[Bug c++/96311] [8/9/10/11 Regression] false positive for -Wunused-but-set-variable (const/constexpr identifier used in generic lambda)

[jakub at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96311

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
If I compare
void
foo ()
{
  constexpr int used = 0;
  [](auto unused)
  {
    return used;
  };
}

void
bar ()
{
  constexpr int used = 0;
  [](int unused)
  {
    return used;
  };
}

then in bar mark_exp_read is called on used from check_return_expr ->
convert_for_initialization -> convert_from_assignment ->
perform_implicit_conversion_flags -> mark_rvalue_use -> mark_use ->
mark_exp_read, but for the generic lambda check_return_expr takes the goto
dependent; and doesn't mark anything.
It is only marked during instantiation of the lambda, such as when (0) as added
in between } and ; and in that case the warning is gone.

More complete testcase:
void
foo ()
{
  constexpr int used = 0;
  [](auto unused)
  {
    return used;
  };
}

void
bar ()
{
  constexpr int used = 0;
  [](int unused)
  {
    return used;
  };
}

void
baz ()
{
  constexpr int used = 0;
  [](auto unused)
  {
    return used;
  } (0);
}

void
qux ()
{
  constexpr int used = 0;
  [](auto unused)
  {
    return used + 1;
  };
}

void
corge ()
{
  constexpr int used = 0;
  [](auto unused)
  {
    return used + unused;
  };
}

This shows that if the variable is used in some non-dependent expression as in
qux (and not directly in return statement), it is marked already without
instantiation, but when it is only used directly in return stmt or used in a
dependent expression, it is marked only during instantiation of the lambda.

Though, when the lambda isn't really instantiated, the variable is not really
used, so not 100% sure the warning is unwanted.