PATCH RFC: dwarf2 EH for Darwin

[Stan Shebs <shebs at apple dot com>]

This patch enables dwarf2-based exception support for Darwin.  While
the patch is mostly non-controversial, there are two changes to
generic code that I would like to get feedback on.

The first is the macro ASM_OUTPUT_DWARF_DELTA, which seems like
it probably ought to be a target hook, even though it only needs
a non-default value for the one target.

The second is the TERMINATE_EH_FRAME_INFO, which looks kludgy to
me, but I'm not clear on how this ought to be done.  The crux of
the problem is that EH_FRAME_SECTION_NAME should not be set for
Darwin, because it enables ELFish bits we don't really want.
Perhaps the zero-termination choice ought to be a generic target
hook too?

In any case, here's the patch.  As is, it bootstraps on Darwin, and
knocks off a couple hundred G++ testsuite fails.

Stan

2002-09-09  Stan Shebs  <shebs@apple.com>

   * config.gcc (i[34567]86-*-darwin*): Define extra_parts.
   (powerpc-*-darwin*): Ditto.
   * crtstuff.c [OBJECT_FORMAT_MACHO]: Update the Mach-O bits
   to work correctly for Darwin.
   * dwarf2asm.c (dw2_asm_output_delta): Use macro
   ASM_OUTPUT_DWARF_DELTA if defined.
   * dwarf2out.c (output_call_frame_info): Use
   TERMINATE_EH_FRAME_INFO to control output of terminating zero.
   * config/darwin.h (OBJECT_FORMAT_MACHO): Define.
   (STARTFILE_SPEC): Add crtbegin.o.
   (ENDFILE_SPEC): Define.
   (EXTRA_SECTION_FUNCTIONS): Put gcc_except_tab in data segment.
   (ASM_PREFERRED_EH_DATA_FORMAT): Handle more cases.
   (ASM_OUTPUT_DWARF_DELTA): Define.
   (TERMINATE_EH_FRAME_INFO): Define.
   * config/darwin.c (darwin_asm_output_dwarf_delta): New function.

Index: config.gcc
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config.gcc,v
retrieving revision 1.244
diff -p -r1.244 config.gcc
*** config.gcc    3 Sep 2002 20:21:32 -0000    1.244
--- config.gcc    9 Sep 2002 23:02:27 -0000
*************** i[34567]86-*-darwin*)
*** 980,985 ****
--- 980,986 ----
     target_gtfiles="\$(srcdir)/config/darwin.c"
     c_target_objs="darwin-c.o"
     cxx_target_objs="darwin-c.o"
+     extra_parts="crtbegin.o crtend.o"
     # Darwin linker does collect2 functionality
     use_collect2=no
     ;;
*************** powerpc-*-darwin*)
*** 1946,1951 ****
--- 1947,1953 ----
     target_gtfiles="\$(srcdir)/config/darwin.c"
     c_target_objs="darwin-c.o"
     cxx_target_objs="darwin-c.o"
+     extra_parts="crtbegin.o crtend.o"
     # Darwin linker does collect2 functionality
     use_collect2=no
     extra_headers=altivec.h
Index: crtstuff.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/crtstuff.c,v
retrieving revision 1.56
diff -p -r1.56 crtstuff.c
*** crtstuff.c    4 Jun 2002 02:55:28 -0000    1.56
--- crtstuff.c    9 Sep 2002 23:02:28 -0000
*************** __do_global_ctors (void)
*** 545,608 ****

 #else  /* OBJECT_FORMAT_MACHO */

! /* For Mach-O format executables, we assume that the system's runtime is
!    smart enough to handle constructors and destructors, but doesn't have
!    an init section (if it can't even handle constructors/destructors
!    you should be using INVOKE__main, not crtstuff). All we need to do
!    is install/deinstall the frame information for exceptions. We do this
!    by putting a constructor in crtbegin.o and a destructor in crtend.o.
!
!    crtend.o also puts in the terminating zero in the frame information
!    segment.  */
!
! /* The crtstuff for other object formats use the symbol __EH_FRAME_BEGIN__
!    to figure out the start of the exception frame, but here we use
!    getsectbynamefromheader to find this value. Either method would work,
!    but this method avoids creating any global symbols, which seems
!    cleaner.  */
!
! #include <mach-o/ldsyms.h>
! extern const struct section *
!   getsectbynamefromheader (const struct mach_header *,
!                const char *, const char *);

