Fix PR middle-end/87574

[Eric Botcazou]

It turns out that, even if cgraph_node::expand_thunk happily overrides the 
DECL_IGNORED_P setting on the thunk from the front-end, this is necessary when 
the thunk is initially a back-end thunk and then turned into a GIMPLE thunk, 
e.g. because of inlining, to play nice with early debug info generation.

So the patch reinstates the overriding, but only when expanding to GIMPLE.

Tested on x86-64/Linux, applied on the mainline as obvious.


2018-10-10  Eric Botcazou  <ebotca...@adacore.com>

        PR middle-end/87574
        * cgraphunit.c (cgraph_node::expand_thunk): Force DECL_IGNORED_P on
        the thunk when expanding to GIMPLE.


2018-10-10  Eric Botcazou  <ebotca...@adacore.com>

        * g++.dg/other/pr87574.C: New test.

-- 
Eric Botcazou
Index: cgraphunit.c
===================================================================
--- cgraphunit.c	(revision 264986)
+++ cgraphunit.c	(working copy)
@@ -1862,6 +1862,12 @@ cgraph_node::expand_thunk (bool output_a
 	 DECL_ARGUMENTS.  In this case force_gimple_thunk is true.  */
       if (in_lto_p && !force_gimple_thunk)
 	get_untransformed_body ();
+
+      /* We need to force DECL_IGNORED_P when the thunk is created
+	 after early debug was run.  */
+      if (force_gimple_thunk)
+	DECL_IGNORED_P (thunk_fndecl) = 1;
+
       a = DECL_ARGUMENTS (thunk_fndecl);
 
       current_function_decl = thunk_fndecl;

// PR middle-end/87574
// Testcase by David Binderman <dcb...@hotmail.com>

// { dg-do compile }
// { dg-options "-O2 -g -Wno-return-type" }

class a {
public:
  virtual ~a();
};
class c {
public:
  enum j {};
  virtual j d() = 0;
};
class e : a, c {
  j d();
};
class f;
class g {
public:
  static g *h();
  f *i();
};
class f {
public:
  template <class b> b *l(int);
};
c::j e::d() {}
void m() {
  for (int k;;)
    g::h()->i()->l<c>(k)->d();
}