All
I'd like to give a (short?) explanation of what I said in yesterday's
meeting about the impossibility of expressing a directive level forall
on a cover directive in terms of a parameterized operator inside the
cover directive.
(In the following let 'forall' and 'forsome' be imaginary directive
level quantors, 'and' and 'or' be formula level conjunction and
disjunction and '&&','&' and '|' be sere operators.)
It is the case that
forall i in S: assert F(i) <=> assert for i in S: and(F(i))
but not
forall i in S: cover r(i) <=> cover for i in S: op(r(i))
regardless whether op is '&&', '&' or '|'
It is the case however that
forsome i in S: cover r(i) <=> cover for i in S: |(r(i))
The reason for this is apparent if we supply a formal semantics for the
directives (which we haven't done yet).
I think satisfaction for directives should be defined for a set of runs
rather than for a single run because a directive represents a proof
obligation w.r.t. a design and a design defines a set of possible runs.
So if RUNS is a set of runs we would want to have something like the
following, where '|=d' is directive level and '|=' is property level
truth.
RUNS|=d cover r
<=>
exists w in RUNS s.t.
w|= {[*];r}!
RUNS|=d assert F
<=>
for all w in RUNS
w|= F
A forall on directive level is really an indication that the proof
obligation should be multiplied, so for instance
forall i in S: assert F(i) is true w.r.t. a set of runs RUNS iff
for all i in S:
RUNS|=d assert F(i)
<=>
for all i in S:
for all w in RUNS
w|= F(i)
<=>
for all w in RUNS
for all i in S:
w|= F(i)
<=>
for all w in RUNS
w|= for i in S: and(F(i))
<=>
RUNS|=d assert for i in S: and(F(i))
and
forall i in S: cover r(i) is true w.r.t. a set of runs RUNS iff
for all i in S:
RUNS|=d cover r
<=>
for all i in S:
exists w in RUNS s.t.
w|= {[*];r}!
Howeve this is not the same as
exists w in RUNS s.t.
for all i in S:
w|= {[*];r}!
<=>
exists w in RUNS s.t.
w|= for i in S: and{[*];r}!
nor is it the same as
exists w in RUNS s.t.
exists i in S:
w|= {[*];r}!
<=>
exists w in RUNS s.t.
w|= for i in S: |{[*];r}!
we have however
exists i in S:
RUNS|=d cover r(i)
<=>
exists i in S:
exists w in RUNS s.t.
w|= {[*];r(i)}!
<=>
exists w in RUNS s.t.
exists i in S:
w|= {[*];r(i)}!
<=>
exists w in RUNS s.t.
w|= for i in S: |{[*];r(i)}!
<=> [this should be true, have to check!]
exists w in RUNS s.t.
w|= {[*];for i in S: |r(i)}!
<=>
RUNS|=d cover for i in S: |(r(i))
A way to explain this informally is that a directive level forall on a
cover directive instructs the tool to check if there is for every
instantiation of the variable a possible run in which the sere is
covered. This means that possibly different runs are required for
different instances of the cover directive. No operator inside a cover
directive can achieve this. A single cover directive is either
(entirely) satisfied in a some single run or not at all.
With directive level forall on an assert directive the tool is
instructed to check whether for all instantiations and for all runs the
property is satisfied, and that is the same as to check whether for all
runs the corresponding conjunction is satisfied.
Best Regards, Johan M
On Tue, Dec 21, 2004 at 06:26:12AM -0800, Erich Marschner wrote:
> Dana,
>
> I would argue that such an "automatic reduction" is no less a tool implementation choice than the implementation choices I mentioned - and for 'cover', where the goal is to get detailed information about what has been tested, rather than to verify the design, the tool implementation choices I suggested might be preferable for a given implementor. Also, the language will be simpler if we avoid "multiplying entities" unnecessarily, to quote Occam. If the issue is indeed a tool implementation choice - whether to optimize a parameterized operator, or whether to elaborate it to collect more information - then let's leave that decision to the tool.
>
> Perhaps others on the working group have an opinion about this?
>
> Regards,
>
> Erich
>
> | -----Original Message-----
> | From: Dana Fisman [mailto:dana.fisman@gmail.com]
> | Sent: Tuesday, December 21, 2004 3:48 AM
> | To: Erich Marschner
> | Cc: Cindy Eisner; ieee-1850-extensions@eda.org
> | Subject: Re: Proposal for Parameterized And/Or operators
> |
> | Erich,
> |
> | I still think we cannot use the parameterized-or to do what
> | you want for cover. You are relying too much on the
> | implementation, which may vary from tool to tool. Note for
> | example, that you are disallowing an automatic reduction that
> | replaces (for j in {0..15}: | {data=j}) with
> | (true) in cases where indeed data can only assume values in 0..15.
> | This is unacceptable ??? since semantically they are the same.
> | If you have a different replicator for the directive, this
> | problem does not occur.
