Subject: Re: [sv-ec] Re: $wait_all/any/...
From: Kevin Cameron x3251 (Kevin.Cameron@nsc.com)
Date: Fri Dec 20 2002 - 15:48:27 PST
> From: "Michael McNamara" <mac@verisity.com>
>
> While we are being creative here, why not:
>
> always @(a or b) ; // equivalnet to wait any
> always @(a and b); // equivalent to wait all
> always @(a then b); // equivalent to wait priority
I like that :-)
Should allow:
always @((a then b) or (c then d)) ...
> None of these need a new keyword
>
> The new ones are currently illegal verilog, so no current designs
> would be broken by this addition.
>
> [By the way, the previous two questions are a criteria that must be
> carefully considered before making any change to a live language, in
> my opinion].
>
> Of course, number two [ @ (a and b) ] can already be done for entities
> things other than events by:
>
> always @(a && b);
>
> As for events: this may be a suprise to some; but events do not
> happen silmultaneously! Events are triggered one by one, and
> things that are blocking on an event are released; and then
> other events may be triggered. Specifically:
>
> always @(a) -> e1;
> always @(a) -> e2;
>
> always @( e1 and e2 ) ; // This will never fire, even if it was
> // legal syntax, as while e1 and e2 will
> // both spike true in the cycles that 'a'
> // changes, the spikes will NOT overlap.
It would work with the new persistent events - I presume the compiler would
complain if you used an old-style event with "and".
IMO there is insufficient syntactic difference between "event" and "event
bit" (particularly when things can be buried in typedefs) to make code
easily readable, which is why I prefer "<event>.active":
always @( e1.active and e2.active ) ; // fires for code above
That would work with old stlye events as well as new.
> Of course, number three [@ (a then b ) ] can already be done today by:
>
> always begin
> @(a);
> @(b);
> end
>
> So again I return to what is the value of adding these constructs?
>
> -> They do not give any more power to Verilog's capability;
>
> -> They do not as Vassilous hopes replace four lines of 1364 Verilog
> with 1 line of System Verilog;
>
> -> and instead they add a redundant, slightly different way of doing
> something that is already possible in Verilog 1364-1995.
>
I think there is value in the functionality (though it's not my proposal
so I have no examples), I just think the syntax and semantics could be
better.
Kev.
> Stuart Sutherland writes:
> > Vassilios makes a good point (in a separate e-mail) that we need to be
> > careful about sacrificing language clarity just to avoid adding
> > keywords. Using operators in an event sensitivity list instead of using
> > keywords would make the constructs more obscure, and therefore more error
> > prone.
> >
> > I propose that we narrow this discussion down to two options:
> >
> > Option One - multiple use of "any", "all" and "priority" keywords:
> > wait any
> > wait all
> > wait priority (instead of "wait order")
> > join none
> > join any
> > join all
> > join priority
> >
> > Note that "priority" is already reserved in SystemVerilog,
> > and with a similar meaning of prioritized order. Using
> > "priority" avoids having to reserve "order" as a keyword.
> >
> > Option Two - add unique, non-common keywords:
> > wait_any
> > wait_all
> > wait_priority
> > join_none
> > join_any
> > join_all
> > join_priority
> >
> > My personal preference is Option Two. Either option clearly documents
> > intent and is straight forward to use. The multiple use of "any", "all",
> > gives some justification to making them keywords, but I am still concerned
> > about reserving keywords that are common words. That will very likely
> > cause compatibility issues with existing Verilog models. Option Two
> > reserves more keywords than Option One, but the keywords are not likely to
> > have any impact on existing models.
> >
> > We should keep in mind that user's cannot not always easily change an
> > existing model to make it compatible with SystemVerilog. The model may be
> > an IP model or part of library which should only be changed by the model
> > provider. The model may be source code protected, making it impossible to
> > change without a key. The model may be locked from changes by version
> > control software, etc. SystemVerilog 3.0 already adds some keywords that
> > make this a serious issue. But that is not a good reason to exacerbate the
> > problem. Any time we can add the functionality and simplicity we desire
> > reserving non-common keywords, that should be the preferred
> > approach. There are times when a more common word is the best choice (e.g.
