Re: FW: New Mantis Issue #667 for V-1364

From: Steven Sharp (sharp@cadence.com)
Date: Wed Apr 20 2005 - 18:54:13 PDT

  • Next message: Steven Sharp: "Re: Proposal for ETF erratum 350"

    Steve Meyer writes:
    > It turns out that "port collapsing on input" is not sufficient to
    >model Verilog tran (switch level) gates because a relaxation algorithm
    >is needed.

    The data terminals of Verilog tran primitives are not inputs. They
    are inouts. Inouts are effectively required to be collapsible, so
    the extra rule Shalom suggested for switch inputs is not required for
    them. Inouts also cannot be connected to constants (or 2-state variables)
    so they don't enter into the time-zero constant propagation issue. The
    control input of a tranif can be connected to a constant. If so, the
    implicit continuous assignment mechanism would trigger it the same way
    as any primitive input.

    The rules for terminal connections for tran primitives don't require
    getting into the details of how tran networks are evaluated. Tran
    networks can pass through inout ports of modules and still evaluate
    properly, so it is sufficient to regard bidirectional tran terminals
    as inouts. It doesn't matter that tools have to do some rather complex
    things with the resulting networks to make them evaluate properly. That
    is a separable issue.

    > Related to this is the time 0 (we call it time minus infinity although
    >it really is the delta just before time 0 starts) gate
    >evaluation algorithm in Cver not only evaluates constants tied to
    >gates and module input ports, but propagates the changes through the
    >hierarchy, i.e. evaluate constants connected to gates (and continuous
    >assignments), if the gate output connects to a port, treat it as a
    >constant port connection and evaluation the strength preserving continuous
    >assignment, etc. If the port, connects to a gate elsewhere in the hierarchy,
    >treat it as a constant connection, etc.
    >
    > I think many old bipolar designs required this behavior. It evolved
    >since for Cver we needed to match behavior from bug reports.

    I can assure you that it is sufficient to treat the original constant
    connections as implicit continuous assignments, and evaluate all continuous
    assignments that reference constants in the first delta cycle. The rest
    of the propagation you describe will happen with normal event simulation.

    There may be performance advantages to propagating constants before
    simulation starts, because you can then take advantage of knowing that
    more values are constant. However, it isn't necessary, and can cause
    differences in simulation behavior by changing the order of evaluation.

    Steven Sharp
    sharp@cadence.com

    This archive was generated by hypermail 2.1.4 : Wed Apr 20 2005 - 18:31:01 PDT and
    sponsored by Boyd Technology, Inc.