[project @ 2003-12-17 17:18:27 by sof]
[ghc-hetmet.git] / ghc / rts / Linker.c
index a465dc7..85d5809 100644 (file)
@@ -1,7 +1,6 @@
 /* -----------------------------------------------------------------------------
- * $Id: Linker.c,v 1.104 2002/10/02 09:36:00 wolfgang Exp $
  *
- * (c) The GHC Team, 2000, 2001
+ * (c) The GHC Team, 2000-2003
  *
  * RTS Object Linker
  *
@@ -31,7 +30,9 @@
 #include <sys/stat.h>
 #endif
 
-#ifdef HAVE_DLFCN_H
+#if defined(HAVE_FRAMEWORK_HASKELLSUPPORT)
+#include <HaskellSupport/dlfcn.h>
+#elif defined(HAVE_DLFCN_H)
 #include <dlfcn.h>
 #endif
 
@@ -57,7 +58,7 @@
 #include <sys/mman.h>
 #endif
 
-#if defined(linux_TARGET_OS) || defined(solaris2_TARGET_OS) || defined(freebsd_TARGET_OS)
+#if defined(linux_TARGET_OS) || defined(solaris2_TARGET_OS) || defined(freebsd_TARGET_OS) || defined(netbsd_TARGET_OS) || defined(openbsd_TARGET_OS)
 #  define OBJFORMAT_ELF
 #elif defined(cygwin32_TARGET_OS) || defined (mingw32_TARGET_OS)
 #  define OBJFORMAT_PEi386
 #  include <mach-o/loader.h>
 #  include <mach-o/nlist.h>
 #  include <mach-o/reloc.h>
+#  include <mach-o/dyld.h>
 #endif
 
 /* Hash table mapping symbol names to Symbol */
 static /*Str*/HashTable *symhash;
 
+/* List of currently loaded objects */
+ObjectCode *objects = NULL;    /* initially empty */
+
 #if defined(OBJFORMAT_ELF)
 static int ocVerifyImage_ELF    ( ObjectCode* oc );
 static int ocGetNames_ELF       ( ObjectCode* oc );
@@ -83,9 +88,12 @@ static int ocVerifyImage_PEi386 ( ObjectCode* oc );
 static int ocGetNames_PEi386    ( ObjectCode* oc );
 static int ocResolve_PEi386     ( ObjectCode* oc );
 #elif defined(OBJFORMAT_MACHO)
+static int ocAllocateJumpIslands_MachO ( ObjectCode* oc );
 static int ocVerifyImage_MachO    ( ObjectCode* oc );
 static int ocGetNames_MachO       ( ObjectCode* oc );
 static int ocResolve_MachO        ( ObjectCode* oc );
+
+static void machoInitSymbolsWithoutUnderscore( void );
 #endif
 
 /* -----------------------------------------------------------------------------
@@ -198,11 +206,7 @@ typedef struct _RtsSymbolVal {
       SymX(uname)                               \
       SymX(unlink)                              \
       SymX(utime)                               \
-      SymX(waitpid)                             \
-      Sym(__divdi3)                             \
-      Sym(__udivdi3)                            \
-      Sym(__moddi3)                             \
-      Sym(__umoddi3)
+      SymX(waitpid)
 
 #elif !defined(mingw32_TARGET_OS)
 #define RTS_MINGW_ONLY_SYMBOLS /**/
@@ -211,9 +215,21 @@ typedef struct _RtsSymbolVal {
 #define RTS_POSIX_ONLY_SYMBOLS  /**/
 #define RTS_CYGWIN_ONLY_SYMBOLS /**/
 
+/* Extra syms gen'ed by mingw-2's gcc-3.2: */
+#if __GNUC__>=3
+#define RTS_MINGW_EXTRA_SYMS                    \
+      Sym(_imp____mb_cur_max)                   \
+      Sym(_imp___pctype)            
+#else
+#define RTS_MINGW_EXTRA_SYMS
+#endif
+
 /* These are statically linked from the mingw libraries into the ghc
    executable, so we have to employ this hack. */
 #define RTS_MINGW_ONLY_SYMBOLS                  \
+      SymX(asyncReadzh_fast)                   \
+      SymX(asyncWritezh_fast)                  \
+      SymX(asyncDoProczh_fast)                 \
       SymX(memset)                              \
       SymX(inet_ntoa)                           \
       SymX(inet_addr)                           \
@@ -273,11 +289,8 @@ typedef struct _RtsSymbolVal {
       Sym(opendir)                              \
       Sym(readdir)                              \
       Sym(rewinddir)                            \
-      Sym(closedir)                             \
-      Sym(__divdi3)                             \
-      Sym(__udivdi3)                            \
-      Sym(__moddi3)                             \
-      Sym(__umoddi3)
+      RTS_MINGW_EXTRA_SYMS                      \
+      Sym(closedir)
 #endif
 
