build system: stop .eh_frame generation

From Rich Felker:

By default, modern GCC generates DWARF2 debug/unwind tables in the
.eh_frame section of the object files/binaries. This adds significant
bloat (as much as 15%) to the size of the busybox binary, including
the portion mapped/loaded into memory at runtime (possibly a big issue
for NOMMU targets), and the section is not strippable with the strip
command due to being part of the loaded program text.

I've since done some further checking - both testing and asking the
GCC developers about it - and it seems the solution is to add to the
CFLAGS -fno-unwind-tables and -fno-asynchronous-unwind-tables. If
debugging is disabled, this will prevent GCC from outputting DWARF2
tables entirely. But since busybox builds with -g by default, the
interesting case is what happens then. I originally thought these
options would break debugging, but they don't; instead, they tell GCC
to output the DWARF2 tables in the .debug_frame section instead of
the newish .eh_frame section (used for exception handling). With these
options added, busybox_unstripped is still fully debuggable, and the
final busybox binary loses the 15% bloat factor from the DWARF2
tables.

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>

diff --git a/Makefile.flags b/Makefile.flags
index f5f446598598b781c4c600d4151cba3aa7bca3c2..e77c0e5277ea81ea126160de3d0170ef5970e716 100644 (file)
--- a/Makefile.flags
+++ b/Makefile.flags
@@ -53,6 +53,9 @@ CFLAGS += $(call cc-option,-fno-builtin-strlen -finline-limit=0 -fomit-frame-poi
 CFLAGS += $(call cc-option,-fno-guess-branch-probability,)
 CFLAGS += $(call cc-option,-funsigned-char -static-libgcc,)
 CFLAGS += $(call cc-option,-falign-functions=1 -falign-jumps=1 -falign-labels=1 -falign-loops=1,)
+# Defeat .eh_frame bloat (gcc 4.6.3 x86-32 defconfig: 20% smaller busybox binary):
+CFLAGS += $(call cc-option,-fno-unwind-tables,)
+CFLAGS += $(call cc-option,-fno-asynchronous-unwind-tables,)
 
 # FIXME: These warnings are at least partially to be concerned about and should
 # be fixed..