 #ifdef CRT_BEGIN

! static void __reg_frame_ctor (void) __attribute__ ((constructor));

 static void
! __reg_frame_ctor (void)
 {
!   static struct object object;
!   const struct section *eh_frame;
!
!   eh_frame = getsectbynamefromheader (&_mh_execute_header,
!                       "__TEXT", "__eh_frame");
!   __register_frame_info ((void *) eh_frame->addr, &object);
! }

! #elif defined(CRT_END)

! static void __dereg_frame_dtor (void) __attribute__ ((destructor));

 static void
! __dereg_frame_dtor (void)
 {
!   const struct section *eh_frame;

!   eh_frame = getsectbynamefromheader (&_mh_execute_header,
!                       "__TEXT", "__eh_frame");
!   __deregister_frame_info ((void *) eh_frame->addr);
 }

! /* Terminate the frame section with a final zero.  */
! STATIC int __FRAME_END__[]
!      __attribute__ ((unused, mode(SI), section(EH_FRAME_SECTION_NAME),
!              aligned(4)))
!      = { 0 };

 #else /* ! CRT_BEGIN && ! CRT_END */
 #error "One of CRT_BEGIN or CRT_END must be defined."
 #endif

! #endif /* OBJECT_FORMAT_MACHO */
--- 545,624 ----

 #else  /* OBJECT_FORMAT_MACHO */

! /* Crt stuff for Mach-O (NeXT and Darwin).
!
!    The theory of this is that each dynamically-loadable module,
!    including the main program itself, must have been positioned by
!    dyld before any frame info can be registered.  So we set up the
!    registration functions as dyld hooks, using a "preregistration"
!    function that is called directly from the system crt1.o.  */

 #ifdef CRT_BEGIN

! /* Homemade decls substituting for getsect.h and dyld.h, so cross
!    compilation works.  */
! struct mach_header;
! extern char *getsectdatafromheader (struct mach_header *, const char *,
!                     const char *, unsigned long *);
! extern void _dyld_register_func_for_add_image
!   (void (*) (struct mach_header *, unsigned long));
! extern void _dyld_register_func_for_remove_image
!   (void (*) (struct mach_header *, unsigned long));
!
! extern void __darwin_gcc3_preregister_frame_info (void);

 static void
! unwind_dyld_add_image_hook (struct mach_header *mh,
!                 unsigned long vm_slide)
 {
!   unsigned long sz;
!   char *fde;

!   fde = getsectdatafromheader (mh, "__TEXT", "__eh_frame", &sz);
!   if (fde)
!     {
!       struct object *ob = (struct object *) malloc (sizeof (struct 
object));

!       __register_frame_info (fde + vm_slide, ob);
!     }
! }

 static void
! unwind_dyld_remove_image_hook (struct mach_header *mh,
!                    unsigned long vm_slide)
 {
!   unsigned long sz;
!   char *fde;
!
!   fde = getsectdatafromheader (mh, "__TEXT", "__eh_frame", &sz);

!   if (fde)
!     __deregister_frame_info (fde + vm_slide);
 }

! /* Call this routine from the system crt1.o.  The call is standard in
!    Darwin 6.x (Mac OS X 10.2) and later; for earlier systems, you
!    would need to modify crt.c in the Csu project.  (This isn't great,
!    but other alternatives run afoul of linker semantics.  This
!    function is declared as common and tested before being called, so
!    that programs compiled by older GCCs still link and run.)  */
!
! void
! __darwin_gcc3_preregister_frame_info ()
! {
!   _dyld_register_func_for_add_image (unwind_dyld_add_image_hook);
!   _dyld_register_func_for_remove_image (unwind_dyld_remove_image_hook);
! }
!
! #elif defined(CRT_END) /* ! CRT_BEGIN */
!
! /* Install a single zero word at the end of the __eh_frame section.  */
!
! asm (".section __TEXT,__eh_frame");
! asm (".long 0");