 #ifndef SMP
@@ -290,40 +303,40 @@ typedef struct _RtsSymbolVal {
       Maybe_ForeignObj                         \
       Maybe_Stable_Names                       \
       Sym(StgReturn)                           \
-      Sym(init_stack)                          \
-      SymX(__stg_chk_0)                                \
-      SymX(__stg_chk_1)                                \
-      SymX(stg_chk_2)                          \
-      SymX(stg_chk_3)                          \
-      SymX(stg_chk_4)                          \
-      SymX(stg_chk_5)                          \
-      SymX(stg_chk_6)                          \
-      SymX(stg_chk_7)                          \
-      SymX(stg_chk_8)                          \
-      Sym(stg_enterStackTop)                   \
-      SymX(stg_gc_d1)                          \
-      SymX(stg_gc_l1)                          \
+      SymX(stg_enter_info)                     \
+      SymX(stg_enter_ret)                      \
+      SymX(stg_gc_void_info)                   \
       SymX(__stg_gc_enter_1)                   \
-      SymX(stg_gc_enter_2)                     \
-      SymX(stg_gc_enter_3)                     \
-      SymX(stg_gc_enter_4)                     \
-      SymX(stg_gc_enter_5)                     \
-      SymX(stg_gc_enter_6)                     \
-      SymX(stg_gc_enter_7)                     \
-      SymX(stg_gc_enter_8)                     \
-      SymX(stg_gc_f1)                          \
       SymX(stg_gc_noregs)                      \
-      SymX(stg_gc_seq_1)                       \
-      SymX(stg_gc_unbx_r1)                     \
+      SymX(stg_gc_unpt_r1_info)                        \
       SymX(stg_gc_unpt_r1)                     \
-      SymX(stg_gc_ut_0_1)                      \
-      SymX(stg_gc_ut_1_0)                      \
-      SymX(stg_gen_chk)                                \
+      SymX(stg_gc_unbx_r1_info)                        \
+      SymX(stg_gc_unbx_r1)                     \
+      SymX(stg_gc_f1_info)                     \
+      SymX(stg_gc_f1)                          \
+      SymX(stg_gc_d1_info)                     \
+      SymX(stg_gc_d1)                          \
+      SymX(stg_gc_l1_info)                     \
+      SymX(stg_gc_l1)                          \
+      SymX(__stg_gc_fun)                       \
+      SymX(stg_gc_fun_info)                    \
+      SymX(stg_gc_fun_ret)                     \
+      SymX(stg_gc_gen)                         \
+      SymX(stg_gc_gen_info)                    \
+      SymX(stg_gc_gen_hp)                      \
+      SymX(stg_gc_ut)                          \
+      SymX(stg_gen_yield)                      \
+      SymX(stg_yield_noregs)                   \
       SymX(stg_yield_to_interpreter)           \
+      SymX(stg_gen_block)                      \
+      SymX(stg_block_noregs)                   \
+      SymX(stg_block_1)                                \
+      SymX(stg_block_takemvar)                 \
+      SymX(stg_block_putmvar)                  \
+      SymX(stg_seq_frame_info)                 \
       SymX(ErrorHdrHook)                       \
       MAIN_CAP_SYM                              \
       SymX(MallocFailHook)                     \
-      SymX(NoRunnableThreadsHook)              \
       SymX(OnExitHook)                         \
       SymX(OutOfHeapHook)                      \
       SymX(PatErrorHdrHook)                    \
@@ -357,17 +370,20 @@ typedef struct _RtsSymbolVal {
       SymX(divExactIntegerzh_fast)             \
       SymX(divModIntegerzh_fast)               \
       SymX(forkzh_fast)                                \
-      SymX(forkProcesszh_fast)                  \
+      SymX(forkProcess)                                \
+      SymX(forkOS_createThread)                        \
       SymX(freeHaskellFunctionPtr)             \
       SymX(freeStablePtr)                      \
       SymX(gcdIntegerzh_fast)                  \
       SymX(gcdIntegerIntzh_fast)               \
       SymX(gcdIntzh_fast)                      \
+      SymX(genSymZh)                           \
       SymX(getProgArgv)                                \
       SymX(getStablePtr)                       \
       SymX(int2Integerzh_fast)                 \
       SymX(integer2Intzh_fast)                 \
       SymX(integer2Wordzh_fast)                        \
+      SymX(isCurrentThreadBoundzh_fast)                \
       SymX(isDoubleDenormalized)               \
       SymX(isDoubleInfinite)                   \
       SymX(isDoubleNaN)                                \
@@ -386,9 +402,10 @@ typedef struct _RtsSymbolVal {
       SymX(newArrayzh_fast)                    \
       SymX(newBCOzh_fast)                      \
       SymX(newByteArrayzh_fast)                        \
-      SymX(newCAF)                             \
+      SymX_redirect(newCAF, newDynCAF)         \
       SymX(newMVarzh_fast)                     \
       SymX(newMutVarzh_fast)                   \
+      SymX(atomicModifyMutVarzh_fast)          \
       SymX(newPinnedByteArrayzh_fast)          \
       SymX(orIntegerzh_fast)                   \
       SymX(performGC)                          \
@@ -399,6 +416,7 @@ typedef struct _RtsSymbolVal {
       SymX(quotIntegerzh_fast)                 \
       SymX(quotRemIntegerzh_fast)              \
       SymX(raisezh_fast)                       \
+      SymX(raiseIOzh_fast)                     \
       SymX(remIntegerzh_fast)                  \
       SymX(resetNonBlockingFd)                 \
       SymX(resumeThread)                       \
@@ -407,6 +425,7 @@ typedef struct _RtsSymbolVal {
       SymX(rts_eval)                           \
       SymX(rts_evalIO)                         \
       SymX(rts_evalLazyIO)                     \
+      SymX(rts_evalStableIO)                   \
       SymX(rts_eval_)                          \
       SymX(rts_getBool)                                \
       SymX(rts_getChar)                                \
@@ -415,10 +434,12 @@ typedef struct _RtsSymbolVal {
       SymX(rts_getInt)                         \
       SymX(rts_getInt32)                       \
       SymX(rts_getPtr)                         \
+      SymX(rts_getFunPtr)                      \
       SymX(rts_getStablePtr)                   \
       SymX(rts_getThreadId)                    \
       SymX(rts_getWord)                                \
       SymX(rts_getWord32)                      \
+      SymX(rts_lock)                           \
       SymX(rts_mkBool)                         \
       SymX(rts_mkChar)                         \
       SymX(rts_mkDouble)                       \
@@ -429,6 +450,7 @@ typedef struct _RtsSymbolVal {
       SymX(rts_mkInt64)                                \
       SymX(rts_mkInt8)                         \
       SymX(rts_mkPtr)                          \
+      SymX(rts_mkFunPtr)                       \
       SymX(rts_mkStablePtr)                    \
       SymX(rts_mkString)                       \
       SymX(rts_mkWord)                         \
@@ -436,18 +458,55 @@ typedef struct _RtsSymbolVal {
       SymX(rts_mkWord32)                       \
       SymX(rts_mkWord64)                       \
       SymX(rts_mkWord8)                                \
+      SymX(rts_unlock)                         \
+      SymX(rtsSupportsBoundThreads)            \
       SymX(run_queue_hd)                       \
+      SymX(__hscore_get_saved_termios)         \
+      SymX(__hscore_set_saved_termios)         \
       SymX(setProgArgv)                                \
+      SymX(startupHaskell)                     \
+      SymX(shutdownHaskell)                    \
       SymX(shutdownHaskellAndExit)             \
       SymX(stable_ptr_table)                   \
       SymX(stackOverflow)                      \
       SymX(stg_CAF_BLACKHOLE_info)             \
+      SymX(stg_BLACKHOLE_BQ_info)              \
+      SymX(awakenBlockedQueue)                 \
       SymX(stg_CHARLIKE_closure)               \
       SymX(stg_EMPTY_MVAR_info)                        \
       SymX(stg_IND_STATIC_info)                        \
       SymX(stg_INTLIKE_closure)                        \
       SymX(stg_MUT_ARR_PTRS_FROZEN_info)       \
       SymX(stg_WEAK_info)                       \
+      SymX(stg_ap_v_info)                      \
+      SymX(stg_ap_f_info)                      \
+      SymX(stg_ap_d_info)                      \
+      SymX(stg_ap_l_info)                      \
+      SymX(stg_ap_n_info)                      \
+      SymX(stg_ap_p_info)                      \
+      SymX(stg_ap_pv_info)                     \
+      SymX(stg_ap_pp_info)                     \
+      SymX(stg_ap_ppv_info)                    \
+      SymX(stg_ap_ppp_info)                    \
+      SymX(stg_ap_pppp_info)                   \
+      SymX(stg_ap_ppppp_info)                  \
+      SymX(stg_ap_pppppp_info)                 \
+      SymX(stg_ap_ppppppp_info)                        \
+      SymX(stg_ap_0_ret)                       \
+      SymX(stg_ap_v_ret)                       \
+      SymX(stg_ap_f_ret)                       \
+      SymX(stg_ap_d_ret)                       \
+      SymX(stg_ap_l_ret)                       \
+      SymX(stg_ap_n_ret)                       \
+      SymX(stg_ap_p_ret)                       \
+      SymX(stg_ap_pv_ret)                      \
+      SymX(stg_ap_pp_ret)                      \
+      SymX(stg_ap_ppv_ret)                     \
+      SymX(stg_ap_ppp_ret)                     \
+      SymX(stg_ap_pppp_ret)                    \
+      SymX(stg_ap_ppppp_ret)                   \
+      SymX(stg_ap_pppppp_ret)                  \
+      SymX(stg_ap_ppppppp_ret)                 \
       SymX(stg_ap_1_upd_info)                  \
       SymX(stg_ap_2_upd_info)                  \
       SymX(stg_ap_3_upd_info)                  \
@@ -473,9 +532,7 @@ typedef struct _RtsSymbolVal {
       SymX(stg_sel_7_upd_info)                 \
       SymX(stg_sel_8_upd_info)                 \
       SymX(stg_sel_9_upd_info)                 \
-      SymX(stg_seq_frame_info)                 \
       SymX(stg_upd_frame_info)                 \
-      SymX(__stg_update_PAP)                   \
       SymX(suspendThread)                      \
       SymX(takeMVarzh_fast)                    \
       SymX(timesIntegerzh_fast)                        \
@@ -497,35 +554,59 @@ typedef struct _RtsSymbolVal {
 #define RTS_LONG_LONG_SYMS /* nothing */
 #endif
 
