• Arthur Loiret <arthur.loiret gmail.com> Nov 16 2007
• Ingo Oeser <ioe-news rameria.de> Nov 24 2007
• Arthur Loiret <arthur.loiret gmail.com> Nov 24 2007

Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable


Hi,


I've been able to build gdc with gcc-4.2, it stills segfault sometimes
and is unable to build phobos/tango for now, but it's a beginning and
can compile short D files.

I've put the patches in: http://people.dunnewind.net/arthur/gdc-4.2/patches/

gdc-4.2-build.dpatch updates gcc/d/Make-lang.in for the new gcc-4.2 objs
(pr26885) and gcc/d/d-builtins.c copied from the new c-common.c.

gdc-4.2.dpatch is mostly taked from patch-gcc-4.1.x, adding this:

--- gcc/tree.h  2007-11-15 14:21:01.000000000 +0100
+++ gcc/tree.h  2007-11-15 14:22:05.000000000 +0100
   -3535,6 +3535,19   
=20
 extern tree build_nt (enum tree_code, ...);
=20
+/* Still required by gdc */
+#if GCC_VERSION >=3D 3000 || __STDC_VERSION__ >=3D 199901L
+/* Use preprocessor trickery to map "build" to "buildN" where N is the
+   expected number of arguments.  This is used for both efficiency (no
+   varargs), and checking (verifying number of passed arguments).  */
+#define build(code, ...) \
+  _buildN1(build, _buildC1(__VA_ARGS__))(code, __VA_ARGS__)
+#define _buildN1(BASE, X)       _buildN2(BASE, X)
+#define _buildN2(BASE, X)       BASE##X
+#define _buildC1(...)           _buildC2(__VA_ARGS__,9,8,7,6,5,4,3,2,1,0,0)
+#define _buildC2(x,a1,a2,a3,a4,a5,a6,a7,a8,a9,c,...) c
+#endif
+
 extern tree build0_stat (enum tree_code, tree MEM_STAT_DECL);
 #define build0(c,t) build0_stat (c,t MEM_STAT_INFO)
 extern tree build1_stat (enum tree_code, tree, tree MEM_STAT_DECL);



I'm currently trying to fix ICEs, the problems seems to come from
gimple.


Any help would be welcome, there is still a lot of work to get it
working. :-)


Arthur.

Hi Arthur,

nice to see that happening!

Arthur Loiret wrote:

 I've been able to build gdc with gcc-4.2, it stills segfault sometimes
 and is unable to build phobos/tango for now, but it's a beginning and
 can compile short D files.


I've been trying to do the same some months ago.
 
 --- gcc/tree.h  2007-11-15 14:21:01.000000000 +0100
 +++ gcc/tree.h  2007-11-15 14:22:05.000000000 +0100
    -3535,6 +3535,19   
  
  extern tree build_nt (enum tree_code, ...);
  
 +/* Still required by gdc */
 +#if GCC_VERSION >= 3000 || __STDC_VERSION__ >= 199901L
 +/* Use preprocessor trickery to map "build" to "buildN" where N is the
 +   expected number of arguments.  This is used for both efficiency (no
 +   varargs), and checking (verifying number of passed arguments).  */
 +#define build(code, ...) \
 +  _buildN1(build, _buildC1(__VA_ARGS__))(code, __VA_ARGS__)
 +#define _buildN1(BASE, X)       _buildN2(BASE, X)
 +#define _buildN2(BASE, X)       BASE##X
 +#define _buildC1(...)          
 _buildC2(__VA_ARGS__,9,8,7,6,5,4,3,2,1,0,0) +#define
 _buildC2(x,a1,a2,a3,a4,a5,a6,a7,a8,a9,c,...) c +#endif
 +
  extern tree build0_stat (enum tree_code, tree MEM_STAT_DECL);
  #define build0(c,t) build0_stat (c,t MEM_STAT_INFO)
  extern tree build1_stat (enum tree_code, tree, tree MEM_STAT_DECL);


That's more (short term) maintainable than my approach. I patched the D glue
layer, fixed up the callers and provided build2() and build3() for compilers
below GCC 4.0.x. 

David, would you accept a patch from me, which does ONLY that?

 I'm currently trying to fix ICEs, the problems seems to come from
 gimple.


Yes, I have seen the same and my vacation was over, so I stopped it.

I have some questions for some parts of that patch. I'll mail them to you.


