Re: linux-xscale port, SYSFS device support

[Peter Milne <Peter.Milne@d-tacq.com>]

Hi Jeff

> Jeff Hill wrote:
>
> A patch (overlapping with the one provided below) for arm archetecture
> impure little endian was provided earlier by Larry Doolittle. It was
> integrated into EPICS R3.1.4.8. The preprocessor switch in that patch was
> _armv4l_. In contrast, the patch below is using just __arm__).
>
> http://www.d-tacq.com/swrel/EPICS/base-linux-xscale.patch
>
> However, TBOMK config file changes for XSCALE have *not* yet been
> integrated.
Thanks for the feedback.

_armv4l_ was most likely an older StrongArm processor. It will have the 
generic __arm__ symbol defined on build as well, and since it's likely 
that all the ARM processors will use the same floating point lib, the 
generic __arm__ should work for both and hopefully all cases, and least 
until someone tries to make a big-endian ARM port.

To cater for that eventuality, a full solution is suggested below, but 
please note I do _not_ want to make work for Janet by pushing any of 
this into the mainline, the patch is easy enough for me to maintain, a 
merge is really only justified if Larry or others are actively using 
similar code.

Cheers

Peter.

Problem:
Implement BASE port to allow generic ARM little endian with impure float 
as well as ARM big endian (assumed pure).

Solution:
Predefine CA_LITTLE_ENDIAN, CA_FLOAT_IEEE in the 
CONFIG.Common.linux-ARCH file, and to check this before setting the 
values according to arch.

This means a change to net_convert.h, but it can also lose the reference 
to __arm__, and the file does not have to keep
growing to cater for new processors. (maybe the H8 / PPC / uCDIMM people 
already do this?).

diff -X dontdiff -urN base-3.14.7.orig/src/ca/net_convert.h 
base-3.14.7/src/ca/net_convert.h
--- base-3.14.7.orig/src/ca/net_convert.h 2004-10-04 19:55:38.000000000 
+0100
+++ base-3.14.7/src/ca/net_convert.h 2005-11-09 09:11:32.000000000 +0000
@@ -27,7 +27,11 @@
* Here are the definitions for architecture dependent byte ordering
* and floating point format
*/
-#if defined (_M_IX86) || defined (_X86_) || defined (__i386__)
+#if defined (CA_FLOAT_IEEE) && defined (CA_LITTLE_ENDIAN)
+/*
+ * CONFIG has already defined the byte ordering, so we don't need to do 
it here
+ */
+#elif defined (_M_IX86) || defined (_X86_) || defined (__i386__)
# define CA_FLOAT_IEEE
# define CA_LITTLE_ENDIAN
#elif defined (VAX)

We still need the reference to __arm__ in the code to actually sort out 
the problem, __arm__ doesn't tell you anything about endian-ness, but 
since the point of use is guarded by "#ifdef CA_LITTLE_ENDIAN" the 
endian-ness is already known.

We have to add this line to CONFIG.Common.linux-xscale         
	ARCH_DEP_CFLAGS += -DCA_FLOAT_IEEE -DCA_LITTLE_ENDIAN

and similarly for other CONFIG.linux-ARM machines.

To be pedantic, I guess this file should be called  CONFIG.Common.linux-xscale-le to allow the possibility of a future 
linux-xscale-be option, but if the default arch "linux-xscale" is little-endian (I think most of them are, anyway, we got there first :-) ), then 
if a big-endian port is ever implemented in the future, it can be called "linux-xscale-be", the corresponding CONFIG file sets the 
CA_ defines appropriately and there is no clash.



--
Peter Milne Peter.Milne@d-tacq.com
D-TACQ Solutions Ltd www.d-tacq.com