hal::k20: Add DSPI registers

[bgamari]

No description provided.

[bgamari]

This example brings up an interesting limitation of the current ioregs! implementation:

PUSHR is the transmit data register of the DSPI peripheral. It is backed by a FIFO and contains a data field (TXDATA) as well as a bitfield (PCS[5:0]) to specify which chip-select(s) should be targeted by the transaction. I currently represent this register with,

0x34 => reg32 pushr {
  0..15 => txdata, //= Transmitted data
  16..21 => pcs[6], //= Chip select state
}

Unfortunately, reads to this register fetch the values associated with the head of the FIFO. This means that queuing a new transaction would require that we first clear all of the pcs flags, then set the one we intend along with txdata. This is a rather tricky task with the current interface.

Another solution would be mark the pushr register as write-only, since write-only registers are no longer read from on update. This unfortunately means we loose the ability read the TX FIFO head (although I'm frankly not sure how valuable this capability is).

This does suggest, however, that the current scheme isn't quite expressive enough. Perhaps we want a new register type which has the same no-read-on-update semantics as wo but preserves the read accessors? What would such a thing be called?

Thoughts?

[farcaller]

Do I get it right that you want to write into txdata while preserving pcs value and not doing a read? Then a "cached" register might be a reasonable (although quite tricky to implement) option.

[bgamari]

The opposite actually, we don't want to preserve any of the former contents of the register across the update.

[bharrisau]

So there is no problem with the read? Just that when you write, you don't care about the previous value. So the current implementation will look like

pushr.set_txdata(data)
     .set_pcs(0, false)
     .set_pcs(1, true)
     .set_pcs(2, false)
     .set_pcs(3, false)
     .set_pcs(4, false)
     .set_pcs(5, false);

[bharrisau]

Perhaps add a clear() that initialised with a full mask and a value of 0?

pushr.clear()
     .set_txdata(data)
     .set_pcs(1, true);

Otherwise just live with the slightly verbose language above.

[bgamari]

@bharrisau correct.

clear is one solution, although it feels like a bit of a hack. I should probably just get over it though.

[farcaller]

clear() sounds like writing zeroes to it.

[bharrisau]

The other solution is a always-clear flag which indicates that there is write only, but still with reads.

[farcaller]

@bgamari can you please update the state of this or move out from in progress?

[bgamari]

I'll have a look tonight.