[RFC] "indirect" DCR access (40x, BookE)
ebs at ebshome.net
Fri Mar 12 12:48:00 EST 2004
As some people know using DCR is sometimes PITA ;-).
The reason is simple, DCR number is encoded in instruction itself, so
one cannot write code like:
for (i = 0; i < LAST_REG; ++i)
mtdcr(DCR_XXX_BASE + i, 0);
Because of this limitation we have to use a lot of ugly #ifdefs and/or
explicit switch/if statements when accessing DCR from device drivers
(see for example include/asm-ppc/ppc405-dma.h,
drivers/net/ibm_emac/ibm_ocp_mal.c (2.4 tree), etc)
There is a simple solution for this problem (I'm sure everybody
thought of this one, but just didn't have time to code it :).
This approach trades space (16K of code) for convenience.
Here is the short snippet which demonstrates the idea (full patch is
quite big and can be found at http://kernel.ebshome.net/dcr-2.6.diff):
#define DCR_ACCESS_PROLOG(table) \
rlwinm r3,r3,4,18,27; \
lis r5,table at h; \
ori r5,r5,table at l; \
add r3,r3,r5; \
mtctr r3; \
mfdcr r3,0; blr
mtdcr 0,r4; blr
mfdcr r3,1; blr
mtdcr 1,r4; blr
mfdcr r3,2; blr
mtdcr 2,r4; blr
mfdcr r3,3; blr
I grouped reading & writing of the same DCR together to make these
functions more cache friendly. I'm not sure about the names, though
This patch is against 2.6, I can easily make for 2.4 as well if there
is an interest.
Please note, I'm not proposing _removing_ current "mfdcr/mtdcr"
macros, they have their use. These "indirect" versions can be used in
some drivers to get more clean and readable code.
Comments, suggestions are welcome.
** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/
More information about the Linuxppc-embedded