-#ifdef ia64_TARGET_ARCH
-/* force these symbols to be present */
-#define RTS_EXTRA_SYMBOLS                      \
-      Sym(__divsf3)
-#elif defined(powerpc_TARGET_ARCH)
-#define RTS_EXTRA_SYMBOLS                                              \
+#ifdef HAVE_TERMIOS_H
+#define RTS_TERMIOS_SYMS                        \
+      Sym(saved_termios)
+#else
+#define RTS_TERMIOS_SYMS /* nothing */
+#endif
+
+// 64-bit support functions in libgcc.a
+#if defined(__GNUC__) && SIZEOF_VOID_P <= 4
+#define RTS_LIBGCC_SYMBOLS                     \
       Sym(__divdi3)                             \
       Sym(__udivdi3)                            \
       Sym(__moddi3)                             \
-      Sym(__umoddi3)                                                   \
-         Sym(__ashldi3)                                                        \
-         Sym(__ashrdi3)                                                        \
-         Sym(__lshrdi3)                                                        \
-         SymX(__eprintf)
+      Sym(__umoddi3)                           \
+      Sym(__ashldi3)                           \
+      Sym(__ashrdi3)                           \
+      Sym(__lshrdi3)                           \
+      Sym(__eprintf)
+#elif defined(ia64_TARGET_ARCH)
+#define RTS_LIBGCC_SYMBOLS                     \
+      Sym(__divdi3)                            \
+      Sym(__udivdi3)                            \
+      Sym(__moddi3)                            \
+      Sym(__umoddi3)                           \
+      Sym(__divsf3)                            \
+      Sym(__divdf3)
 #else
-#define RTS_EXTRA_SYMBOLS /* nothing */
+#define RTS_LIBGCC_SYMBOLS
+#endif
+
+#ifdef darwin_TARGET_OS
+      // Symbols that don't have a leading underscore
+      // on Mac OS X. They have to receive special treatment,
+      // see machoInitSymbolsWithoutUnderscore()
+#define RTS_MACHO_NOUNDERLINE_SYMBOLS          \
+      Sym(saveFP)                              \
+      Sym(restFP)
 #endif
 
