Before this commit, the C backend would put the translated definitions
of an interface file A.epi into A.{h,c}. This is inconsistent with the
C code generated for source files, which expects to find A_types.{h,c}.
For a clock (ck on x) to be well-formed in an environment H, we must
have H x = ck, i.e., the clock of x is the same as the clock ck of the
stream being sampled.
This constraint is guaranteed by construction for fully inferred clocks
(by the rules for when and merge), but nothing guarantees that user
declarations be well-formed. This can lead to problems.
For instance, this invalid program is incorrectly accepted:
node f (x : bool; a : bool :: . on b;
b : bool :: . on a) returns (y:bool);
let
y = true;
tel
as is this one:
node f(a: bool :: . on a; b: bool :: . on a)
returns (z: bool);
var w : bool;
let
w = a when b;
z = false fby w;
tel
This invalid program is incorrectly accepted and leads to an internal
compiler error:
node f (x : bool) returns (y:bool);
var a : bool :: . on b;
b : bool :: . on a;
let
y = true;
a = true;
b = true;
tel
This patch enforces the well-formedness constraint. It gives a sensible
error message when the constraint cannot be satisfied.
The parameter or local declaration:
x : bool :: . on y
Can now be made using the 'standard' Lustre syntax:
x : bool when y
In this case, the translation gives x the clock:
'a on y
and relies on the (MiniLS) Clocking pass to instantiate the fresh clock
variables.
Minils.CtrlNbacGen relies on such type annotations to generat
Controllable-Nbac nodes. This fix allows the use of ReaX to enforce
contracts of nodes involving merge operations on tuples over multiple
data types, such as:
(a, b) = merge c
(true -> ((true when c), (0 when c)))
(false -> ((false whenot c), (2 whenot c)));
This kind of code previously led to erroneous Controllable-Nbac code.
Leave only comments about compilation passes.
Remove output of intermediate code on standard output: added
generation of <module>.log containing all intermediate code.
The following node was accepted by the causality analysis:
node m(x:int) = (y,z:int)
let
automaton
state A
do
y = x + z;
z = x + 1;
until x = 3 then B
state B
do
y = x + 3;
z = y * x;
until x = 10 then A
end
tel
Each state is indeed causal, but once the automaton is translated to equations
(which is the systematic way in the current version), the node is not
schedulable.
Correction: all "Or" of dependency constraints translated to "And".
This constraint could be relaxed if code generation is done from Heptagon code,
before translation to minils equations.
Changes in CtrlNbac back-end :
- systematic sink state, only in ctrln code
- sink state "ok" removed (useless as was coded)
- no sink state in minils code
- outputs in ctrln are local variables, no longer state variables
