Merge branch 'sverker/hipe-code-loadnpurge/OTP-13968'

* sverker/hipe-code-loadnpurge: (35 commits)
  erts: Cleanup dead code
  kernel,hipe: Fix dialyzer warnings
  erts: Replace unsafe Module.first_hipe_ref
  erts: Disable DBG_TRACE_MFA for debug build
  kernel: Fix code_SUITE:upgrade for non-hipe
  erts: Cleanup hipe trampoline code
  erts: Remove dead alloc stats in hipe_amd64.c
  erts: Remove code_SUITE:make_stub and make_stub_many_funs
  erts: Let code:make_stub_module raise 'notsup'
  erts: Fix bug in stack walk on risc
  erts: Fix old leak for ppc hipe code
  erts: Fix old leak for arm hipe code
  erts: Fix old leak of sparc hipe code
  erts: Fix old leak of hipe code on x86 32-bit
  erts: Enable exec_alloc for all hipe architectures
  erts: Remove debug printout for hipe loader state
  erts: Free hipe_refs and hipe_sdesc of a failed load
  erts: Refactor out hipe_purge_refs/sdesc
  erts: Refactor hipe_loader_state_dtor into a true destructor
  hipe: TRY fix llvm external calls to own module
  ...

13 files changed, 401 insertions, 112 deletions

diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index bce57fb7c1..099abad4e1 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -36,6 +36,12 @@
 #include "erl_nif.h"
 #include "erl_bits.h"
 #include "erl_thr_progress.h"
+#ifdef HIPE
+#  include "hipe_bif0.h"
+#  define IF_HIPE(X) (X)
+#else
+#  define IF_HIPE(X) (0)
+#endif
 
 #ifdef HIPE
 #  include "hipe_stack.h"
@@ -143,8 +149,17 @@ BIF_RETTYPE code_is_module_native_1(BIF_ALIST_1)
 
 BIF_RETTYPE code_make_stub_module_3(BIF_ALIST_3)
 {
+#if !defined(HIPE)
+    BIF_ERROR(BIF_P, EXC_NOTSUP);
+#else
     Module* modp;
-    Eterm res;
+    Eterm res, mod;
+
+    if (!ERTS_TERM_IS_MAGIC_BINARY(BIF_ARG_1) ||
+	is_not_atom(mod = erts_module_for_prepared_code
+		    (((ProcBin*)binary_val(BIF_ARG_1))->val))) {
+	BIF_ERROR(BIF_P, BADARG);
+    }
 
     if (!erts_try_seize_code_write_permission(BIF_P)) {
 	ERTS_BIF_YIELD3(bif_export[BIF_code_make_stub_module_3],
@@ -154,7 +169,7 @@ BIF_RETTYPE code_make_stub_module_3(BIF_ALIST_3)
     erts_smp_proc_unlock(BIF_P, ERTS_PROC_LOCK_MAIN);
     erts_smp_thr_progress_block();
 
-    modp = erts_get_module(BIF_ARG_1, erts_active_code_ix());
+    modp = erts_get_module(mod, erts_active_code_ix());
 
     if (modp && modp->curr.num_breakpoints > 0) {
 	ASSERT(modp->curr.code_hdr != NULL);
@@ -166,9 +181,12 @@ BIF_RETTYPE code_make_stub_module_3(BIF_ALIST_3)
 
     res = erts_make_stub_module(BIF_P, BIF_ARG_1, BIF_ARG_2, BIF_ARG_3);
 
-    if (res == BIF_ARG_1) {
+    if (res == mod) {
 	erts_end_staging_code_ix();
 	erts_commit_staging_code_ix();
+        if (!modp)
+	    modp = erts_get_module(mod, erts_active_code_ix());
+        hipe_redirect_to_module(modp);
     }
     else {
 	erts_abort_staging_code_ix();
@@ -177,6 +195,7 @@ BIF_RETTYPE code_make_stub_module_3(BIF_ALIST_3)
     erts_smp_proc_lock(BIF_P, ERTS_PROC_LOCK_MAIN);
     erts_release_code_write_permission();
     return res;
+#endif
 }
 
 BIF_RETTYPE
@@ -241,7 +260,7 @@ struct m {
     Uint exception;
 };
 
-static Eterm staging_epilogue(Process* c_p, int, Eterm res, int, struct m*, int);
+static Eterm staging_epilogue(Process* c_p, int, Eterm res, int, struct m*, int, int);
 #ifdef ERTS_SMP
 static void smp_code_ix_commiter(void*);
 