> |
> | Regards,
> | Dana.
> |
> |
> |
> | On Mon, 20 Dec 2004 13:43:23 -0800, Erich Marschner
> | <erichm@cadence.com> wrote:
> | > Hi Dana,
> | >
> | > I understand your point, but I would argue that your key
> | assumption is false.
> | >
> | > First, it would rarely be the case that the range given for
> | the parameter is the entire range of the data value it is
> | being compared to. For example, I may have an address value
> | whose range is 0..2^32-1, for which I might write the
> | following cover directives:
> | >
> | > A_0: cover {addr==0};
> | > A_1: cover for i in {1:4095} {addr==i};
> | > A_2: cover for j in {4096:65535} {addr==j};
> | > A_3: cover {65536 <= addr};
> | >
> | > This effectively creates 4 "bins" that we want to cover,
> | corresponding to ranges of address values that have
> | particular meanings - e.g., address==0 is the null address,
> | perhaps addresses in the range up to 4K are in ROM, and
> | addresses above that, up to 2^16-1, are reserved for some
> | special usage. In doing data coverage, we want to know if
> | any of the addresses in a given bin have been hit, so the
> | 'or' semantics is exactly what we want here.
> | >
> | > The first level of coverage only tells us whether or not we
> | have hit some bin. Here, the PSL semantics of cover
> | directives suffice. However, in production verification
> | flows, users want to know not only whether, but how many
> | times, a given coverage bin has been hit. So the first and
> | most obvious addition for PSL implementors is to count how
> | many times a cover directive has been satisfied, rather than
> | just registering yes or no.
> | >
> | > Beyond that, if we have defined parameters for cover
> | directives, the next obvious thing to do would be to record
> | the parameter values for which the cover directive was
> | satisfied. Doing this would allow an implementation to track
> | individual values within a bin, and chart the distribution of
> | those values over the range. So even if a cover directive
> | specified the entire possible range of values for a variable
> | (creating one bin for the entire range), thereby making the
> | directive equivalent to "cover {True}", a tool could use the
> | parameter values to give more detailed information about
> | *how* the cover directive was satisfied in a given cycle.
> | This goes beyond PSL cover directive semantics, but it is
> | entirely consistent with the disjunction semantics that I
> | proposed - it simply provides more detail.
> | >
> | > So the key point here is that we are NOT assuming that the
> | range specified for a cover directive is all-inclusive, i.e.,
> | that by definition a parameterized cover directive (with
> | disjunction semantics) will reduce to cover {True}. Instead,
> | this proposed capability assumes a typical use model in which
> | different cover statements are provided for different subsets
> | or subranges of the values of a variable, and furthermore
> | that tools may elect to use the parameter values that satisfy
> | a cover directive to provide more detail.
> | >
> | > Regards,
> | >
> | > Erich
> | >
> | > | -----Original Message-----
> | > | From: Dana Fisman [mailto:dana.fisman@gmail.com]
> | > | Sent: Monday, December 20, 2004 4:02 PM
> | > | To: Erich Marschner
> | > | Cc: Cindy Eisner; ieee-1850-extensions@eda.org
> | > | Subject: Re: Proposal for Parameterized And/Or operators
> | > |
> | > | Erich,
> | > |
> | > | I guess I understand your motivation, but I don't think
> | the details
> | > | work.
> | > |
> | > | > cover for i in {0:4095} : |({data==i}); ...
> | > | >
> | > | > is essentially shorthand for
> | > | >
> | > | > cover {data==0};
> | > | > cover {data==1};
> | > | > cover {data==2};
> | > | > ...
> | > | > cover {data==4095};
> | > |
> | > | I disagree.
> | > |
> | > | cover for i in {0:4095} : |({data==i});
> | > |
> | > | is equivalent to
> | > |
> | > | cover {data==0} | {data==1} | {data==2} | ... | {data==4095}
> | > |
> | > | which is in turn
> | > |
> | > | cover {true} assuming data can only assume values
> | > | between 0 to 4095.
> | > |
> | > | (the dual of what you noted. i.e. that with the conjunction
> | > | semantics we get cover{false}).
> | > |
> | > | This is why I still think we need a replication operator
> | that can be
> | > | applied to directive. So that for example
> | > |
> | > | > forall i in {0:4095} : cover {data==i};
> | > |
> | > | will be equivalent to 4096 cover statements (as above).
> | > |
> | > | Regards,
> | > | Dana.
> | > |
> | > |
> | > | On Mon, 20 Dec 2004 10:16:16 -0800, Erich Marschner
> | > | <erichm@cadence.com> wrote:
> | > | > Hi Dana,
> | > | >
> | > | > This is the separate response I mentioned regarding the
> | > | motivation for disjunction of sequences.