 /* entirely bogus claims about types of these symbols */
-#define Sym(vvv)  extern void (vvv);
+#define Sym(vvv)  extern void vvv(void);
 #define SymX(vvv) /**/
+#define SymX_redirect(vvv,xxx) /**/
 RTS_SYMBOLS
 RTS_LONG_LONG_SYMS
-RTS_EXTRA_SYMBOLS
 RTS_POSIX_ONLY_SYMBOLS
 RTS_MINGW_ONLY_SYMBOLS
 RTS_CYGWIN_ONLY_SYMBOLS
+RTS_TERMIOS_SYMS
+RTS_LIBGCC_SYMBOLS
 #undef Sym
 #undef SymX
+#undef SymX_redirect
 
 #ifdef LEADING_UNDERSCORE
 #define MAYBE_LEADING_UNDERSCORE_STR(s) ("_" s)
@@ -537,13 +618,19 @@ RTS_CYGWIN_ONLY_SYMBOLS
                     (void*)(&(vvv)) },
 #define SymX(vvv) Sym(vvv)
 
+// SymX_redirect allows us to redirect references to one symbol to
+// another symbol.  See newCAF/newDynCAF for an example.
+#define SymX_redirect(vvv,xxx) \
+    { MAYBE_LEADING_UNDERSCORE_STR(#vvv), \
+      (void*)(&(xxx)) },
+
 static RtsSymbolVal rtsSyms[] = {
       RTS_SYMBOLS
       RTS_LONG_LONG_SYMS
-      RTS_EXTRA_SYMBOLS
       RTS_POSIX_ONLY_SYMBOLS
       RTS_MINGW_ONLY_SYMBOLS
       RTS_CYGWIN_ONLY_SYMBOLS
+      RTS_LIBGCC_SYMBOLS
       { 0, 0 } /* sentinel */
 };
 
@@ -611,6 +698,10 @@ initLinker( void )
        ghciInsertStrHashTable("(GHCi built-in symbols)",
                                symhash, sym->lbl, sym->addr);
     }
+#   if defined(OBJFORMAT_MACHO)
+    machoInitSymbolsWithoutUnderscore();
+#   endif
+
 #   if defined(OBJFORMAT_ELF) || defined(OBJFORMAT_MACHO)
     dl_prog_handle = dlopen(NULL, RTLD_LAZY);
 #   endif
@@ -662,6 +753,7 @@ addDLL( char *dll_name )
    initLinker();
 
    hdl= dlopen(dll_name, RTLD_NOW | RTLD_GLOBAL);
+
    if (hdl == NULL) {
       /* dlopen failed; return a ptr to the error msg. */
       errmsg = dlerror();
@@ -706,13 +798,13 @@ addDLL( char *dll_name )
         sprintf(buf, "%s.DRV", dll_name);      // KAA: allow loading of drivers (like winspool.drv)
         instance = LoadLibrary(buf);
         if (instance == NULL) {
-               free(buf);
+               stgFree(buf);
 
            /* LoadLibrary failed; return a ptr to the error msg. */
            return "addDLL: unknown error";
         }
    }
-   free(buf);
+   stgFree(buf);
 
    /* Add this DLL to the list of DLLs in which to search for symbols. */
    o_dll = stgMallocBytes( sizeof(OpenedDLL), "addDLL" );
@@ -740,8 +832,15 @@ lookupSymbol( char *lbl )
     val = lookupStrHashTable(symhash, lbl);
 
     if (val == NULL) {
-#       if defined(OBJFORMAT_ELF) || defined(OBJFORMAT_MACHO)
+#       if defined(OBJFORMAT_ELF)
        return dlsym(dl_prog_handle, lbl);
+#       elif defined(OBJFORMAT_MACHO)
+       if(NSIsSymbolNameDefined(lbl)) {
+           NSSymbol symbol = NSLookupAndBindSymbol(lbl);
+           return NSAddressOfSymbol(symbol);
+       } else {
+           return NULL;
+       }
 #       elif defined(OBJFORMAT_PEi386)
         OpenedDLL* o_dll;
         void* sym;
@@ -813,14 +912,16 @@ void ghci_enquire ( char* addr )
       for (i = 0; i < oc->n_symbols; i++) {
          sym = oc->symbols[i];
          if (sym == NULL) continue;
-         /* fprintf(stderr, "enquire %p %p\n", sym, oc->lochash); */
+         // fprintf(stderr, "enquire %p %p\n", sym, oc->lochash);
          a = NULL;
-         if (oc->lochash != NULL)
+         if (oc->lochash != NULL) {
             a = lookupStrHashTable(oc->lochash, sym);
-         if (a == NULL)
+        }
+         if (a == NULL) {
             a = lookupStrHashTable(symhash, sym);
+        }
          if (a == NULL) {
-            /* fprintf(stderr, "ghci_enquire: can't find %s\n", sym); */
+            // fprintf(stderr, "ghci_enquire: can't find %s\n", sym);
          }
          else if (addr-DELTA <= a && a <= addr+DELTA) {
             fprintf(stderr, "%p + %3d  ==  `%s'\n", addr, a - addr, sym);
@@ -886,7 +987,7 @@ loadObj( char *path )
 #  elif defined(OBJFORMAT_MACHO)
    oc->formatName = "Mach-O";
 #  else
-   free(oc);
+   stgFree(oc);
    barf("loadObj: not implemented on this platform");
 #  endif
 
