Hi Albert,
On 20.06.2013 18:42, Albert ARIBAUD wrote:
diff --git a/arch/arm/lib/crt0.S b/arch/arm/lib/crt0.S index a9657d1..b05f66a 100644 --- a/arch/arm/lib/crt0.S +++ b/arch/arm/lib/crt0.S @@ -85,7 +85,13 @@ ENTRY(_main) bic sp, sp, #7 /* 8-byte alignment for ABI compliance */ sub sp, #GD_SIZE /* allocate one GD above SP */ bic sp, sp, #7 /* 8-byte alignment for ABI compliance */ +#if !defined(CONFIG_SPL_BUILD) +/*
- SPL already has GD set to the correct location (in s_init), we mustn't
- move it around now since some data (clocks etc) is already present.
- */ mov r8, sp /* GD is above SP */
+#endif mov r0, #0 bl board_init_f
NAK in this form. I don't want gd to be set "somewhere in the code" depending on the actual target; I want it set in crt0.S, period.
I see there are several locations in ARM architecture or board code which set up GD themselves in the same manner as OMAP does. Luckily all these locations set it to the same value, the address of gdata.
The correct fix (read: the one I won't NAK) is thus to add a #else clause in the code above, in which r8 will be set to =gdata, and to remove the corresponding assignments in the various places where they reside.
Here's the problem. Setting r8 in _main is too late. As it has already been used (in the current implementation) to store some data (e.g. clocks for baudrate generation etc). Here the code from arch/arm/cpu/armv7/omap3/board.c:s_init():
#ifdef CONFIG_SPL_BUILD gd = &gdata;
preloader_console_init();
timer_init(); #endif
Note that this is done *before* _main() is called (we are talking about SPL for OMAP here). And it did cost me quite some time to find this problem, that r8 was re-configured in _main() and all the already set values disappeared again (no serial output etc).
Yes, this needs some cleanup/fixup. Unfortunately I won't find the time to look into such a cleanup in the next days. Perhaps somebody else might jump in...
Thanks, Stefan