gcc loads register from flash with two 16Bit moves

arne Guest 2010-01-10 12:27 #1548552

Rate this post

•	▲ useful ▼ not useful

Hello,
I am using an LM3s9b90 Controller with cortex M3 Core.

I wrote a simple test which toggles a Port pin by accessing the 
appropriate Bitband-Alias.
In following example EL_PF_3 is an Bitband Alias to Port F bit 3.
(Address 0x424A7F8C ,1112178572 dec)

In the assembler listing, you can see that two 16Bit moves are used to 
load the address,
I don't understand why. I think, this could be done with one 32Bit move.

ssi.c ****   EL_PF_3 = 0;
 21538                  .loc 1 19 0
 21539 0000 47F68C73     movw  r3, #:lower16:1112178572
 21540 0004 C4F24A23     movt  r3, #:upper16:1112178572
 21541 0008 0021         movs  r1, #0
ssi.c ****   EL_PF_3 = 1;
 21542                  .loc 1 20 0
 21543 000a 0122         movs  r2, #1
 21544                  .loc 1 19 0
 21545 000c 1960         str  r1, [r3, #0]
 21546                  .loc 1 20 0
 21547 000e 1A60         str  r2, [r3, #0]


I am using arm-elf-gcc, gcc-Version 4.4.0 with following flags:
-O3 -g3 -Wall -L lib -mcpu=cortex-m3 -mthumb

Has someone an idea, what might be wrong?

Best regards
 Arne

Reply Quote Edit Move Thread sperren Anmeldepflicht aktivieren Anpinnen Delete topic Thread mit anderem zusammenführen Report post

Giovanni D. (gdisirio) 2010-01-17 09:00 #1557759

Rate this post

•	▲ useful ▼ not useful

I think it is because the flash prefetch buffers, fetching two 
consecutive instructions is faster than fetching an instruction and then 
a 32 bits value displaced in memory which would require refilling the 
buffer with the insertion of wait states.

Giovanni
---
ChibiOS/RT http://chibios.sourceforge.net

Reply Quote Edit Delete Report post

(prx) A. K. (prx) 2010-01-17 09:13 #1557763

Rate this post

•	▲ useful ▼ not useful

arne wrote:

> In the assembler listing, you can see that two 16Bit moves are used to
> load the address,

There are 2 alternative ways to load a 32bit constant. PC relative from 
a constant pool like the old ARM core had to do, and the way shown here.

The PC relative load needs 6 bytes total, but depending on the machine 
it could cost a lot of clock cycles due to slow flash memory (~6 cycles 
on a LPC1700 at max freq). Embedding the constant in the sequential 
instruction stream needs 8 bytes, but runs in 2 clock cycles as 
sequential operations are largely unaffected by flash speed for the 
reason mentioned previously.

In GCC the taken alternative depends on the optimization setting. With 
-Os you get the smaller code.