@@ -953,6 +1054,11 @@ loadObj( char *path )
 
 #endif /* USE_MMAP */
 
+#  if defined(OBJFORMAT_MACHO)
+   r = ocAllocateJumpIslands_MachO ( oc );
+   if (!r) { return r; }
+#endif
+
    /* verify the in-memory image */
 #  if defined(OBJFORMAT_ELF)
    r = ocVerifyImage_ELF ( oc );
@@ -1051,14 +1157,14 @@ unloadObj( char *path )
 
            /* We're going to leave this in place, in case there are
               any pointers from the heap into it: */
-           /* free(oc->image); */
-           free(oc->fileName);
-           free(oc->symbols);
-           free(oc->sections);
+           /* stgFree(oc->image); */
+           stgFree(oc->fileName);
+           stgFree(oc->symbols);
+           stgFree(oc->sections);
            /* The local hash table should have been freed at the end
                of the ocResolve_ call on it. */
             ASSERT(oc->lochash == NULL);
-           free(oc);
+           stgFree(oc);
            return 1;
        }
     }
@@ -1812,11 +1918,16 @@ ocResolve_PEi386 ( ObjectCode* oc )
         * real count can be found in the first reloc entry.
          *
         * See Section 4.1 (last para) of the PE spec (rev6.0).
+        *
+        * Nov2003 update: the GNU linker still doesn't correctly 
+        * handle the generation of relocatable object files with 
+        * overflown relocations. Hence the output to warn of potential 
+        * troubles.
         */
         COFF_reloc* rel = (COFF_reloc*)
                            myindex ( sizeof_COFF_reloc, reltab, 0 );
        noRelocs = rel->VirtualAddress;
-       fprintf(stderr, "Overflown relocs: %u\n", noRelocs);
+       fprintf(stderr, "WARNING: Overflown relocation field (# relocs found: %u)\n", noRelocs); fflush(stderr);
        j = 1;
       } else {
        noRelocs = sectab_i->NumberOfRelocations;
@@ -1928,6 +2039,9 @@ ocResolve_PEi386 ( ObjectCode* oc )
 #  define ELF_TARGET_SPARC  /* Used inside <elf.h> */
 #elif defined(i386_TARGET_ARCH)
 #  define ELF_TARGET_386    /* Used inside <elf.h> */
+#elif defined(x86_64_TARGET_ARCH)
+#  define ELF_TARGET_X64_64
+#  define ELF_64BIT
 #elif defined (ia64_TARGET_ARCH)
 #  define ELF_TARGET_IA64   /* Used inside <elf.h> */
 #  define ELF_64BIT
@@ -1936,7 +2050,15 @@ ocResolve_PEi386 ( ObjectCode* oc )
 #  define ELF_NEED_PLT      /* needs Procedure Linkage Tables */
 #endif
 
+#if !defined(openbsd_TARGET_OS)
 #include <elf.h>
+#else
+/* openbsd elf has things in different places, with diff names */
+#include <elf_abi.h>
+#include <machine/reloc.h>
+#define R_386_32    RELOC_32
+#define R_386_PC32  RELOC_PC32
+#endif
 
 /*
  * Define a set of types which can be used for both ELF32 and ELF64
@@ -1968,11 +2090,19 @@ ocResolve_PEi386 ( ObjectCode* oc )
 #define Elf_Sym     Elf32_Sym
 #define Elf_Rel     Elf32_Rel
 #define Elf_Rela    Elf32_Rela
+#ifndef ELF_ST_TYPE
 #define ELF_ST_TYPE ELF32_ST_TYPE
+#endif
+#ifndef ELF_ST_BIND
 #define ELF_ST_BIND ELF32_ST_BIND
+#endif
+#ifndef ELF_R_TYPE
 #define ELF_R_TYPE  ELF32_R_TYPE
+#endif
+#ifndef ELF_R_SYM
 #define ELF_R_SYM   ELF32_R_SYM
 #endif
+#endif
 
 
 /*
@@ -2633,8 +2763,8 @@ do_Elf_Rela_relocations ( ObjectCode* oc, char* ehdrC,
 #ifdef ELF_FUNCTION_DESC
            /* If a function, already a function descriptor - we would
               have to copy it to add an offset. */
-            if (S && ELF_ST_TYPE(sym.st_info) == STT_FUNC)
-               assert(A == 0);
+            if (S && (ELF_ST_TYPE(sym.st_info) == STT_FUNC) && (A != 0))
+               belch("%s: function %s with addend %p", oc->fileName, symbol, (void *)A);
 #endif
         }
          if (!S) {
@@ -2693,6 +2823,9 @@ do_Elf_Rela_relocations ( ObjectCode* oc, char* ehdrC,
         case R_IA64_FPTR64LSB:
            *pP = value;
            break;
+        case R_IA64_PCREL64LSB:
+           *pP = value - P;
+           break;
         case R_IA64_SEGREL64LSB:
            addr = findElfSegment(ehdrC, value);
            *pP = value - addr;
@@ -2701,6 +2834,7 @@ do_Elf_Rela_relocations ( ObjectCode* oc, char* ehdrC,
            ia64_reloc_gprel22(P, value);
            break;
         case R_IA64_LTOFF22:
+        case R_IA64_LTOFF22X:
         case R_IA64_LTOFF_FPTR22:
            addr = allocateGOTEntry(value);
            ia64_reloc_gprel22(P, addr);
@@ -2708,6 +2842,10 @@ do_Elf_Rela_relocations ( ObjectCode* oc, char* ehdrC,
         case R_IA64_PCREL21B:
            ia64_reloc_pcrel21(P, S, oc);
            break;
+        case R_IA64_LDXMOV:
+           /* This goes with R_IA64_LTOFF22X and points to the load to
+            * convert into a move.  We don't implement relaxation. */
+           break;
 #        endif
          default:
             belch("%s: unhandled ELF relocation(RelA) type %d\n",
@@ -2872,25 +3010,77 @@ ia64_reloc_pcrel21(Elf_Addr target, Elf_Addr value, ObjectCode *oc)
 #if defined(OBJFORMAT_MACHO)
 