> | > | >
> | > | > | I also don't understand the motivation for disjunction of
> | > | a sequence
> | > | > | (neither I understand exactly the proposal: does it apply to
> | > | > | standalone sequences? what about applying it to the
> | > | entire left hand
> | > | > | side or right hand side of a suffix implication operator?).
> | > | >
> | > | > The notion of disjunction on sequences came from the idea
> | > | of using a
> | > | > forall-like parameter in a cover directive, to say "cover
> | > | any sequence
> | > | > that includes a value in the range of the parameter" - e.g.,
> | > | >
> | > | > cover forall i in {0:4095} : {data==i};
> | > | >
> | > | > However, the fact that forall is hardwired to mean
> | > | conjunction makes this meaningless - clearly the
> | conjunction (either
> | > | & or &&) of the sequences implied by the above cannot
> | hold, because
> | > | data cannot take on any more than one value in any given cycle.
> | > | What we want is the disjunction of the above sequences, i.e., to
> | > | cover the case where any one of the sequences occurs.
> | This is what
> | > | originally caused to me to suggest a "for some" operator,
> | which led
> | > | to the discussion of parameterized and/or operators on properties
> | > | and sequences.
> | > | >
> | > | > With the proposed parameterized and/or operators, we can say
> | > | > instead
> | > | >
> | > | > cover for i in {0:4095} : |({data==i});
> | > | >
> | > | > (It is not clear to me that we need the parentheses around
> | > | {data==i}
> | > | > in addition to the braces, but I'll leave that to be
> | > | discussed later.)
> | > | >
> | > | > This is essentially shorthand for
> | > | >
> | > | > cover {data==0};
> | > | > cover {data==1};
> | > | > cover {data==2};
> | > | > ...
> | > | > cover {data==4095};
> | > | >
> | > | > Note that disjunction is in some sense a natural semantics
> | > | for cover directives (and sequences), whereas conjunction is a
> | > | natural semantics for assert directives (and properties).
> | > | For the cover directive, which detects a sequence that may or may
> | > | not occur in a given verification run, we want to detect
> | ANY (a_0 or
> | > | a_1 or ... a_n) occurrence of any form of the
> | parameterized sequence
> | > | - but for assert directives, which are expected to hold at every
> | > | cycle in every verification run, we want to require EVERY
> | (a_0 and
> | > | a_1 and ... a_n) form of the parameterized property to hold.
> | > | >
> | > | > Note also that, if we can say the above, then this is
> | > | hardly different
> | > | > from applying forall to a directive, e.g.,
> | > | >
> | > | > forall i in {0:4095} : cover {data==i};
> | > | >
> | > | > In fact, since 'forall' suggests conjunction, it seems odd
> | > | to apply it to cover directives, which as I've just mentioned are
> | > | not all required to hold at any given time; the semantic sense of
> | > | 'cover' is more disjunction-oriented. So I believe allowing
> | > | disjunction on sequences, using the syntax that has been
> | proposed,
> | > | gives us the ability to 'apply forall to directives' as has been
> | > | requested, but with a more natural semantics.
> | > | >
> | > | > The proposed syntax could also be applied to the LHS or
> | RHS of a
> | > | > suffix implication operator, e.g.,
> | > | >
> | > | > for i in {0:3} : {a[*i]} |=> for j in {4:7} : {b[*j]}
> | > | >
> | > | > One question that remains to be addressed is the scope
> | of the 'for'
> | > | > parameter - does it only extend over the property or
> | > | sequence to which
> | > | > it applies, or does it extend beyond that? For example,
> | > | would it be
> | > | > possible to write
> | > | >
> | > | > for i in {0:3} : {a[*i]} |=> for j in {4:7} : {b[*j-i]}
> | > | >
> | > | > so that the number of b's is a function of the number of
> | > | a's? At the very least, we would need the value of the
> | parameter to
> | > | be reported as part of any message associated with an assert or
> | > | cover directive.
> | > | >
> | > | > Regards,
> | > | >
> | > | > Erich
> | > | >
> | > |
> | > |
> | >
> |
> |
-- ------------------------------------------------------------ Johan Martensson, PhD Office: +46 31 7451913 R&D Mobile: +46 703749681 Safelogic AB Fax: +46 31 7451939 Arvid Hedvalls Backe 4 johan.martensson@safelogic.se SE-411 33 Gothenburg, SWEDEN PGP key ID A8857A60 www.safelogic.se ------------------------------------------------------------ -- ------------------------------------------------------------ Johan Martensson, PhD Office: +46 31 7451913 R&D Mobile: +46 703749681 Safelogic AB Fax: +46 31 7451939 Arvid Hedvalls Backe 4 johan.martensson@safelogic.se SE-411 33 Gothenburg, SWEDEN PGP key ID A8857A60 www.safelogic.se ------------------------------------------------------------Received on Wed Dec 22 01:01:34 2004
This archive was generated by hypermail 2.1.8 : Wed Dec 22 2004 - 01:01:41 PST