> > "class") because of its clarity for those coming from a C/C++
> > background. In the case of modifiers for "wait" and "join", however, we
> > can achieve clarity with non-common words.
> >
> > I would also like to point out that "join priority" or "wait priority" are
> > not orthogonal with "priority if" and "priority case". Should the modifier
> > be before or after the object? Maybe it's not a big deal, but it is different.
> >
> > Stu
> >
> >
> > At 11:54 PM 12/19/2002, Arturo Salz wrote:
> > >Mac,
> > >
> > >You make an excellent case for keeping these constructs in the language and
> > >not via system tasks. Also, the wait_order is clearly the most useful of
> > >the 3
> > >constructs. The other 2 wait-all/wait-any are there for orthogonality and
> > >completeness. If we can support all of these using existing operators that
> > >would be good.
> > >
> > > Arturo
> > >
> > >----- Original Message -----
> > >From: "Michael McNamara" <mac@verisity.com>
> > >To: "Arturo Salz" <Arturo.Salz@synopsys.COM>
> > >Cc: "Kevin Cameron" <Kevin.Cameron@nsc.com>; <sv-ec@eda.org>
> > >Sent: Thursday, December 19, 2002 3:43 PM
> > >Subject: Re: [sv-ec] Re: $wait_all/any/... (Forward of bounced emailfrom
> > >Arturo Salz)
> > >
> > >
> > >
> > >Introducing a system task ($wait_all) for a scheduling task, which
> > >replicates what can already be done in every verilog 1364-1995
> > >compliant simulator:
> > >
> > >always begin fork @a; @b; @c; join
> > >
> > >end
> > >
> > >seems to be
> > >
> > >1) unnecessarily introducing an additional way to do something that
> > > can already be done with even 1995 compliant simulators.
> > >
> > >2) slower (to implement this one needs to cross the pli boundry, set
> > > call backs on all of the operands to the task, and then somehow
> > > block further execution of the statement guarded by the $wait_all,
> > > yet allow execution of the rest of the simulation so that these
> > > events might occur)
> > >
> > >3) require semantic changes (how does a system task request the
> > > simulator to block execution of the guarded statement (other than by
> > > refusing to return), while also retruning control to the simulator
> > > so that the guard events may occur?
> > >
> > >4) can be overridden (users can define their own user task that
> > > overrides a system task, and perhaps already have a task called
> > > $wait_all).
> > >
> > >Arturo Salz writes:
> > > > Kev,
> > > >
> > > > I appreciate your point, but I miss where you are going with this.
> > > > The persistency property needs to be part of the declaration since
> > > > it's not possible to determine this from the usage (since events can
> > > > appear anywhere in a dynamic context). Otherwise, we run the
> > > > risk of forcing all events to implement persistency, something that
> > > > can have adverse performance consequences.
> > > >
> > > > Arturo
> > > >
> > > > ----- Original Message -----
> > > > From: "Kevin Cameron" <Kevin.Cameron@nsc.com>
> > > > Cc: <sv-ec@eda.org>
> > > > Sent: Thursday, December 19, 2002 2:45 PM
> > > > Subject: Re: [sv-ec] Re: $wait_all/any/... (Forward of bounced
> > > emailfrom Arturo Salz)
> > > >
> > > >
> > > >
> > > > "Arturo" wrote:
> > > >
> > > > > Stu,
> > > > >
> > > > > The difference between $wait_all(x,y,z) and @(x && y && z) is that =
> > > > > $wait_all actually works whereas the second form doesn't work. Most
> > > Verilog
> > > > > compilers will = flag that expression as an error, and even if they
> > > don't,
> > > > > it won't work because = the events will not trigger at the same
> > > time. The
> > > > > $wait_all with non-persistent events = is semantically more like this:
> > > > > fork
> > > > > @ x;
> > > > > @ y;
> > > > > @ z;
> > > > > join
> > > > >
> > > > > The second form, $wait_any is equivalent to @(x or b or z) with the =
> > > > > exception that both $wait_all and $wait_any will work with either
> > > persistent
> > > > > and = non-persistent events. =20
> > > >
> > > > I understand why @(x && y) doesn't work for non-persistent events, but
> > > I don't
> > > > see why it can't work with persistant events.
