The backtrace is now nested, and should be used together with the stack
pointer array to construct the full backtrace for each exception.
We now allocate exception objects in a stack, but their names are still
not allocated. This is fine for exceptions raised in the driver or artiq
code, but we will have to implement allocation for names of exceptions
raised in RPC calls. The compiler should also emit code to store the
exception names once they catch it, to prepare for later reraising.
This patches ports the LLVM libunwind newly added forced unwinding
function. This enables us to run forced unwinding to obtain correct
backtrace when uncaught exceptions occur.
This patch also changes the exception handling scheme from the standard
two-phase unwinding to single phase using forced unwinding. This brings
some performance improvement and prepared for later nested exception
support. For nested exceptions, we will have to record the backtrace
regardless if the exception is an uncaught exception, as there can be
another exception being thrown while executing the finally block for
caught exceptions, and we will lose the backtrace if we don't store it
earlier before running the cleanup pads.
Port of 4a6bea479a
Co-authored-by: Steve Fan <sf@m-labs.hk>
Reviewed-on: M-Labs/artiq-zynq#156
Co-authored-by: stevefan1999 <sf@m-labs.hk>
Co-committed-by: stevefan1999 <sf@m-labs.hk>
si5324 driver in runtime should be now equal in function to the one in artiq.
kasli-soc has no way of doing a hard reset on the peripheral, but zc706 does.
Reviewed-on: M-Labs/artiq-zynq#132
Co-authored-by: mwojcik <mw@m-labs.hk>
Co-committed-by: mwojcik <mw@m-labs.hk>
The previous method of taking the channel could cause deadlock, we now
use semaphore to signal if the channel is available instead of busy
polling the mutex.