c++: Fix up -fstrong-eval-order handling of call arguments [PR70796]

c++: Fix up -fstrong-eval-order handling of call arguments [PR70796]

For -fstrong-eval-order (default for C++17 and later) we make sure to
gimplify arguments in the right order, but as the following testcase
shows that is not enough.
The problem is that some lvalues can satisfy the is_gimple_val / fb_rvalue
predicate used by gimplify_arg for is_gimple_reg_type typed expressions,
or is_gimple_lvalue / fb_either used for other types.
E.g. in foo we have:
  C::C (&p,  ++i,  ++i)
before gimplification where i is an automatic int variable and without this
patch gimplify that as:
  i = i + 1;
  i = i + 1;
  C::C (&p, i, i);
which means that the ctor is called with the original i value incremented
by 2 in both arguments, while because the call is CALL_EXPR_ORDERED_ARGS
the first argument should be different.  Similarly in qux we have:
  B::B (&p, TARGET_EXPR <D.2274, *(const struct A &) A::operator++ (&i)>,
        TARGET_EXPR <D.2275, *(const struct A &) A::operator++ (&i)>)
and gimplify it as:
      _1 = A::operator++ (&i);
      _2 = A::operator++ (&i);
      B::B (&p, MEM[(const struct A &)_1], MEM[(const struct A &)_2]);
but because A::operator++ returns the passed in argument, again we have
the same value in both cases due to gimplify_arg doing:
      /* Also strip a TARGET_EXPR that would force an extra copy.  */
      if (TREE_CODE (*arg_p) == TARGET_EXPR)
        {
          tree init = TARGET_EXPR_INITIAL (*arg_p);
          if (init
              && !VOID_TYPE_P (TREE_TYPE (init)))
            *arg_p = init;
        }
which is perfectly fine optimization for calls with unordered arguments,
but breaks the ordered ones.
Lastly, in corge, we have before gimplification:
  D::foo (NON_LVALUE_EXPR <p>, 3,  ++p)
and gimplify it as
  p = p + 4;
  D::foo (p, 3, p);
which is again wrong, because the this argument isn't before the
side-effects but after it.
The following patch adds cp_gimplify_arg wrapper, which if ordered
and is_gimple_reg_type forces non-SSA_NAME is_gimple_variable
result into a temporary, and if ordered, not is_gimple_reg_type
and argument is TARGET_EXPR bypasses the gimplify_arg optimization.
So, in foo with this patch we gimplify it as:
  i = i + 1;
  i.0_1 = i;
  i = i + 1;
  C::C (&p, i.0_1, i);
in qux as:
      _1 = A::operator++ (&i);
      D.2312 = MEM[(const struct A &)_1];
      _2 = A::operator++ (&i);
      B::B (&p, D.2312, MEM[(const struct A &)_2]);
where D.2312 is a temporary and in corge as:
  p.9_1 = p;
  p = p + 4;
  D::foo (p.9_1, 3, p);
The is_gimple_reg_type forcing into a temporary should be really cheap
(I think even at -O0 it should be optimized if there is no modification in
between), the aggregate copies might be more expensive but I think e.g. SRA
or FRE should be able to deal with those if there are no intervening
changes.  But still, the patch tries to avoid those when it is cheaply
provable that nothing bad happens (if no argument following it in the
strong evaluation order doesn't have TREE_SIDE_EFFECTS, then even VAR_DECLs
etc. shouldn't be modified after it).  There is also an optimization to
avoid doing that for this or for arguments with reference types as nothing
can modify the parameter values during evaluation of other argument's
side-effects.

I've tried if e.g.
  int i = 1;
  return i << ++i;
doesn't suffer from this problem as well, but it doesn't, the FE uses
  SAVE_EXPR <i>, SAVE_EXPR <i> << ++i;
in that case which gimplifies the way we want (temporary in the first
operand).

2021-11-19  Jakub Jelinek  <jakub@redhat.com>

	PR c++/70796
	* cp-gimplify.c (cp_gimplify_arg): New function.
	(cp_gimplify_expr): Use cp_gimplify_arg instead of gimplify_arg,
	pass true as last argument to it if there are any following
	arguments in strong evaluation order with side-effects.

	* g++.dg/cpp1z/eval-order11.C: New test.