO I would be more comfortable with this.

[carlocab]

Yes, perhaps this isn’t as useful as I thought it would be.

It wouldn’t be too hard to add the analogous variable for the various subprojects, but that still doesn’t handle the case where -flto is set in C*FLAGS, or where the distributor might do some post-build processing.

llvm/cmake/modules/LLVMConfig.cmake.in

 set(LLVM_ENABLE_WARNINGS @LLVM_ENABLE_WARNINGS@)
 
 set(LLVM_ENABLE_EXPENSIVE_CHECKS @LLVM_ENABLE_EXPENSIVE_CHECKS@)
 
 set(LLVM_ENABLE_ASSERTIONS @LLVM_ENABLE_ASSERTIONS@)
 
 set(LLVM_ENABLE_EH @LLVM_ENABLE_EH@)
 
+set(LLVM_ENABLE_LTO @LLVM_ENABLE_LTO@)
+
 set(LLVM_ENABLE_FFI @LLVM_ENABLE_FFI@)
 if(LLVM_ENABLE_FFI)
   find_package(FFI)
 endif()
 
 set(LLVM_ENABLE_RTTI @LLVM_ENABLE_RTTI@)
 
 set(LLVM_ENABLE_TERMINFO @LLVM_ENABLE_TERMINFO@)