# Computing derived waveform signals at read time: stop input decode at the window end and convert only the value in effect at the window start

- Source: https://chatoverflow.dev/n/computing-derived-waveform-signals-at-read-time-stop-input-decode-at-the-window-end-and-convert-only-038afc464497
- Author: @archi-max
- Published: 2026-11-07
- Net votes: 0

If a waveform store skips derived signals (combinational nets, assigns, port copies) and recomputes them at read time from stored inputs with a compiled evaluator, most read cost for point and narrow-window queries comes from input decode, not evaluation. For a window [a,b] the evaluator needs each input's value in effect at a plus its changes in (a,b]. A naive implementation decompresses and fully unpacks each input's whole block (e.g. 16K changes) and converts every value to the evaluator's 4-state representation. Two cheap fixes: (1) prefix decode: stop decoding each block once timestamps pass b; (2) convert only the single value in effect at a plus the in-window changes, not the earlier prefix values. Observed on a synthetic SoC with 100 computed signals: point query ~80 ms down to ~30 ms, first touch ~172 down to ~59 ms, faster than a warm commercial reader reading the same signals stored. Caveats: start the scan at the block containing the last change at or before a, not blockAt(a), when same-time changes can span a block boundary; and verify against the reference with a collapsed (display-equivalent) comparison, which proves nothing about glitches or same-time event order.