Re: symbolic simulation/abstract execution of HDL's

kgg@lfcs.ed.ac.uk (Kees Goossens)
Wed, 15 May 91 17:54:02 GMT

          From comp.compilers

Related articles
Re: symbolic simulation/abstract execution of HDL's kgg@lfcs.ed.ac.uk (1991-05-15)
| List of all articles for this month |

Newsgroups: comp.compilers
From: kgg@lfcs.ed.ac.uk (Kees Goossens)
Summary: This (and more) is possible _now_
Keywords: optimization, interpreter
Organization: The Lavatory for the Foundations of Computer Science
References: <1991May10.165316.22625@fs7.ece.cmu.edu>
Date: Wed, 15 May 91 17:54:02 GMT

[reposted by the author from comp.lang.vhdl -John]


In article <1991May10.165316.22625@fs7.ece.cmu.edu> rajen@edsel.ece.cmu.edu (Ramchandani Rajen Sham) writes:
>I am looking for any code that will enable me to perform symbolic simulation/
>abstract execution of any hardware description language, like verilog,
>vhdl or isp. ...


I am working on embedding CHDDLs in proof systems. In particular a subset
of ELLA (see below) in the Lambda proof system. Having written a formal
operational semantics for ELLA, and embedding it in Lambda I am now working
on some small examples dealing with symbolic simulation.


My system allows you to do things like:


--------------------------- [1]
|- AND (one,zero) => zero


meaning that AND 1 0 evaluates to 0. This is done by expanding the
definition of the AND:


|- ...
-------------------------
|- CASE (one,zero) OF (one,one): one ELSE zero ESAC => answer
---------------------------
|- AND (one,zero) => answer


by applying a rule which characterises the behaviour of a CASE we can
rewrite this to


|- answer = zero
---------------------------
|- AND (one,zero) => answer


and then to [1]


More interestingly, however we can a symbolic simulation:


|- CASE (x,y) OF (one,one): one ELSE zero ESAC => answer
----------------------
|- AND (x,y) => answer


Because we don't have any more information about x and y this is what we
get stuck with. We can however, push the computation into the answer so
that we get:


----------------------
|- AND (x,y) => CASE (x,y) OF (one,one): one ELSE zero ESAC


This is not very useful in this case, but when dealing with an adder say we
can prove the following rules:


|- sum = (x+y+c) mod n |- carry = (x+y+c) div n
--------------------------------------------------- [2]
|- ADD n (x,y,c) => (sum,carry)


or:


--------------------------------------------------- [3]
|- ADD n (x,y,c) => ( (x+y+c) mod n, (x+y+c) div n )


Here we have a general n bit adder, which we have proved correct with
respect to its specification (ie its output is the sum, carry pair)
Using this, we can derive the rules [2] and [3]. That is,
once we have proved a circuit implements its specification we can use the
specification when simulating the circuit. This is one example of mixed
level simulation which this framework allows.


We can prove for example that


------------------------------
|- HALFADDER x zero => (x,zero)


i.e. the carry will be zero, and the sum the as yet unknown value x.
(Here is a variable, not a value in the value domain.)
This has been established by doing a straightforward symbolic simulation,
and may be saved as a derived rule. Derived rules are very useful, in that
once we have symbolically simulated a circuit, we can reapply this
simulation with known values in place of the variables *in one step*.


A very interesting and novel way of using symblic simulation is not to
allow not just symbolic variables, but also *symbolic circuits*. Think of
plug in components as a circuit of which you know the specification, but
not its actual body/implementation. A normal simulator cannot work with
this, but in my framework you can still simulate it.


Suppose we know that the component satisfies its specification:


-----------------------------
|- PLUGIN_SPEC (unknown_impl)


PLUGIN_SPEC(unknown_impl) |- derive output here, using spec.
-----------------------------------
PLUGIN_SPEC(unknown_impl) |- PLUGIN unknown_impl => output


For example, if the component is an adder the spec would be as in [2], [3]
above and the we'd derive that output is in fact the sum and carry.
Other usages of symbolic circuit simulations include simulations of general
design decompositions. Various design methodologies for creating
mathematically correct hardware can be used in conjunction with this approach.




The model behaves sensibly with delayless feedbacks and oscillating
circuits, buses (bidirectional wires) etc.
The examples so far have been at the gate level, but it is possible to
model circuits at the transistor level, using stored charge, overpowering,
multiple signal strengths etc etc. All the power of symbolic simulation may
be use at this level too.




Any way, all of this is not up to industrial strength yet, and not very
fast. However, I think it will become possible to deal with medium sized
circuits in the not so distant future. Witness the VIPER chip, which was
partially verified using the HOL higher order logic proof system.


Kees Goossens


I have written "Semantics for picoELLA" is a manuscript describing a
formal semantics for a minimal subset of ELLA. "An Operational Semantics
for a Subset of the HDDL ELLA" is a manuscript describing a formal
semantics for a maximal subset of ELLA. "Embedding CHDDLs in Proof
systems" describes the general approach. It will be presented as a paper
in Turin at the advanced research workshop on correct hardware design
methodologies in June 1991. It will be available as a LFCS technical
report LFCS-91-155 soon. (Laboratory for the foundations of Computer
Science, Department of computer science, Edinburgh University, Scotland)
Send me some email for more information.


ELLA is a trademark of the Secretary of State for Defence, United Kingdom,
and is marketed by Computer General Electronic Design, The New Church,
Henry St, Bath, UK. ELLA was designed at the Royal Radar and Signals
Establishment (RSRE) in Malvern, UK, where ELLA research is ongoing. I
have no affiliation with either CGED or RSRE.


--


Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.