 #else /* ! CRT_BEGIN && ! CRT_END */
 #error "One of CRT_BEGIN or CRT_END must be defined."
 #endif

! #endif  /* OBJECT_FORMAT_MACHO  */
Index: dwarf2asm.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/dwarf2asm.c,v
retrieving revision 1.27
diff -p -r1.27 dwarf2asm.c
*** dwarf2asm.c    4 Jun 2002 07:07:22 -0000    1.27
--- dwarf2asm.c    9 Sep 2002 23:02:28 -0000
*************** dw2_asm_output_delta VPARAMS ((int size,
*** 104,114 ****
   VA_FIXEDARG (ap, const char *, lab2);
   VA_FIXEDARG (ap, const char *, comment);

   dw2_assemble_integer (size,
             gen_rtx_MINUS (Pmode,
                        gen_rtx_SYMBOL_REF (Pmode, lab1),
                        gen_rtx_SYMBOL_REF (Pmode, lab2)));
!
   if (flag_debug_asm && comment)
     {
       fprintf (asm_out_file, "\t%s ", ASM_COMMENT_START);
--- 104,117 ----
   VA_FIXEDARG (ap, const char *, lab2);
   VA_FIXEDARG (ap, const char *, comment);

+ #ifdef ASM_OUTPUT_DWARF_DELTA
+   ASM_OUTPUT_DWARF_DELTA (asm_out_file, size, lab1, lab2);
+ #else
   dw2_assemble_integer (size,
             gen_rtx_MINUS (Pmode,
                        gen_rtx_SYMBOL_REF (Pmode, lab1),
                        gen_rtx_SYMBOL_REF (Pmode, lab2)));
! #endif
   if (flag_debug_asm && comment)
     {
       fprintf (asm_out_file, "\t%s ", ASM_COMMENT_START);
Index: dwarf2out.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/dwarf2out.c,v
retrieving revision 1.385
diff -p -r1.385 dwarf2out.c
*** dwarf2out.c    8 Sep 2002 18:36:53 -0000    1.385
--- dwarf2out.c    9 Sep 2002 23:02:33 -0000
*************** output_call_frame_info (for_eh)
*** 2043,2049 ****
--- 2043,2056 ----
       ASM_OUTPUT_LABEL (asm_out_file, l2);
     }

+ #ifndef TERMINATE_EH_FRAME_INFO
+ #define TERMINATE_EH_FRAME_INFO 0
 #ifndef EH_FRAME_SECTION_NAME
+ #define TERMINATE_EH_FRAME_INFO 1
+ #endif
+ #endif
+
+ #if TERMINATE_EH_FRAME_INFO
   if (for_eh)
     dw2_asm_output_data (4, 0, "End of Table");
 #endif
Index: config/darwin.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/darwin.c,v
retrieving revision 1.26
diff -p -r1.26 darwin.c
*** config/darwin.c    21 Aug 2002 02:41:46 -0000    1.26
--- config/darwin.c    9 Sep 2002 23:02:35 -0000
*************** darwin_globalize_label (stream, name)
*** 1294,1298 ****
--- 1294,1322 ----
     default_globalize_label (stream, name);
 }

+ /* Output a difference of two labels that is relocatable.  */
+
+ static int dwarf_label_counter;
+
+ void
+ darwin_asm_output_dwarf_delta (file, size, lab1, lab2)
+      FILE *file;
+      int size ATTRIBUTE_UNUSED;
+      const char *lab1, *lab2;
+ {
+   const char *p = lab1 + (lab1[0] == '*');
+   int islocaldiff = (p[0] == 'L');
+
+   if (islocaldiff)
+     fprintf (file, "\t.set L$set$%d,", dwarf_label_counter);
+   else
+     fprintf (file, "\t%s\t", ".long");
+   assemble_name (file, lab1);
+   fprintf (file, "-");
+   assemble_name (file, lab2);
+   if (islocaldiff)
+     fprintf (file, "\n\t.long L$set$%d", dwarf_label_counter++);
+ }
+
 #include "gt-darwin.h"

