The expression 'not (8 > 0 or 42 > 0)' is compiled incorrectly by the GDSL compiler

[kranzj]

The expression 'not (8 > 0 or 42 > 0)' is compiled incorrectly by the GDSL compiler. Consider the following GDSL program:

export test_right : (int) -> |1|
export test_wrong : (int) -> |1|

val test_right x = not (8 > 0)
val test_wrong x = not (8 > 0 or 42 > 0)

The code is compiled into the C code shown below:

...
/* test_right */
static int_t test_right(state_t s,int_t x) {
  return vec_not(s,gen_vec(1, 0<8)).data;
}
/* test_wrong */
static int_t test_wrong(state_t s,int_t x) {
  return vec_not(s,gen_vec(0, (0<8) | (0<42))).data;
}
...

Here, the size of the vector is 1 in case of the test_right function, but 0 in case of test_wrong function. The size of the vector should be 1 in both cases.

[axel-simon]

Hi Julian,

that’s bad. There is a problem in the internal representation in that an INT2VECexp sometimes does not have a size which is then set by the type inference. That is brittle. I have not good fix since I don’t understand which kind of extra type propagation is needed. It boils down to the following:

In
https://github.com/gdslang/gdsl-toolkit/blob/master/detail/semantic/primitives.sml#L337 https://github.com/gdslang/gdsl-toolkit/blob/master/detail/semantic/primitives.sml#L337
the primitives “and” and “or” are translated using boxVany which uses an INT2VEC (0,arg) which means that it converts an Int to a vector of 0 bits. That’s always wrong. The way this is repaired later is in

https://github.com/gdslang/gdsl-toolkit/blob/master/detail/imp/imp-opt.sml#L1231 https://github.com/gdslang/gdsl-toolkit/blob/master/detail/imp/imp-opt.sml#L1231

where these special translations are given for the 0-bit case:

    | PRIexp (ANDprim,_,[VEC2INTexp (NONE,e1),VEC2INTexp (NONE,e2)]) =>
        lub (s, visitExp s e1, visitExp s e2)

Here, the size is (I think) somehow regenerated by unifying the type which seems to fail if the outer expression’s type is generic. I would have to dig into this which is difficult at the moment (I frist have to get approval to work on this open source project).

Stay tuned,
Axel

On Nov 12, 2015, at 2:27 AM, Julian Kranz notifications@github.com wrote:

The expression 'not (8 > 0 or 42 > 0)' is compiled incorrectly by the GDSL compiler. Consider the following GDSL program:

export test_right : (int) -> |1|
export test_wrong : (int) -> |1|

val test_right x = not (8 > 0)
val test_wrong x = not (8 > 0 or 42 > 0)
´´´

The code is compiled into the following C code:

...
/* test_right */
static int_t test_right(state_t s,int_t x) {
  return vec_not(s,gen_vec(1, 0<8)).data;
}
/* test_wrong */
static int_t test_wrong(state_t s,int_t x) {
  return vec_not(s,gen_vec(0, (0<8) | (0<42))).data;
}
...
´´´

Here, the size of the vector is 1 in case of the ```test_right´´´ function, but 0 in case of ```test_wrong´´´ function. The size of the vector should be one in both bases.
—
Reply to this email directly or view it on GitHub <https://github.com/gdslang/gdsl-toolkit/issues/20>.

[axel-simon]

I think this bug is due to the simplifier removing an VEC2INT(INT2VEC()) pair (or the other way round). I think in the complex expression this results in something that is not an AND or and OR node to be under an INT2VEC(0, ...) node. The way to debug this is to print the expression in this line:
https://github.com/gdslang/gdsl-toolkit/blob/master/detail/imp/imp-opt.sml#L1233