bad code generation for arrays (2)

[Lennart Ochel]

Bad simulation code gets generated for the example bug_2286 from our testsuite.

location: /testsuite/simulation/modelica/algorithms_functions/bug_2286.mos

error message

Simulation failed for model: bug_2286
Error: Error building simulator. Build log: gcc-4.4   -falign-functions -march=native -mfpmath=sse -fPIC   -I\"/home/lochel/Documents/workspace/OpenModelica/trunk/bui$
bug_2286.c: In function 'eqFunction_3':
bug_2286.c:237: error: used struct type value where scalar is required
bug_2286.c: In function 'eqFunction_7':
bug_2286.c:335: error: used struct type value where scalar is required
make: *** [bug_2286.o] Error 1

[Lennart Ochel]

This should be fixed as soon as possible, because it blocks the Petri net library.

[Mahder Alemseged Gebremedhin]

This was an issue with vector operations on boolean arrays. There was no support for logical operations on arrays of booleans.

Fixed in r16715 and r16719.

[Lennart Ochel]

The following example does still not simulate:

model bug_2286
  Boolean a[1] = {false};
  Boolean b[1] = {true};
  Boolean c[1];
algorithm
  c := a and b;
end bug_2286;

output:

bug_2286.c: In function ‘eqFunction_3’:
bug_2286.c:240: error: used struct type value where scalar is required
bug_2286.c: In function ‘eqFunction_5’:
bug_2286.c:329: error: used struct type value where scalar is required

bug_2286.c (line 240 contains the call of copy_boolean_array_data_mem):

/*
 equation index: 3
 type: ALGORITHM

   c := {false} and {true};
 */
static void eqFunction_3(DATA *data)
{
  boolean_array tmp0;
  boolean_array tmp1;
  array_alloc_scalar_boolean_array(&tmp0, 1, (modelica_boolean)(0));
  array_alloc_scalar_boolean_array(&tmp1, 1, (modelica_boolean)(1));
  copy_boolean_array_data_mem(&(tmp0 && tmp1), &$Pc);
}

(Maybe this is already mentioned in another ticket?!)

[Mahder Alemseged Gebremedhin]

This last issue should not have reached the code gneration. It should have been removed by constant evaluation in the front-end somewhere in ceval.

c := {false} and {true};

should have been reduced to

c := {false};

Anyways this issue could have happened for other cases that can not be reduced with constant folding.

It is handled now. Fixed in r16724.

[Henning Kiel]

I implemented the same code for the unary 'not' operator in r16727.

Still code is not generated as expected:

model model_bug_2286

Boolean a[1] = {time > 0.4};
Boolean b[1] = {time > 0.5};
Boolean c[1] = {time > 0.6};
Boolean d[1];
Boolean e[1];
Boolean f[1];
Boolean g[1];

algorithm

d:=a and b and c; /* fails */
e:=not (a and b); /* fails */
f:=not a and b; /* fails */
g:=a and not b; /* succeeds */

end model_bug_2286;

[Martin Sjölund]

Your code only looks for CREF and array. But in this case it needs to look for a binary expression. There exists functions so you can match against the type of the expression instead. Or better yet: Match against the type inside the record AND. Then it will work for any expression.

[Mahder Alemseged Gebremedhin]

Fixed in r16756 using the type in the operators instead of the types from the lhs and rhs. It should work fine since there is no operator in modelica that can take boolean arrays and result in boolean scalar and vice versa. If there are more issues we should open a new ticket since the original issue is already fixed.