On Tue, Apr 25, 2017 at 05:59:43PM +0800, Chen-Yu Tsai wrote:
On Tue, Apr 25, 2017 at 4:46 PM, Chen-Yu Tsai wens@csie.org wrote:
Hi,
On Tue, Apr 25, 2017 at 1:56 PM, Jagan Teki jagannadh.teki@gmail.com wrote:
Hi,
We're unable to build with buildman with gcc-4.9.x
arm: + Bananapi_M2_Ultra+arm-unknown-linux-gnueabi-ld.bfd: u-boot-spl section `.rodata' will not fit in region `.sram' +arm-unknown-linux-gnueabi-ld.bfd: region `.sram' overflowed by 716 bytes +make[2]: *** [spl/u-boot-spl] Error 1 +make[1]: *** [spl/u-boot-spl] Error 2
Any inputs?
CC-ing Andre as he's probably interested in this as well.
I normally build my stuff with Debian's 6.3 cross compiler. It seems that people don't always use the latest stuff, So I did a comparison between Linaro's 2016.02 5.3 and 2016.11 6.2.1 toolchains. I don't have a 4.x toolchain at the moment.
Long story short, the latter builds successfully, while the former fails. However, the object file sizes don't vary much between the two. I think the real difference is that the 6.2.1 toolchain comes with the GOLD linker. It probably does a much better job at pruning out dead code.
Appended is a list of object file sizes for the 5.3 toolchain, with total sizes for the 6.2.1 toolchain for comparison.
A few things pop up:
spl/arch/arm/lib/interrupts.o We are not using interrupts at all. Is this even necessary? IIRC there was a patch to get rid of this from Tom (CC-ed).
spl/common/dlmalloc.o Why do we still need this when we have SYS_MALLOC_SIMPLE enabled?
spl/common/image.o Almost half of this is description strings for U-boot legacy images, which probably isn't used in SPL? Could we somehow use macros to get rid of them?
Digging a bit deeper, I found that interrupts and dlmalloc were discarded by the linker. As far as C strings go, GCC 5.3 just dumps them all in the same section, which basically means all of them will get included, regardless of whether they are actually used or referenced. GCC 6 does a good job of splitting them by function, so all the C strings in one function get allocated to one section. If a function isn't used, it's text and rodata sections get dropped.
In a project I worked on in the past, a solution (or workaround) for this was to use named array (not pointer) symbols for each string. Each string would get tied to a symbol, and -fdata-sections would work properly with these.
IMHO upgrading to a newer toolchain is easier though.
Right and agreed. This is https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54303 which was resolved with gcc-6.x. That we keep hitting this form of problem (board X is OK with gcc-6.x or later, fails to link with older) is why I finally did https://patchwork.ozlabs.org/patch/751309/ and have moved travis-ci over to gcc-6.3.1 for ARM. I will be applying my patch for this release.
This in turn leads to needing to address lots of new warnings, and I need to find some time to add a real testcase to https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80428 as that is the bug I filed over the warnings about "tab_seq[]" and "erase_seq[]" in cli_readline.c