On iCE40, use SB_IO PIN_TYPE to avoid unnecessary placement constraint on buffers

[whitequark]

Right now it appears that regardless of PIN_TYPE, the CLOCK_ENABLE and {INPUT,OUTPUT}_CLOCK in a pin pair with z=0,1 must match. This is mostly fine for clocks but gets extremely annoying for CLOCK_ENABLE, and worse, the constraint still has to be preserved if you use the pin as a comb input/output.

Nextpnr could take this into account and ignore the nets tied to CLOCK_ENABLE/*_CLOCK if the PIN_TYPE instructs it so.

[gatecat]

I believe this should be implemented via this function already - can you provide an example where this isn't working?

[whitequark]

@wanda-phi Do you recall where we hit this?

Actually, I remember now, it was in the Glasgow RGMII applet. I'd have to change the code a little to create a reproducer but it shouldn't be too hard.

[wanda-phi]

wasn't that hyperram, actually?

[whitequark]

Oh it would have been!