https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85599
Thomas Koenig <tkoenig at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|WAITING |NEW
Component|fortran |middle-end
Summary|invalid optimization: |Function need not be
|function not always |evaluated in logical
|evaluated in logical |expression
|expression |
--- Comment #13 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
In the original test case, it is not necessary to evaluate
the function at all. A function evaluation should always be
cheaper than just checking a variable. Constant propagation
would then eliminate all of the test case, leading only
to the effect of
flag = .false.
As a more realistic example, one should change
subroutine foo(a,b)
logical, intent(in) :: a
logical, intent(out) :: b
b = func() .and. a
end subroutine foo
to
subroutine foo(a,b)
logical, intent(in) :: a
logical, intent(out) :: b
if (a) then
b = func()
else
b = .false.
end if
end subroutine
We are passing TRUTH_AND_EXPR to the middle end to allow such
optimizations. Is there something in the semantics of TRUTH_AND_EXPR
(for example needed for other languages) that says that the
function check() needs to be called? Or is this simply some
optimization opportunity that is currently not taken?
