On 24/11/15 21:39, Florian Weimer wrote: > On 11/24/2015 10:14 PM, Segher Boessenkool wrote: > >>> However, there is still a subtle difference for non-static objects. >>> Adding the initializer turns a tentative definition into a definite one, >>> and it seems that this can't go into .bss. >> >> extern int a; >> int a = 0; >> >> and >> >> int b; >> int b = 0; >> >> go to .bss just fine. Did you mean something different? > > This is what I meant. I still see a subtle difference in the generated > executable, but it is not what I thought. One should be using proper > tools. This is the diff after adding a couple of zero initializers: > > @@ -810,7 +810,7 @@ > 00000000003958f8 d > __elf_set___libc_thread_subfreeres_element___rpc_thread_destroy__ > 00000000003958e8 d > __elf_set___libc_thread_subfreeres_element_strerror_thread_freeres__ > 00000000001699a0 r empty_path_elem > -000000000039ea00 b _end > +000000000039e9f8 b _end > 00000000000fd6a0 T endaliasent > 00000000000dfaa0 T endfsent > 00000000000b3690 T endgrent > @@ -1287,14 +1287,14 @@ > 00000000000216c0 T __gconv_get_modules_db > 0000000000022d10 t __gconv_get_path > 0000000000028c00 t __gconv_load_cache > -000000000039e378 b __gconv_lock > +000000000039a8e8 b __gconv_lock > 0000000000028eb0 t __gconv_lookup_cache > -000000000039e388 b __gconv_max_path_elem_len > +000000000039e380 b __gconv_max_path_elem_len > 00000000001620a9 r gconv_module_ext > -000000000039e380 b __gconv_modules_db > +000000000039e378 b __gconv_modules_db > 0000000000020f20 t __gconv_open > -000000000039e398 b __gconv_path_elem > -000000000039e390 b __gconv_path_envvar > +000000000039e390 b __gconv_path_elem > +000000000039e388 b __gconv_path_envvar > 00000000000230f0 t __gconv_read_conf > 0000000000029310 t __gconv_release_cache > 00000000000293c0 t __gconv_release_shlib > @@ -3412,7 +3412,7 @@ > 0000000000397920 d __libc_utmp_file_functions > 000000000039a438 d __libc_utmp_file_name > 000000000039a430 d __libc_utmp_jump_table > -000000000039e9f8 b __libc_utmp_lock > +000000000039e164 b __libc_utmp_lock > 00000000003978e0 d __libc_utmp_unknown_functions > 000000000007a350 T __libc_valloc > 00000000001656d0 r __libc_version > @@ -4155,7 +4155,7 @@ > 000000000039b880 b old_realloc_hook > 000000000011b980 t __old_realpath > 000000000039bad0 b old_tz > -000000000039a8e8 b once > +000000000039a8ec b once > 000000000039d77c b once.11214 > 000000000039dae4 b once.13803 > 000000000039bf64 b once.14680 > > So __gconv_lock moves towards the end, somehow improving the overall > shared object size: > > 1659534 20312 17192 1697038 19e50e build-pristine/libc.so.6 > 1659534 20312 17184 1697030 19e506 build/libc.so.6 > > (In the diff above, - is the pristine variant (without initializer) and > + has the initializer.) > > Anyway, it means I can submit my patch because the GCC optimization does > what we need. > > Thanks, > Florian > Don't forget that by default GCC uses the common model for C, meaning that 'int x' is treated as a tentative definition. If you want the compiler to behave more like C++ in this regard, use -fno-common; but that might break some legacy programs. R.
