3.0.1 PATCH: Allow building shared libobjc on Solaris 2

[Rainer Orth <ro at TechFak dot Uni-Bielefeld dot DE>]

As already reported in

	http://gcc.gnu.org/ml/gcc-patches/2001-03/msg00125.html

and several PRs, GCC 3.0 on Solaris 2 (both SPARC and Intel) fails to
bootstrap when configured with --enable-shared (which only changes the way
libobjc is built which has AC_DISABLE_SHARED in configure.in; all other
libraries supporting shared libs are already built shared by default).

A very conservative patch to fix this only for HP-UX went into 3.0, but
this fix is needed for Solaris 2 as well.

Here's what I've done to validate it is safe on a couple of platforms:

* apply the patch below

* configure --enable-shared --enable-languages=c,objc; make bootstrap

* cd gcc; make check-objc (this is sufficient since only objc files are
  touched by the patch)

The platforms involved are

* sparc-sun-solaris2.5.1

* sparc-sun-solaris2.8 (bi-arch 32x64 configuration)

* mips-sgi-irix6.2 (bi-arch n32, n64 setup)

* alpha-dec-osf4.0f

* alpha-dec-osf5.1

The patch allows libobjc to build shared on Solaris 2, where bootstrap
failed before, and bootstrap continues to complete on the other platforms.

check-objc differences between static and shared results are as follows:

* sparc-sun-solaris2.5.1:

-PASS: objc/execute/class-13.m execution,  -O 
+FAIL: objc/execute/class-13.m execution,  -O 

-FAIL: objc/execute/class-3.m execution,  -O 
+PASS: objc/execute/class-3.m execution,  -O 

-PASS: objc/execute/class-5.m execution,  -O 
+FAIL: objc/execute/class-5.m execution,  -O 

-PASS: objc/execute/class-6.m execution,  -O 
+FAIL: objc/execute/class-6.m execution,  -O 

-PASS: objc/execute/class-8.m execution,  -O 
+FAIL: objc/execute/class-8.m execution,  -O 

-PASS: objc/execute/class-9.m execution,  -O 
+FAIL: objc/execute/class-9.m execution,  -O 

-# of expected passes           148
-# of unexpected failures       1
+# of expected passes           144
+# of unexpected failures       5
 # of expected failures         1

  All of the new failures abort with

  test_instances failed

  Still it's much better than before when the bootstrap would fail (and we
  receive lots of bug reports about this).

* sparc-sun-solaris2.8:

  -m32 (default):

-PASS: objc/execute/class-10.m execution,  -O 
+FAIL: objc/execute/class-10.m execution,  -O 

-PASS: objc/execute/class-11.m execution,  -O 
+FAIL: objc/execute/class-11.m execution,  -O 

-PASS: objc/execute/class-14.m execution,  -O 
+FAIL: objc/execute/class-14.m execution,  -O 

-# of expected passes           149
+# of expected passes           146
+# of unexpected failures       3
 # of expected failures         1

  Same error as on Solaris 2.5.1 above.

  -m64:

  unchanged:

FAIL: objc/execute/bycopy-3.m execution,  -O 
FAIL: objc/execute/formal_protocol-5.m execution,  -O 

  Bus Error in both cases.

  dbx gives

signal BUS (invalid address alignment) in objc_msg_lookup at 0xffffffff7f2144d0
0xffffffff7f2144d0: objc_msg_lookup+0x0034:     ldx     [%o5 + 0x40], %o2
(/opt/SUNWspro/bin/../WS6U1/bin/sparcv9/dbx) where
=>[1] objc_msg_lookup(0x1001016d8, 0x1001016a0, 0xffffffff7efb34e8, 0x100000b6c, 0x0, 0xffffffff7f600000), at 0xffffffff7f2144d0
  [2] main(0x1, 0xffffffff7ffff3a8, 0xffffffff7ffff3b8, 0x100101798, 0x100000000, 0x0), at 0x100000d70

  Since Solaris 2/SPARC V9 gdb doesn't exist yet, I'll have to dig into
  gcc's assembler code to find what's wrong here, but I'll defer this since
  this is no regression introduced by this patch.

