`compile_to_hardware` mangles sequences

[matthewware]

Is there any danger in doing a deep copy of seqs before we get here:

QGL/QGL/Compiler.py

Lines 327 to 331 in 446ff4f

	# all sequences should start with a WAIT for synchronization
	for seq in seqs:
	if not isinstance(seq[0], ControlFlow.Wait):
	logger.debug("Adding a WAIT - first sequence element was %s", seq[0])
	seq.insert(0, ControlFlow.Wait())

This mangles the seqs and even makes lines like

tomo_block = state_tomo([X(q1)], (q1,))
cals = create_cal_seqs((q1,), numRepeats=4)
seqs = tomo_block + cals

axis_descriptor = [{
    'name': 'segment num',
    'unit': None,
    'points': list(range(0,len(tomo_block))),
    'partition': 1
}]
axis_descriptor.append(cal_descriptor((q1,), 4))

metafile = compile_to_hardware(seqs, 'Tomo/Tomo', axis_descriptor=axis_descriptor, 
                               extra_meta = {'sequences':tomo_block})

crash hard since the metadata construction is happening after this initial loop. This seems to work for me with minor changes to QGL's Pulse object:

compiler_seqs = copy.deepcopy(seqs)
metafile = compile_to_hardware(compiler_seqs, 'Tomo/Tomo', axis_descriptor=axis_descriptor, 
                               extra_meta = {'sequences':str(tomo_block)})

Assuming we're not doing something like GST (a case we can work around) is there any reason not to do this? Seems this would be an issue anytime you needed seqs or adata about them after you'd run compile_to_hardware

[matthewware]

I basically just added a copy and deepcopy method

[dieris]

I think we can indeed copy the sequence (but do we need to define copy methods for lower layers? Note that tests are not passing now). This question was also brought up in #150

Note however that for this application, the sequence is already saved in Tomo-code.json before the WAIT is added

QGL/QGL/Compiler.py

Line 325 in 446ff4f

save_code(seqs, fileName + suffix)

[matthewware]

In my experience you need to have a __deepcopy__ method defined for all the objects to be deep copied. Here Pulse is the base object. The code above doesn't work without a deepcopy method. I guess I'm fine with not copying and loading the sequences from json as long as we document that process and put up some examples. My one concern is it starts to make creating axis descriptors comically obtuse. @grahamrow I think it's the extra meta data Auspex needs down the line that causes issues here. Do you have strong opinions on save-v-load?