$dffsr semantics as implemented differ from models and docs

[widlarizer]

Version

main

On which OS did this happen?

Linux

Reproduction Steps

Liberty file format allows the specification of the value internal variables attain when a DFFSR is being both set and reset with variables clear_preset_var1 and clear_preset_var2. Yosys internal cells canonically prioritize reset over set in all DFFSR and dfflibmap ignores the relevant liberty variables clear_preset_var1 and clear_preset_var2.

Some examples of the interpretation of these cells:

• async2sync - reset has priority
• verilog, proc_dff - no assumption about s&r behavior
• dfflibmap - assumes s&r behavior is free to be replaced with any other (disrespects clear_preset_var*)
• read_liberty - incorrectly always presents liberty dffsr as "reset has priority" (disrespects clear_preset_var*)
• filterlib -verilogsim - correctly models liberty cell behavior (respects clear_preset_var*)

Expected Behavior

Either:

• documentation declares dffsr cells to have an undefined output and state when set&reset
• or dfflibmap gets way more complex and worse at mapping because it only picks cells with matching set&reset behavior
• or we add more internal flop cell types or parameters to be able to specify this behavior

Actual Behavior

Some passes secretly assume a different definition of $dffsr and related cells than is documented