funs done:
MakeArgNode
MakeBasetypeArg

This is important since the additions of Rouland do not allow for eexternals to be nested without screwing the code generation up......
We can tackle this later!

actually not much was missing here. However, the treatment of T_hidden received a massive conceptual overhaul.
This was triggered by the observation that SACarg needs to be treated like a nested data structure.....
When ironing that out in compile.c we can equally well make sure we add proper support for nesting throughout....

Fix mt fold sbs

See merge request !129

streamlined the explanation; added a section on the implementation and extracted the two Handle-functions as helpes...

and put quite some more detail into the main comment of MTSPMDF.
This is now feature complete....

started re-writing SPMDF lifting to include a dec_rc on the neutral element after the lifted function

By setting the rc in the stack copy of the descriptor to 2 we are safe now. I also injected tracing info so that it easier to see what is going on just from the trace.
As a consequence of this, the MT version now leaks one copy of the neutral element! This needs another fix.
Finally, I added some comments in MTRMI to explain what exactly it does and to understand that traversal more quickly :-)

[hwloc] fix error on no hwloc.h header

See merge request !127

The declarations in cpubind.h still need the hwloc.h header file,
regardless if we are compiling with HWLOC support or not. This commit
fixes this.

Hotfix for CUDA profiling

See merge request !126

Fix cuda mech (for cudaManaged) memcpy ICMs

See merge request !109

This traversal was replaced by EMR-related traversals.

and add some better documentation

We now also add timers (using GPU timer) to measure the time for
certain events on the GPU (kernel launches, memcpys, allocs, etc.).

When moving to an ad-hoc macro cyclical traversal mechanism, the latest
counter value(s) were never stored, meaning that we never cycled to a
fix-point. This caused several transfers to be left in place which could
otherwise have been elided.

This bug also affected the algebraic wlfi traversal.

on some systems (change in linux kernel maybe?), filter over
HWLOC_OBJ_OS_DEVICE objects leads to seqfault. For HWLOC we do not rely
on any _system_ devices so leaving the filtering to NONE should be fine.

Also updated .gitignore

We also add the count of kernel calls (might be useful).

We can now measure the runtime (wall-clock time) of CUDA kernels using
sac2c's inbuilt profiling system. We use CUDA events systems (GPU/device
counters) to make the measurements.

Performing the measurement itself is fairly cheap, and has little effect
on runtime of main() function. However, we do perform some costly
summing up within libsac/runtime libraries after we've reached the end
of the program, which can take up several whole seconds.

At the moment, this CUDA timer feature only provides a total time for
the entire program run, not on a per-function basis.

NOTE: we store the start/stop values within a linked-list. We do this as
      we don't statically know how many times the kernel function will be
      called (in a conditional loop for example).  As such, if we have a
      programming launching thousands of kernels, this could consume a
      lot of memory. Clearly we could for instance at the point of the
      kernel launch make our measurement, and just store the elapsed
      time. This would require a sync on the device, which could destroy
      any performance gains from using asynchronise or managed backend.