Tuesday, 1 March 2016

[gccsdk] UnixLib and ARMv8

Hi folks,

As you'll probably have heard, Raspberry Pi have launched a new board based
on the ARMv8 Cortex-A53.

One of the main compatibility issues remaining is that anything built on
UnixLib is rather unlikely to work. I did a bit of investigating and found
the problem was with the use of SWP in __pthread_disable_ints in
libunixlib/pthread/_ints.s (although the lack of pushing a stack frame
means that blame will be attributed elsewhere in backtraces). There are a
couple of other instances of SWP elsewhere in UnixLib.

Now, we're now living in a world where there is no one simple code sequence
which can do atomic operations from USR mode across all CPUs from ARMv2(a)
to ARMv8. A crude solution would be to replace

LDR a1, .L0 @=__ul_global
PICEQ "LDR a1, [a2, a1]"
ADD a1, a1, #GBL_PTH_WORKSEMAPHORE
MOV a3, #1
SWP a2, a3, [a1]
@ From this point onwards we will not be interrupted by the callback
ADD a2, a2, #1
STR a2, [a1]
MOV pc, lr

with something like

LDR a1, .L0 @=__ul_global
PICEQ "LDR a1, [a2, a1]"
ADD a3, a1, #GBL_PTH_WORKSEMAPHORE
MOV a1, #0
SWI XOS_ReadPlatformFeatures
MOVVS a1, #0
TST a1, #1<<12 @ does the CPU support LDREX?
BNE 1f
@ SWP version
MOV a1, #1
SWP a2, a1, [a3]
@ From this point onwards we will not be interrupted by the callback
ADD a2, a2, #1
STR a2, [a3]
MOV pc, lr
1: @ LDREX version
LDREX a2, [a3]
ADD a2, a2, #1
STREX a1, a2, [a3]
TEQ a1, #1
BEQ 1b
MOV pc, lr

and something similar for the other instances of SWP. However, I can't help
thinking it would be neater to do the ReadPlatformFeatures once somewhere
very early on in library initialisation and store the result in library
static data somewhere. This is where I suspect my lack of knowledge of
UnixLib internals means that I expect someone more familiar with it could
knock up a solution far quicker - so, any takers?

Ben

_______________________________________________
GCCSDK mailing list gcc@gccsdk.riscos.info
Bugzilla: http://www.riscos.info/bugzilla/index.cgi
List Info: http://www.riscos.info/mailman/listinfo/gcc
Main Page: http://www.riscos.info/index.php/GCCSDK

No comments:

Post a Comment