Waveform driver tracing: a self-holding register stalls a capped backward walk
In a backward 'trace X / trace to root' walk over a waveform plus design model, a register that holds its own value (q <= en ? d : q, or a plain feedback flop) makes each step's driver the same signal one clock edge earlier. A walk capped at N steps then stops N cycles back and never reaches the real origin (e.g. reset). Detect the self-hold: when the next input to follow is the signal itself, or reaches it through a hold mux, don't step edge by edge. Do a backward edge search for the start of the current value run (the last change into this value) and jump there, labelling the step 'held since t0'. That keeps cost O(log changes) and gets the earliest origin. A related FSDB gotcha: dumps written by several $fsdbDumpvars calls can expose multiple top scopes with the same name, each a partial tree, so merge them by name (recursively for children) before showing a hierarchy.
Correction from implementing this: jump to the start of the X RUN, not the start of the current value run. A signature/LFSR-style register (sig <= rot(sig) ^ in) changes value every cycle while its X moves around, so 'last change into this value' is still one cycle back. Search for the last change before which the value had no x/z bits. If the waveform backend supports derived signals, derive a 1-bit 'any bit unknown' signal ((^s) === 1'bx, since an xor-reduce is x when any bit is x/z) and take its last rising edge at or before t: one search. Otherwise walk changes back with a cap, and mark a capped jump 'held since t or earlier' so it can jump again. Two more cases worth handling: the hold that loops through a combinational next-value signal (q <= nx; nx = en ? f(q) : q) revisits nx, so apply the jump to any input already on the chain, not only a direct self-input. And an x-run start of 0 means uninitialised storage, not a loop.