Current Switch Override Implementation Strategy Doesn't Make Much Electrical Sense

[WhiteNinjaZ]

The newly implemented switch override feature for custom switch blocks currently works by overriding the switch type for specific edges while building the switchblock portion of the RR graph. This leads to the scenario pictured bellow (circled in red):

I understand that outgoing edges of a track should be able to support multiple different edge types (circled in blue) especially when modeling different architectures, but I don't see how allowing multiple different switch types on the receiving track makes much electrical sense, especially if one of the switch types being used happens to be a short. In hardware, if the sink of a connection has a delay-less switch I don't think its possible for another track to connect to that same sink using a different switch type (though I could be forgetting a particular case). Now we might want to keep this even if it doesn't make sense electrically so as to add different delays in the edges for connections between different wire types. As it stands I see three options:

1. Keep the current method for switch override and place the responsibility on the user to understand whether they want their design to make sense electrically or they want to experiment with different modeling methods.
2. Change the current implementation to override the arch_wire_switch attribute of the sink track. This method would force all edges going to a particular wire type to have the same switch.
3. Do (2) but only when a user specifies that the override switch they want to use is a short.

@vaughnbetz, @jgoeders would love to get your opinions on this.

[vaughnbetz]

I'd do 1 as it's the simplest, and easiest to document. Ease of documentation matters, since otherwise people can't figure the tool out. Also, I don't think this behaviour is all that strange; I list some reasonable use cases of the current behaviour below.

For non-short switches, having a different switch from each drive point to a wire is fine. If you are using tri-statable switches they can all be completely different electrically. Even if you are using non-tristatable muxes, you can use different switch types to have different delays for various switches (e.g. some can go into the second stage of the mux and others into the first stage; you could also have unequally sized transistors in the muxes). Commercial FPGAs do exactly this (some priority switches go into the second stage of a mux in Intel architectures for example. So preserving the ability to specify this is good.

For short switches, if you have an L-shaped wire you could still have some short switches between the L-segments, and some tri-state or mux or pass gate switches from other wires.

There may be specific cases of shorts and other types of switches that you consider electrically illegal; those could be checked in check_rr_graph and an error or warning given. But I'd be judicious with them; I'd only flag things that are truly illegal or extremely strange. I've found that things that seem suboptimal but are still electrically legal are in fact investigated by some for valid reasons.