LTO on libraries

Project Infrastructure LLVM Dev List Archives
1

Hello,

Is there a combination of the current tools would would allow me to apply LTO to a set of object files where are less than an entire application? Similar to “ld –r” for doing a partial link.
I assume it would start with “clang –flto –c {a,b,c}.c” but the a subsequent link step fails do to unresolved references of course.
Using llvm-link will linke the bitcode files but does not trigger compilation.

Thanks
david

2

You can llvm-link and then run clang. Note that the result of LTO in this case is not as powerful as when it is linker driver, since you can’t internalize.
You may also want to look at llvm-lto, but it can be annoying because it requires (AFAIK) an export list.

What is your use-case?

3

Thanks for the response.
To clarify in your suggestion, llvm-link will combine the modules but not run the optimization pass, that is still delayed until the final binary is built, correct?

My use case is apply LTO to roughly program subsets; sacrificing effectiveness to avoid scaling problems and to allow the artifacts to be reused like archives and cached like .o’s.
I need to trigger the optimizer on the intermediate rather than defer to final link for these goals.

—david

4

Thanks for the response.
To clarify in your suggestion, llvm-link will combine the modules but not run the optimization pass, that is still delayed until the final binary is built, correct?

Yes.
You can use clang to generate a .o (as a real binary), that you can then wrap in an archive library.

My use case is apply LTO to roughly program subsets; sacrificing effectiveness to avoid scaling problems and to allow the artifacts to be reused like archives and cached like .o’s.

You’re probably be interested in ThinLTO that is currently being implemented: scalability and incrementally are scheduled.

I need to trigger the optimizer on the intermediate rather than defer to final link for these goals.

Just run clang on the output of llvm-link :slight_smile:

Note: you’re not having the same linker semantic with an archive containing one .o compared to an archive with multiple .o. So you need to be careful.

5

Thanks

I had two small learnings.

  1. Don’t name the output of llvm-link as a “.o” (the subsequent clang ignores it)
  2. You need to have optimization flags like –O3 on the second clang, the on on the first is dropped.
    But this did as I had hoped:

bin/clang -flto -O3 -c {a,b,c}.o
bin/llvm-link -o abc.bc {a,b,c}.o
bin/clang -O3 -o abc.o -c abc.bc
bin/clang -o test main.c abc.o

6

Note that you may want to disable the optimization on the first run of clang:

bin/clang -flto -O3 -c {a,b,c}.o -mllvm -disable-llvm-optzns