> > > >
> > > > Since we are moving to using <object>.<method> syntax for other things
> > > we could
> > > > add methods to events so that you wouldn't need to differentiate e.g.:
> > > >
> > > > event x,y;
> > > > always @(x.active && y.active) ...
> > > >
> > > > "active" would be true for any event for the simulation cycle in which
> > > the event occurs.
> > > > No need to add "event bit".
> > > >
> > > > There are lots of alternatives to using system tasks and keywords.
> > > >
> > > > Kev.
> > > >
> > > > > We've also been thinking about the dual usage of the all, any
> > > keywords and
> > > > > treat this as wait all( ... ) wait any ( ... ) and possibly wait
> > > seq( ... )
> > > > > for wait_order (seq is the keyword for defining sequences in sv-ac).
> > > >
> > > > > Regarding $wait_order, I think you all missed one important property (it
> > > > > may be that the document could stress this better). A $wait_order
> > > requires
> > > > > that the events occur in strict order, that is if ANY event is received
> > > > > out-of-order the call = fails. It is this distinction that makes
> > > wait_order
> > > > > special and not at all like:
> > > > > @d @e @f <statement>
> > > > >
> > > > > The assertions committee has syntax that allows easy specification
> > > of some
> > > > > sequences like the above, but they have no syntax for specifying
> > > this strict
> > > >
> > > > > ordering in a succinct way. It is a lot of code in both assertions and
> > > > > regular Verilog. That's why we proposed it. And, finally, wait_order is
> > > > > guaranteed to work with both persistent and non-persistent events by
> > > > > enforcing that only the first event in the sequence can be triggered at
> > > > > the time of the call.
> > > > >
> > > > > Arturo
> > > > >
> > > > > ----- Original Message -----=20
> > > > > From: Stuart Sutherland=20
> > > > > To: Kevin Cameron ; sv-ec@eda.org=20
> > > > > Sent: Wednesday, December 18, 2002 3:42 PM
> > > > > Subject: Re: [sv-ec] Re: $wait_all/any/...
> > > > >
> > > > > My thoughts are inserted below...
> > > > >
> > > > > Stu
> > > > >
> > > > > At 03:02 PM 12/18/2002, Kevin Cameron wrote:
> > > > >
> > > > > Link: Replace 12=20
> > > > >
> > > > > Having posted a list of keywords that shouldn't be, I have to say =
> > > > > these=20
> > > > > look like they should be keywords rather than system tasks (if
> > > they=20
> > > > > are actually needed).=20
> > > > >
> > > > > I agree with the use of keywords here, instead of system tasks
> > > that = can
> > > > > be redefined by the PLI. If we are stuck with the "any" and "all" =
> > > > > keywords from the "join any" and "join all", then no new additional =
> > > > > keywords are needed. One could do "wait any" and "wait all". I do
> > > not =
> > > > > like reserving "any" and "all" as keywords, but a dual usage might
> > > sway =
> > > > > me--maybe.
> > > > >
> > > > > What's the difference between:=20
> > > > >
> > > > > $wait_all(x,y,z)=20
> > > > >
> > > > > and=20
> > > > >
> > > > > @(x && y && z)=20
> > > > >
> > > > > ?=20
> > > > >
> > > > > At least one difference in Verilog is the ambiguity of whether to =
> > > > > trigger on a change on the operand or a change on the result. =
> > > > > SystemVerilog 3.0 adds the "changed" keyword, which would resolve that =
> > > > > ambiguity.
> > > > >
> > > > > I'm not clear if the $wait_and means at least two of the event types =
> > > > > must be true before the third one is triggered (which would require
> > > the =
> > > > > events have persistence) or that all three events must trigger, but
> > > it = can
> > > > > happen in any order (which would require the @ emulate a state =
> > > machine).
> > > > > If it is the former, then @(changed (x&&y&&z)) would do what = I need
> > > > > without $wait_all or a new keyword. If it is the latter, then =
> > > $wait_all
> > > > > or a new keyword is necessary. That's how I see it, at least.
