Re: PATCH: PR target/63815: [5 Regression] g++.dg/other/pr53811.C fails with -mcmodel=large -fpic

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Wed, Nov 12, 2014 at 2:00 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Wed, Nov 12, 2014 at 01:51:01PM -0800, H.J. Lu wrote:
>> > So, all in all, SET_REGNO sounds inappropriate to me, everywhere else
>> > gen_raw_REG or gen_rtx_REG is used instead, so IMNSHO you should do the
>> > same.
>> >
>> >         Jakub
>>
>> It makes sense.  I checked in the following patch.
>
> gen_raw_REG would be better, the difference is that gen_rtx_REG might
> in theory return you pic_offset_table_rtx again.

The whole  x86_output_mi_thunk function uses gen_rtx_REG.  Using
gen_raw_REG is out of space.

> Anyway, I don't think anything is needed for release branches,
> pic_offset_table_rtx should be initialized there just once and not changed
> afterwards.

Try this with released GCC:

[hjl@gnu-6 tmp]$ cat foo.h
struct ICCStringClass
{
  virtual int CreateString (int) = 0;
};

struct AGSCCDynamicObject
{
  virtual void Unserialize () = 0;
};


struct ScriptString:AGSCCDynamicObject, ICCStringClass
{
  virtual int CreateString (int);
  virtual void Unserialize ();
};
[hjl@gnu-6 tmp]$ cat foo1.cc
#include "foo.h"

int
__attribute__ ((noinline))
CreateNewScriptString (int fromText, bool reAllocate = true)
{
  return fromText;
}

int
__attribute__ ((noinline))
ScriptString::CreateString (int fromText)
{
  return CreateNewScriptString (fromText);
}

void
__attribute__ ((noinline))
ScriptString::Unserialize ()
{
}
[hjl@gnu-6 tmp]$ cat foo2.cc
#include "foo.h"

int
main ()
{
  ICCStringClass *x = new ScriptString;
  if (x->CreateString (1) != 1)
    __builtin_abort ();
  return 0;
}
[hjl@gnu-6 tmp]$ g++ -shared -fpic -O2 -mcmodel=large -o libfoo.so foo1.cc
[hjl@gnu-6 tmp]$ g++ foo2.cc ./libfoo.so
[hjl@gnu-6 tmp]$ ./a.out
Segmentation fault
[hjl@gnu-6 tmp]$




-- 
H.J.