 /*
-  Initial support for MachO linking on Darwin/MacOS X on PowerPC chips
+  Support for MachO linking on Darwin/MacOS X on PowerPC chips
   by Wolfgang Thaller (wolfgang.thaller@gmx.net)
   
   I hereby formally apologize for the hackish nature of this code.
   Things that need to be done:
-  *) get common symbols and .bss sections to work properly.
-       Haskell modules seem to work, but C modules can cause problems
   *) implement ocVerifyImage_MachO
-  *) add more sanity checks. The current code just has to segfault if there's a
-     broken .o file.
+  *) add still more sanity checks.
 */
 
+
+/*
+  ocAllocateJumpIslands_MachO
+  
+  Allocate additional space at the end of the object file image to make room
+  for jump islands.
+  
+  PowerPC relative branch instructions have a 24 bit displacement field.
+  As PPC code is always 4-byte-aligned, this yields a +-32MB range.
+  If a particular imported symbol is outside this range, we have to redirect
+  the jump to a short piece of new code that just loads the 32bit absolute
+  address and jumps there.
+  This function just allocates space for one 16 byte jump island for every
+  undefined symbol in the object file. The code for the islands is filled in by
+  makeJumpIsland below.
+*/
+
+static const int islandSize = 16;
+
+static int ocAllocateJumpIslands_MachO(ObjectCode* oc)
+{
+    char *image = (char*) oc->image;
+    struct mach_header *header = (struct mach_header*) image;
+    struct load_command *lc = (struct load_command*) (image + sizeof(struct mach_header));
+    unsigned i;
+
+    for(i=0;i<header->ncmds;i++)
+    {
+       if(lc->cmd == LC_DYSYMTAB)
+        {
+           struct dysymtab_command *dsymLC = (struct dysymtab_command*) lc;
+            unsigned long nundefsym = dsymLC->nundefsym;
+            oc->island_start_symbol = dsymLC->iundefsym;
+            oc->n_islands = nundefsym;
+            
+            if(nundefsym > 0)
+            {
+#ifdef USE_MMAP
+                #error ocAllocateJumpIslands_MachO doesnt want USE_MMAP to be defined
+#else
+                oc->image = stgReallocBytes(
+                    image, oc->fileSize + islandSize * nundefsym,
+                    "ocAllocateJumpIslands_MachO");
+#endif                    
+                oc->jump_islands = oc->image + oc->fileSize;
+                memset(oc->jump_islands, 0, islandSize * nundefsym);
+            }
+            
+            break;  // there can be only one LC_DSYMTAB
+        }
+       lc = (struct load_command *) ( ((char*)lc) + lc->cmdsize );
+    }
+    return 1;
+}
+
 static int ocVerifyImage_MachO(ObjectCode* oc)
 {
     // FIXME: do some verifying here
     return 1;
 }
 
-static void resolveImports(
+static int resolveImports(
     ObjectCode* oc,
     char *image,
     struct symtab_command *symLC,
@@ -2916,25 +3106,75 @@ static void resolveImports(
            addr = lookupSymbol(nm);
        if(!addr)
        {
-           fprintf(stderr, "not found: %s\n", nm);
-           abort();
+           belch("\n%s: unknown symbol `%s'", oc->fileName, nm);
+           return 0;
        }
        ASSERT(addr);
+       checkProddableBlock(oc,((void**)(image + sect->offset)) + i);
        ((void**)(image + sect->offset))[i] = addr;
     }
+    
+    return 1;
 }
 
-static void relocateSection(char *image, 
+static void* makeJumpIsland(
+    ObjectCode* oc,
+    unsigned long symbolNumber,
+    void* target)
+{
+    if(symbolNumber < oc->island_start_symbol ||
+        symbolNumber - oc->island_start_symbol > oc->n_islands)
+        return NULL;
+    symbolNumber -= oc->island_start_symbol;
+    
+    void *island = (void*) ((char*)oc->jump_islands + islandSize * symbolNumber);
+    unsigned long *p = (unsigned long*) island;
+    
+        // lis r12, hi16(target)
+    *p++ = 0x3d800000 | ( ((unsigned long) target) >> 16 );
+        // ori r12, r12, lo16(target)
+    *p++ = 0x618c0000 | ( ((unsigned long) target) & 0xFFFF ); 
+        // mtctr r12
+    *p++ = 0x7d8903a6;
+        // bctr
+    *p++ = 0x4e800420;
+    
+    return (void*) island;
+}
+
+static char* relocateAddress(
+    ObjectCode* oc,
+    int nSections,
+    struct section* sections,
+    unsigned long address)
+{  
+    int i;
+    for(i = 0; i < nSections; i++)
+    {
+        if(sections[i].addr <= address
+            && address < sections[i].addr + sections[i].size)
+        {
+            return oc->image + sections[i].offset + address - sections[i].addr;
+        }
+    }
+    barf("Invalid Mach-O file:"
+         "Address out of bounds while relocating object file");
+    return NULL;
+}
+
+static int relocateSection(
+    ObjectCode* oc,
+    char *image, 
     struct symtab_command *symLC, struct nlist *nlist,
-    struct section* sections, struct section *sect)
+    int nSections, struct section* sections, struct section *sect)
 {
     struct relocation_info *relocs;
     int i,n;
     
