[Bug c++/84961] [7/8 Regression] ICE error: SSA_NAME_DEF_STMT is wrong

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84961

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|ice-on-invalid-code         |ice-on-valid-code
           Priority|P3                          |P2
             Status|NEW                         |ASSIGNED
                 CC|                            |jason at gcc dot gnu.org
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
So we do

      tret = gimplify_expr (&TREE_VALUE (link), pre_p, post_p,
                            is_inout ? is_gimple_min_lval : is_gimple_lvalue,
                            fb_lvalue | fb_mayfail);

and gimplify b = (double) a to _3 with is_gimple_lvalue.  Which it isn't
here (the predicate returns true though).

With a similar

int main()
{
  int a,b,c;
  (b = c) = a;
}

the FE already replaces this with <<error_mark_node>> so it doesn't reach
gimplification.  It seems it expects the gimplifier to do this here.

Note the C FE complains:

t.c: In function ‘c’:
t.c:3:12: error: lvalue required in asm statement
 void c() { asm("" : "=a"(b = a)); }
            ^~~

but still throws the invalid asm at the gimplifier.  The C++ FE fails to
complain, but I guess (b = a) is an lvalue for C++?  In which case the
testcase would be valid C++ but invalid C?  G++ 6 accepts it.

The above testcase is genericized to

  <<cleanup_point <<< Unknown tree: expr_stmt
  (void) (b = c;, b = a) >>>>>;

but the asm stays

<<cleanup_point __asm__("":"=a" b = (double) a:)>>;

I think the C++ FE should help us here by genericizing to b = (double) a; b?