> > > > >
> > > > > Or between=20
> > > > >
> > > > > $wait_any(a,b,c)=20
> > > > >
> > > > > and=20
> > > > >
> > > > > @(a or b or c)=20
> > > > >
> > > > > ?=20
> > > > >
> > > > > - Persistance is a property of the event not of the wait, so I
> > > don't =
> > > > > see why=20
> > > > > we need them.=20
> > > > >
> > > > > I agree. Verilog does this already, so neither $wait_any or a new =
> > > > > keyword is needed.
> > > > >
> > > > > $wait_order could be broken down into an order call:=20
> > > > >
> > > > > $wait_order(d,e,f)=20
> > > > >
> > > > > becomes:=20
> > > > >
> > > > > @(ordered(d,e,f))=20
> > > > >
> > > > > which lets you do more complex conditions:=20
> > > > >
> > > > > @(reset or ordered(d,e,f))=20
> > > > >
> > > > > However, there is probably syntax for assertions that already does =
> > > > > that=20
> > > > > without extra keywords/functions - which we should just reuse.=20
> > > > >
> > > > > Ordered events are already possible in plain old Verilog:
> > > > >
> > > > > @d @e @f <statement>
> > > > >
> > > > > But, I like the "ordered" modifier. It is intuitive and fits well
> > > = with
> > > > > @(changed ...). It makes the complex condition above easy to code.
> > > > >
> > > > > Hmmm, does the "iff" keyword help with any of this?
> > > > >
> > > > > =20
> > > > >
> > > > > I still think we should differentiate between actual events =
> > > > > (Verilog) and=20
> > > > > dynamic events (donation) with syntax. E.g. the example in
> > > (12.8.2) =
> > > > > becomes=20
> > > > > something like:=20
> > > > >
> > > > > event a,d, // events=20
> > > > > &b, &c; // event references=20
> > > > >
> > > > > b &=3D a;=20
> > > > > -> c; // null operation=20
> > > > > -> a; // also triggers b=20
> > > > > -> b; // also triggers a=20
> > > > > c &=3D b; // c now refers to a too=20
> > > > > -> a; // also triggers b and c=20
> > > > > -> b; // also triggers a and c=20
> > > > > -> c; // also triggers a and b=20
> > > > >
> > > > > loop1: always @(c) begin=20
> > > > > ...=20
> > > > > c &=3D null; // block at loop1 until reassigned=20
> > > > > end=20
> > > > >
> > > > > always @(reset) c &=3D a; // reactivate loop1=20
> > > > > always @(set) c &=3D d; // reactivate loop1=20
> > > > > =20
> > > > >
> > > > > The proposed scheme defies easy analysis.=20
> > > > >
> > > > > I don't know about defying analysis, but it sure look readable, =
> > > > > intuitive, and Verilog-like to me. I like it!
> > > > >
> > > > > Kev.=20
> > > > >
> > > > > --=20
> > > > > National Semiconductor, Tel: (408) 721 3251=20
> > > > > 2900 Semiconductor Drive, Mail Stop D3-500, Santa Clara, CA =
> > > > > 95052-8090=20
> > > > > =20
> > > > > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > > > Stuart Sutherland Sutherland HDL Inc.
> > > > > stuart@sutherland-hdl.com 22805 SW 92nd Place
> > > > > phone: 503-692-0898 Tualatin, OR 97062
> > > > > www.sutherland-hdl.com
> > > > >
> > > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ =
> > > >
> > > > --
> > > > National Semiconductor, Tel: (408) 721 3251
> > > > 2900 Semiconductor Drive, Mail Stop D3-500, Santa Clara, CA 95052-8090
> > > >
> > > >
> > > >
> > > >
> >
> > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > Stuart Sutherland Sutherland HDL Inc.
> > stuart@sutherland-hdl.com 22805 SW 92nd Place
> > phone: 503-692-0898 Tualatin, OR 97062
> > www.sutherland-hdl.com
> > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >
>
>
This archive was generated by hypermail 2b28 : Fri Dec 20 2002 - 15:49:05 PST