     if(!strcmp(sect->sectname,"__la_symbol_ptr"))
-       return;
+       return 1;
     else if(!strcmp(sect->sectname,"__nl_symbol_ptr"))
-       return;
+       return 1;
 
     n = sect->nreloc;
     relocs = (struct relocation_info*) (image + sect->reloff);
@@ -2948,11 +3188,55 @@ static void relocateSection(char *image,
            
            if(!scat->r_pcrel)
            {
-               if(scat->r_length == 2 && scat->r_type == GENERIC_RELOC_VANILLA)
+               if(scat->r_length == 2)
                {
-                   unsigned long* word = (unsigned long*) (image + sect->offset + scat->r_address);
+                   unsigned long word = 0;
+                   unsigned long* wordPtr = (unsigned long*) (image + sect->offset + scat->r_address);
+                   checkProddableBlock(oc,wordPtr);
                    
-                   *word = scat->r_value + sect->offset + ((long) image);
+                   // Step 1: Figure out what the relocated value should be
+                   if(scat->r_type == GENERIC_RELOC_VANILLA)
+                   {
+                       word = scat->r_value + sect->offset + ((long) image);
+                   }
+                   else if(scat->r_type == PPC_RELOC_SECTDIFF
+                       || scat->r_type == PPC_RELOC_LO16_SECTDIFF
+                       || scat->r_type == PPC_RELOC_HI16_SECTDIFF
+                       || scat->r_type == PPC_RELOC_HA16_SECTDIFF)
+                   {
+                       struct scattered_relocation_info *pair =
+                               (struct scattered_relocation_info*) &relocs[i+1];
+                               
+                       if(!pair->r_scattered || pair->r_type != PPC_RELOC_PAIR)
+                           barf("Invalid Mach-O file: "
+                                "PPC_RELOC_*_SECTDIFF not followed by PPC_RELOC_PAIR");
+                       
+                       word = (unsigned long)
+                              (relocateAddress(oc, nSections, sections, scat->r_value)
+                             - relocateAddress(oc, nSections, sections, pair->r_value));
+                       i++;
+                   }
+                   else
+                       continue;  // ignore the others
+
+                    if(scat->r_type == GENERIC_RELOC_VANILLA
+                        || scat->r_type == PPC_RELOC_SECTDIFF)
+                    {
+                        *wordPtr = word;
+                    }
+                    else if(scat->r_type == PPC_RELOC_LO16_SECTDIFF)
+                    {
+                        ((unsigned short*) wordPtr)[1] = word & 0xFFFF;
+                    }
+                    else if(scat->r_type == PPC_RELOC_HI16_SECTDIFF)
+                    {
+                        ((unsigned short*) wordPtr)[1] = (word >> 16) & 0xFFFF;
+                    }
+                    else if(scat->r_type == PPC_RELOC_HA16_SECTDIFF)
+                    {
+                        ((unsigned short*) wordPtr)[1] = ((word >> 16) & 0xFFFF)
+                            + ((word & (1<<15)) ? 1 : 0);
+                    }
                }
            }
            
@@ -2964,11 +3248,14 @@ static void relocateSection(char *image,
            if(reloc->r_pcrel && !reloc->r_extern)
                continue;
                
-           if(!reloc->r_pcrel && reloc->r_length == 2)
+           if(reloc->r_length == 2)
            {
-               unsigned long word;
-
+               unsigned long word = 0;
+                unsigned long jumpIsland = 0;
+                long offsetToJumpIsland;
+                
                unsigned long* wordPtr = (unsigned long*) (image + sect->offset + reloc->r_address);
+               checkProddableBlock(oc,wordPtr);
                
                if(reloc->r_type == GENERIC_RELOC_VANILLA)
                {
@@ -2989,6 +3276,12 @@ static void relocateSection(char *image,
                    word = ((unsigned short*) wordPtr)[1] << 16;
                    word += ((short)relocs[i+1].r_address & (short)0xFFFF);
                }
+               else if(reloc->r_type == PPC_RELOC_BR24)
+               {
+                   word = *wordPtr;
+                   word = (word & 0x03FFFFFC) | (word & 0x02000000) ? 0xFC000000 : 0;
+               }
+
 
                if(!reloc->r_extern)
                {
@@ -3004,7 +3297,22 @@ static void relocateSection(char *image,
                    struct nlist *symbol = &nlist[reloc->r_symbolnum];
                    char *nm = image + symLC->stroff + symbol->n_un.n_strx;
                    word = (unsigned long) (lookupSymbol(nm));
-                   ASSERT(word);
+                   if(!word)
+                   {
+                       belch("\nunknown symbol `%s'", nm);
+                       return 0;
+                   }
+                   
+                   if(reloc->r_pcrel)
+                    {
+                        jumpIsland = (long) makeJumpIsland(oc,reloc->r_symbolnum,(void*)word);
+                       word -= ((long)image) + sect->offset + reloc->r_address;
+                        if(jumpIsland != 0)
+                        {
+                            offsetToJumpIsland = jumpIsland
+                                - (((long)image) + sect->offset + reloc->r_address);
+                        }
+                    }
                }
                