# of expected passes            147
# of unexpected failures        2
# of expected failures          1

* mips-sgi-irix6.2:

  -mabi=n32 (default):

-PASS: objc/execute/class-2.m execution,  -O 
+FAIL: objc/execute/class-2.m execution,  -O 

  Same error (test_instances failed) as Solaris 2.5.1 above.

-# of expected passes           149
+# of expected passes           148
+# of unexpected failures       1
 # of expected failures         1

  -mabi=64:

  unchanged:

FAIL: objc/execute/bycopy-3.m execution,  -O 
FAIL: objc/execute/formal_protocol-5.m execution,  -O 

  These are the same two failures as on 64-bit Solaris 8 above, so they
  probably happen for all/most 64-bit big-endian targets.  Again, no
  regression compared to the static libobjc case.

# of expected passes            147
# of unexpected failures        2
# of expected failures          1

* alpha-dec-osf4.0f, alpha-dec-osf5.1:

  no differences:

# of expected passes             149
# of expected failures           1

Ok for branch (the mainline already has a cleaner version of handle_class_ref)?

There are a couple of issues on how to proceed from here:

* As explained in

	http://gcc.gnu.org/ml/gcc-patches/2001-05/msg00635.html

  a shared libobjc is close to unusable if installed into $libsubdir.  It
  needs to go into $libdir instead.  I can provide a separate patch for
  this, if it's agreed that this is the way to go.

* Once we have tested more platforms to make sure the shared libobjc works
  on all of them, we could remove the AC_DISABLE_SHARED from configure.in
  and instead add logic to respect --enable-shared, with and without
  explicitly specified libraries.  This is most likely a mainline-only
  change.

Comments?

	Rainer

-----------------------------------------------------------------------------
Rainer Orth, Faculty of Technology, Bielefeld University

Email: ro@TechFak.Uni-Bielefeld.DE


Thu Jun 28 00:47:23 2001  Rainer Orth  <ro@TechFak.Uni-Bielefeld.DE>

	* objc/objc-act.c (handle_class_ref): Always place decl in
	variable section.
	Fixes PRs libobjc/917, libobjc/3237, bootstrap/3251.

	testsuite:
	* lib/objc.exp (objc_target_compile): Declare rootme.  Append to
	ld_library_path.

===================================================================
RCS file: gcc/objc/RCS/objc-act.c,v
retrieving revision 1.1
diff -up -r1.1 gcc/objc/objc-act.c
--- gcc/objc/objc-act.c	2001/06/15 01:08:40	1.1
+++ gcc/objc/objc-act.c	2001/06/14 23:08:40
@@ -8362,13 +8362,8 @@ handle_class_ref (chain)
       pushdecl (decl);
       rest_of_decl_compilation (decl, 0, 0, 0);
 
-#ifdef __hpux__
       /* Put the decl in the variable section.  It may need relocation.  */
       variable_section (decl, 1);
-#else
-      /* Make following constant read-only (why not)?  */
-      readonly_data_section ();
-#endif
 
       exp = build1 (ADDR_EXPR, string_type_node, decl);
 
Index: gcc/testsuite/lib/objc.exp
===================================================================
RCS file: /cvs/gcc/gcc/gcc/testsuite/lib/objc.exp,v
retrieving revision 1.5.6.4
diff -u -p -r1.5.6.4 objc.exp
--- objc.exp	2001/06/28 12:53:48	1.5.6.4
+++ objc.exp	2001/06/28 17:43:59
@@ -113,13 +113,14 @@ proc objc_init { args } {
 }
 
 proc objc_target_compile { source dest type options } {
+    global rootme;
     global tmpdir;
     global gluefile wrap_flags;
     global OBJC_UNDER_TEST
     global TOOL_OPTIONS
     global ld_library_path
 
-    set ld_library_path "."
+    set ld_library_path ".:${rootme}"
     lappend options "libs=-lobjc"
 
     if { [target_info needs_status_wrapper]!="" && [info exists gluefile] } {