Closed
Description
Summary:
There is an inconsistency between the operator precedence as implemented in stanc2 and as described in the manual.
Description:
The manual describes that .* binds more tightly than *. stanc2 seems to bind * more tightly.
Reproducible Steps:
Try to compile
model {
print(1.0 .* [1., 2.]');
}
Get error
SYNTAX ERROR, MESSAGE(S) FROM PARSER:
No matches for:
real .* vector
Available argument signatures for operator.*:
vector .* vector
row_vector .* row_vector
matrix .* matrix
Expression is ill formed.
error in 'stan/src/test/test-models/good/empty.stan' at line 2, column 25
-------------------------------------------------
1: model {
2: print(1.0 .* [1., 2.]');
^
3: }
-------------------------------------------------
So clearly, the signatures for .* are
vector .* vector
row_vector .* row_vector
matrix .* matrix .
Now, if we try to compile
model {
print([3., 4.]' * 1.0 .* [1., 2.]');
}
we would expect that to fail. Indeed, it should be parsed as
model {
print([3., 4.]' * (1.0 .* [1., 2.]'));
}
according to the manual. However, it parses just fine and prints [3, 8] when it runs. This suggests strongly to me that it has parsed the model as
model {
print(([3., 4.]' * 1.0) .* [1., 2.]');
}
instead.
Additional Information:
Rstanarm is currently exploiting this bug.
This also makes me worried about the other stated operator precedences. Has this been resulting in incorrect generated code?
Current Version:
v2.18.0