Hi Tom:
On 2017年07月05日 09:56, Tom Rini wrote:
On Wed, Jul 05, 2017 at 09:08:47AM +0800, Andy Yan wrote:
Hi Tom:
On 2017年07月04日 21:32, Tom Rini wrote:
On Mon, Jul 03, 2017 at 04:02:59PM +0800, Andy Yan wrote:
Hi Philipp:
On 2017年06月30日 16:14, Dr. Philipp Tomsich wrote:
Andy,
On 30 Jun 2017, at 09:47, Andy Yan andy.yan@rock-chips.com wrote:
As Kever mentioned in [0], the RK3036 based boards could't bootup for a long time. After a git bisect, I found the RK3036 SPL code size has increased from patch [1] [2]. Before Tom's patch [1], the SPL size is 3160 bytes, but it becomes 4080 bytes after [1] applied. After a look at this patch, I realised I should disable SPL_USE_ARCH_MEMCPY/MEMSET, and the code size indeed come down after I disabled them. But I got a LD error after apply patch[2]: "undefined reference to memset", RK3036 SPL didn't use lib/string because of the sram space imitation. The compile succeed after CONFIG_SPL_LIBGENERIC_SUPPORT enabled, but the spl code size become 3248 bytes.
Additionally, Simon post patch [3] call printf to print a message before back to bootrom from spl, which make the spl code size increased to nearly 3.7 kb.
RK3036 SPL only has 4kb sram to use, the spl code will use 3.4 ~ 3.5 kb, the last 0.5kb are used for SP and GD, so there is no space for malloc.
gcc-6-arm-linux-gnueabi What version of GCC are you using? If your problem can also be solved by moving to GCC 6.3 (or newer) and the code-size improvements there, I’d rather just require a more recent GCC version.
I default use arm-linux-gnueabe-gcc v5.4. The current upstream kylin-rk3036_defconfig compiled by gcc-5.4is 4384 bytes, the size comes down to 3936 bytes if I use arm-linux-gnueabihf-gcc v6.3 from linaro. But this is still too large for rk3036.
Please note that (and U-Boot should be complaining at you) that with v2018.01 we'll be moving to gcc-6.x or later for ARM.
Disable SPL_USE_ARCH_MEMCPY/MEMSET will make the spl size comesdown to 3042 bytes by gcc v6.3. But I still need some hack: enable CONFIG_SPL_LIBGENERIC to get support for memset, masks Simon's print in bootrom.c, or the code size will become very large. Event though this hack make things work, we still lost a few hundreds bytes by function board_init_f_alloc_reserve, because platforms with very limit sram like rk3036 will return to bootrom after the dram initialized, they never use the malloc space. This few hundreds bytes is a large space for 4kb sarm, it's better to letf them for code or SP.
Since we're really size constrained here maybe it makes sense to move that print to a debug() ?
Yes, move it to debug() will mask the printf in normal. But even so, we still lost a few hundreds bytes in function
board_init_f_alloc_reserve for malloc. RK3036 will returned to bootrom immediately after the sdram initialization, the few hundreds space for malloc pool is never used in SPL stage.
Can you work the code paths out such that board_init_f_alloc_reserve would not be called then?
board_init_f_alloc_reserve is called from crt0.S, in not so easy to work the code path out. And board_init_f_alloc_reserve also reserve space for GD, this is we needed. So I want make this function as weak, then I can override it in my board spl state.