Dear Albert ARIBAUD,
Am 30.11.2010 09:07, schrieb Albert ARIBAUD:
Hi Andreas,
Le 30/11/2010 08:06, Andreas Bießmann a écrit :
diff --git a/arch/arm/cpu/arm920t/start.S b/arch/arm/cpu/arm920t/start.S index 01edb9b..71de373 100644 --- a/arch/arm/cpu/arm920t/start.S +++ b/arch/arm/cpu/arm920t/start.S @@ -208,15 +208,16 @@ stack_setup: mov sp, r4
adr r0, _start
- cmp r0, r6
- beq clear_bss /* skip relocation */
- mov r1, r6
Why use r1?
Use a scratch register here cause stmia Rn! does increment Rn.
ldr r2, _TEXT_BASE ldr r3, _bss_start_ofs add r2, r0, r3 /* r2<- source end address */
cmp r0, r6
beq clear_bss
copy_loop: ldmia r0!, {r9-r10} /* copy from source address
[r0] */
- stmia r6!, {r9-r10} /* copy to target address [r1] */
- stmia r1!, {r9-r10} /* copy to target address [r1] */
Therefore usage of r6 here would destroy the saved 'addr of destination'. Cause that fact we have saved the 'addr of destination' @ beginning of relocate_code twice, once to r6 and once to r7. This is obviously not needed iv we change the register in copy_loop (which was already used in comment .. if you saw that on the end of the line).
I doubt if this a 'speed up' but we can have a cleaner interface. We have r4, r5, r6 as storage of input values to relocate_code. We do only use scratch registers for copy_loop, fixloop for relocation, clear_bss later on. We do decide if we need fixloop for relocation or if we can skip that and then setup the relevant scratch registers.
Well this is only an RFC. I found that plus the NULL pointer stuff worth to mention. It is not that important to get this special patch in cause the current implementation do also work.
regards
Andreas Bießmann