Best Regards

Ingo Oeser

Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable


Hi Ingo,


On Sat, Nov 24, 2007 at 03:13:52PM +0100, Ingo Oeser wrote:
 --- gcc/tree.h  2007-11-15 14:21:01.000000000 +0100
 +++ gcc/tree.h  2007-11-15 14:22:05.000000000 +0100
    -3535,6 +3535,19   
 =20
  extern tree build_nt (enum tree_code, ...);
 =20
 +/* Still required by gdc */
 +#if GCC_VERSION >=3D 3000 || __STDC_VERSION__ >=3D 199901L
 +/* Use preprocessor trickery to map "build" to "buildN" where N is the
 +   expected number of arguments.  This is used for both efficiency (no
 +   varargs), and checking (verifying number of passed arguments).  */
 +#define build(code, ...) \
 +  _buildN1(build, _buildC1(__VA_ARGS__))(code, __VA_ARGS__)
 +#define _buildN1(BASE, X)       _buildN2(BASE, X)
 +#define _buildN2(BASE, X)       BASE##X
 +#define _buildC1(...)         =20
 _buildC2(__VA_ARGS__,9,8,7,6,5,4,3,2,1,0,0) +#define
 _buildC2(x,a1,a2,a3,a4,a5,a6,a7,a8,a9,c,...) c +#endif
 +
  extern tree build0_stat (enum tree_code, tree MEM_STAT_DECL);
  #define build0(c,t) build0_stat (c,t MEM_STAT_INFO)
  extern tree build1_stat (enum tree_code, tree, tree MEM_STAT_DECL);


 That's more (short term) maintainable than my approach. I patched the D g=


 layer, fixed up the callers and provided build2() and build3() for compil=


 below GCC 4.0.x.=20


You're right. :-)


 I'm currently trying to fix ICEs, the problems seems to come from
 gimple.


 Yes, I have seen the same and my vacation was over, so I stopped it.


I made a few other temporary patches to fix ICEs, here is what changed:

diff -ruN gcc/gimplify.c gcc/gimplify.c
--- gcc/gimplify.c      2007-11-19 11:56:50.000000000 +0000
+++ gcc/gimplify.c      2007-11-19 11:58:32.000000000 +0000
   -1804,12 +1804,16   
              /* Divide the offset by its alignment.  */
              offset =3D size_binop (EXACT_DIV_EXPR, offset, factor);
=20
-             if (!is_gimple_min_invariant (offset))
-               {
-                 TREE_OPERAND (t, 2) =3D offset;
-                 tret =3D gimplify_expr (&TREE_OPERAND (t, 2), pre_p, post=
_p,
-                                       is_gimple_formal_tmp_reg, fb_rvalue=
);
-                 ret =3D MIN (ret, tret);
+             /* Would be better to have a look at size_binop ... */
+             if (offset !=3D NULL_TREE)
+               {
+                 if (!is_gimple_min_invariant (offset))
+                   {
+                     TREE_OPERAND (t, 2) =3D offset;
+                     tret =3D gimplify_expr (&TREE_OPERAND (t, 2), pre_p, =
post_p,
+                                           is_gimple_formal_tmp_reg, fb_rv=
alue);
+                     ret =3D MIN (ret, tret);
+                   }
                }
            }
        }
diff -ruN gcc/tree-inline.c gcc/tree-inline.c
--- gcc/tree-inline.c   2007-11-19 15:41:33.000000000 +0000
+++ gcc/tree-inline.c   2007-11-19 15:42:24.000000000 +0000
   -2636,8 +2636,9   
 {
   tree copy, type;
=20
-  gcc_assert (TREE_CODE (decl) =3D=3D PARM_DECL
-             || TREE_CODE (decl) =3D=3D RESULT_DECL);
+  if (cfun->naked) /* Not sure this is right ... */
+    gcc_assert (TREE_CODE (decl) =3D=3D PARM_DECL
+               || TREE_CODE (decl) =3D=3D RESULT_DECL);
=20
   type =3D TREE_TYPE (decl);
=20


gdc-4.2 now builds phobos and tango, see [0] and [1].

[0] http://packages.qa.debian.org/g/gdc-4.2.html
[1] http://people.dunnewind.net/arthur/gdc-4.2/

Here is how to test it:

1/ Add this to your /etc/apt/sources.list

deb http://ftp.fr.debian.org/debian experimental main
deb-src http://ftp.fr.debian.org/debian experimental main

2/ Run sudo aptitude update && sudo aptitude install gdc-4.2


I've cleanup gdc-4.2-build.dpatch: see:
http://people.dunnewind.net/arthur/gdc-4.2/patches/gdc-4.2-build.dpatch

David, could you also have a look at this one?


 I have some questions for some parts of that patch. I'll mail them to you.


Sure, no problem.


Have a nice day,

Arthur.