                if(reloc->r_type == GENERIC_RELOC_VANILLA)
@@ -3028,13 +3336,30 @@ static void relocateSection(char *image,
                        + ((word & (1<<15)) ? 1 : 0);
                    i++; continue;
                }
-               continue;
+               else if(reloc->r_type == PPC_RELOC_BR24)
+               {
+                    if((long)word > (long)0x01FFFFFF || (long)word < (long)0xFFE00000)
+                    {
+                        // The branch offset is too large.
+                        // Therefore, we try to use a jump island.
+                        if(jumpIsland == 0)
+                            barf("unconditional relative branch out of range: "
+                                 "no jump island available");
+                            
+                        word = offsetToJumpIsland;
+                        if((long)word > (long)0x01FFFFFF || (long)word < (long)0xFFE00000)
+                            barf("unconditional relative branch out of range: "
+                                 "jump island out of range");
+                    }
+                   *wordPtr = (*wordPtr & 0xFC000003) | (word & 0x03FFFFFC);
+                   continue;
+               }
            }
-           fprintf(stderr, "unknown reloc\n");
-           abort();
-           ASSERT(2 + 2 == 5);
+           barf("\nunknown relocation %d",reloc->r_type);
+           return 0;
        }
     }
+    return 1;
 }
 
 static int ocGetNames_MachO(ObjectCode* oc)
@@ -3044,7 +3369,7 @@ static int ocGetNames_MachO(ObjectCode* oc)
     struct load_command *lc = (struct load_command*) (image + sizeof(struct mach_header));
     unsigned i,curSymbol;
     struct segment_command *segLC = NULL;
-    struct section *sections, *la_ptrs = NULL, *nl_ptrs = NULL;
+    struct section *sections;
     struct symtab_command *symLC = NULL;
     struct dysymtab_command *dsymLC = NULL;
     struct nlist *nlist;
@@ -3068,13 +3393,17 @@ static int ocGetNames_MachO(ObjectCode* oc)
 
     for(i=0;i<segLC->nsects;i++)
     {
-       if(!strcmp(sections[i].sectname,"__la_symbol_ptr"))
-           la_ptrs = &sections[i];
-       else if(!strcmp(sections[i].sectname,"__nl_symbol_ptr"))
-           nl_ptrs = &sections[i];
-           
-           // for now, only add __text and __const to the sections table
-       else if(!strcmp(sections[i].sectname,"__text"))
+        if(sections[i].size == 0)
+            continue;
+        
+        if((sections[i].flags & SECTION_TYPE) == S_ZEROFILL)
+        {
+            char * zeroFillArea = stgCallocBytes(1,sections[i].size,
+                                      "ocGetNames_MachO(common symbols)");
+            sections[i].offset = zeroFillArea - image;
+        }
+    
+       if(!strcmp(sections[i].sectname,"__text"))
            addSection(oc, SECTIONKIND_CODE_OR_RODATA, 
                (void*) (image + sections[i].offset),
                (void*) (image + sections[i].offset + sections[i].size));
@@ -3086,6 +3415,14 @@ static int ocGetNames_MachO(ObjectCode* oc)
            addSection(oc, SECTIONKIND_RWDATA, 
                (void*) (image + sections[i].offset),
                (void*) (image + sections[i].offset + sections[i].size));
+       else if(!strcmp(sections[i].sectname,"__bss")
+               || !strcmp(sections[i].sectname,"__common"))
+           addSection(oc, SECTIONKIND_RWDATA, 
+               (void*) (image + sections[i].offset),
+               (void*) (image + sections[i].offset + sections[i].size));
+
+        addProddableBlock(oc, (void*) (image + sections[i].offset),
+                                        sections[i].size);
     }
 
        // count external symbols defined here
@@ -3194,19 +3531,60 @@ static int ocResolve_MachO(ObjectCode* oc)
     indirectSyms = (unsigned long*) (image + dsymLC->indirectsymoff);
 
     if(la_ptrs)
-       resolveImports(oc,image,symLC,la_ptrs,indirectSyms,nlist);
+       if(!resolveImports(oc,image,symLC,la_ptrs,indirectSyms,nlist))
+           return 0;
     if(nl_ptrs)
-       resolveImports(oc,image,symLC,nl_ptrs,indirectSyms,nlist);
+       if(!resolveImports(oc,image,symLC,nl_ptrs,indirectSyms,nlist))
+           return 0;
     
     for(i=0;i<segLC->nsects;i++)
     {
-       relocateSection(image,symLC,nlist,sections,&sections[i]);
+       if(!relocateSection(oc,image,symLC,nlist,segLC->nsects,sections,&sections[i]))
+           return 0;
     }
 
     /* Free the local symbol table; we won't need it again. */
     freeHashTable(oc->lochash, NULL);
     oc->lochash = NULL;
+    
+    /*
+        Flush the data & instruction caches.
+        Because the PPC has split data/instruction caches, we have to
+        do that whenever we modify code at runtime.
+    */
+    {
+        int n = (oc->fileSize + islandSize * oc->n_islands) / 4;
+        unsigned long *p = (unsigned long*)oc->image;
+        while(n--)
+        {
+            __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0"
+                                : : "r" (p));
+            p++;
+        }
+        __asm__ volatile ("sync\n\tisync");
+    }
     return 1;
 }
 
+/*
+ * The Mach-O object format uses leading underscores. But not everywhere.
+ * There is a small number of runtime support functions defined in
+ * libcc_dynamic.a whose name does not have a leading underscore.
+ * As a consequence, we can't get their address from C code.
+ * We have to use inline assembler just to take the address of a function.
+ * Yuck.
+ */
+
+static void machoInitSymbolsWithoutUnderscore()
+{
+    void *p;
+
+#undef Sym    
+#define Sym(x)                                         \
+    __asm__ ("lis %0,hi16(" #x ")\n\tori %0,%0,lo16(" #x ")" : "=r" (p));      \
+    ghciInsertStrHashTable("(GHCi built-in symbols)", symhash, #x, p);
+    
+    RTS_MACHO_NOUNDERLINE_SYMBOLS
+
+}
 #endif