https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93401
--- Comment #3 from Jan Hubicka <hubicka at gcc dot gnu.org> --- Just to summarize discussions we had. I think we need to think through setup where there are different build, train and pgo-buid machines and the directories are not known in advance. There -ffile-prefix-map for similar purpose and we could add -fprofile-prefix-map. Or maybe only -fprofile-prefix since at least for uses I have in mind the second part of map is going to be always empty. I guess we could support the following simulated makefile run: instrumented build ================== cd <buildir> mkdir lib1 cd <lib1> gcc foo.c -o <buildir>/ibb1/foo.o -fprofile-prefix-map=<buildir>= -fprofile-generate=/ this will stet up binary to produce /lib1#foo.gcda on startup. At the instrumentation time we can training run ============ GCOV_PREFIX=<my_favorite_profile_directory_during_train> run_program then tar contents of my_favorite_profile_directory_during_train final build =========== untar profile data to <my_favorite_profile_directory_during_build> Possibly use gcov-tool merge to merge data from multiple machines. cd <buildir2> mkdir lib1 gcc foo.c -o <buildir2>/ibb1/foo.o -fprofile-prefix-map=<buildir2>= -fprofile-use=<my_favorite_profile_directory_during_build> But it looks bit convoluted (especially passing the phony / to -fprofile-generate), so easier solution would be welcome. In particular I am not sure how llvmprofdata solve this problem