Index: config/darwin.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/darwin.h,v
retrieving revision 1.38
diff -p -r1.38 darwin.h
*** config/darwin.h    3 Sep 2002 20:09:18 -0000    1.38
--- config/darwin.h    9 Sep 2002 23:02:35 -0000
*************** Boston, MA 02111-1307, USA.  */
*** 35,40 ****
--- 35,46 ----
    leave it undefined and expect system builders to set configure args
    correctly.  */

+ /* One of Darwin's NeXT legacies is the Mach-O format, which is partly
+    like a.out and partly like COFF, with additional features like
+    multi-architecture binary support.  */
+
+ #define OBJECT_FORMAT_MACHO
+
 /* Suppress g++ attempt to link in the math library automatically.
    (Some Darwin versions have a libm, but they seem to cause problems
    for C++ executables.)  */
*************** Boston, MA 02111-1307, USA.  */
*** 92,99 ****

 #undef    STARTFILE_SPEC
 #define STARTFILE_SPEC  \
!   "%{pg:%{static:-lgcrt0.o}%{!static:-lgcrt1.o}} \
!     %{!pg:%{static:-lcrt0.o}%{!static:-lcrt1.o}}"

 #undef    DOLLARS_IN_IDENTIFIERS
 #define DOLLARS_IN_IDENTIFIERS 2
--- 98,109 ----

 #undef    STARTFILE_SPEC
 #define STARTFILE_SPEC  \
!   "%{pg:%{static:-lgcrt0.o}%{!static:-lgcrt1.o -lcrtbegin.o}} \
!     %{!pg:%{static:-lcrt0.o}%{!static:-lcrt1.o -lcrtbegin.o}}"
!
! #undef    ENDFILE_SPEC
! #define ENDFILE_SPEC \
!   "-lcrtend.o"

 #undef    DOLLARS_IN_IDENTIFIERS
 #define DOLLARS_IN_IDENTIFIERS 2
*************** SECTION_FUNCTION (machopic_picsymbol_stu
*** 425,431 ****
         ".picsymbol_stub", 0)              \
 SECTION_FUNCTION (darwin_exception_section,        \
         in_darwin_exception,            \
!         ".section __TEXT,__gcc_except_tab", 0)    \
 SECTION_FUNCTION (darwin_eh_frame_section,        \
         in_darwin_eh_frame,            \
         ".section __TEXT,__eh_frame", 0)    \
--- 435,441 ----
         ".picsymbol_stub", 0)              \
 SECTION_FUNCTION (darwin_exception_section,        \
         in_darwin_exception,            \
!         ".section __DATA,__gcc_except_tab", 0)    \
 SECTION_FUNCTION (darwin_eh_frame_section,        \
         in_darwin_eh_frame,            \
         ".section __TEXT,__eh_frame", 0)    \
*************** enum machopic_addr_class {
*** 597,603 ****
  
 #undef ASM_PREFERRED_EH_DATA_FORMAT
 #define ASM_PREFERRED_EH_DATA_FORMAT(CODE,GLOBAL)  \
!   (((CODE) == 1 || (GLOBAL) == 0) ? DW_EH_PE_pcrel : DW_EH_PE_absptr)

 #define DARWIN_REGISTER_TARGET_PRAGMAS(PFILE)                \
   do {                                    \
--- 607,620 ----
  
 #undef ASM_PREFERRED_EH_DATA_FORMAT
 #define ASM_PREFERRED_EH_DATA_FORMAT(CODE,GLOBAL)  \
!   (((CODE) == 2 && (GLOBAL) == 1) \
!    ? (DW_EH_PE_pcrel | DW_EH_PE_indirect) : \
!      ((CODE) == 1 || (GLOBAL) == 0) ? DW_EH_PE_pcrel : DW_EH_PE_absptr)
!
! #define ASM_OUTPUT_DWARF_DELTA(FILE,SIZE,LABEL1,LABEL2)  \
!   darwin_asm_output_dwarf_delta (FILE, SIZE, LABEL1, LABEL2)
!
! #define TERMINATE_EH_FRAME_INFO 0

 #define DARWIN_REGISTER_TARGET_PRAGMAS(PFILE)                \
   do {                                    \