@@ -377,8 +396,9 @@ finish_loading_1(BIF_ALIST_1)
     for (i = 0; i < n; i++) {
 	if (p[i].modp->curr.num_breakpoints > 0 ||
 	    p[i].modp->curr.num_traced_exports > 0 ||
-	    erts_is_default_trace_enabled()) {
-	    /* tracing involved, fallback with thread blocking */
+	    erts_is_default_trace_enabled() ||
+	    IF_HIPE(hipe_need_blocking(p[i].modp))) {
+	    /* tracing or hipe need thread blocking */
 	    erts_smp_proc_unlock(BIF_P, ERTS_PROC_LOCK_MAIN);
 	    erts_smp_thr_progress_block();
 	    is_blocking = 1;
@@ -436,32 +456,36 @@ finish_loading_1(BIF_ALIST_1)
     }
 
 done:
-    return staging_epilogue(BIF_P, do_commit, res, is_blocking, p, n);
+    return staging_epilogue(BIF_P, do_commit, res, is_blocking, p, n, 1);
 }
 
 static Eterm
 staging_epilogue(Process* c_p, int commit, Eterm res, int is_blocking,
-		 struct m* loaded, int nloaded)
+		 struct m* mods, int nmods, int free_mods)
 {    
 #ifdef ERTS_SMP
     if (is_blocking || !commit)
 #endif
     {
 	if (commit) {
+	    int i;
 	    erts_end_staging_code_ix();
 	    erts_commit_staging_code_ix();
-	    if (loaded) {
-		int i;
-		for (i=0; i < nloaded; i++) {		
-		    set_default_trace_pattern(loaded[i].module);
+
+	    for (i=0; i < nmods; i++) {
+		if (mods[i].modp->curr.code_hdr) {
+		    set_default_trace_pattern(mods[i].module);
 		}
+	      #ifdef HIPE
+		hipe_redirect_to_module(mods[i].modp);
+	      #endif
 	    }
 	}
 	else {
 	    erts_abort_staging_code_ix();
 	}
-	if (loaded) {
-	    erts_free(ERTS_ALC_T_LOADER_TMP, loaded);
+	if (free_mods) {
+	    erts_free(ERTS_ALC_T_LOADER_TMP, mods);
 	}
 	if (is_blocking) {
 	    erts_smp_thr_progress_unblock();
@@ -474,8 +498,8 @@ staging_epilogue(Process* c_p, int commit, Eterm res, int is_blocking,
     else {
 	ASSERT(is_value(res));
 
-	if (loaded) {
-	    erts_free(ERTS_ALC_T_LOADER_TMP, loaded);
+	if (free_mods) {
+	    erts_free(ERTS_ALC_T_LOADER_TMP, mods);
 	}
 	erts_end_staging_code_ix();
 	/*
@@ -653,8 +677,9 @@ BIF_RETTYPE delete_module_1(BIF_ALIST_1)
 	}
 	else {
 	    if (modp->curr.num_breakpoints > 0 ||
-		modp->curr.num_traced_exports > 0) {
-		/* we have tracing, retry single threaded */
+		modp->curr.num_traced_exports > 0 ||
+		IF_HIPE(hipe_need_blocking(modp))) {
+		/* tracing or hipe need to go single threaded */
 		erts_smp_proc_unlock(BIF_P, ERTS_PROC_LOCK_MAIN);
 		erts_smp_thr_progress_block();
 		is_blocking = 1;
@@ -668,7 +693,14 @@ BIF_RETTYPE delete_module_1(BIF_ALIST_1)
 	    success = 1;
 	}
     }
-    return staging_epilogue(BIF_P, success, res, is_blocking, NULL, 0);  
+    {
+	struct m mod;
+	Eterm retval;
+	mod.module = BIF_ARG_1;
+	mod.modp = modp;
+	retval = staging_epilogue(BIF_P, success, res, is_blocking, &mod, 1, 0);
+	return retval;
+    }
 }
 
 BIF_RETTYPE module_loaded_1(BIF_ALIST_1)
@@ -809,6 +841,9 @@ BIF_RETTYPE finish_after_on_load_2(BIF_ALIST_2)
 	}
 	modp->curr.code_hdr->on_load_function_ptr = NULL;
 	set_default_trace_pattern(BIF_ARG_1);
+      #ifdef HIPE
+        hipe_redirect_to_module(modp);
+      #endif
     } else if (BIF_ARG_2 == am_false) {
 	int i;
 
@@ -1037,6 +1072,10 @@ check_process_code(Process* rp, Module* modp, int *redsp, int fcalls)
     char* mod_start;
     Uint mod_size;
     Eterm* sp;
+#ifdef HIPE
+    void *nat_start = NULL;
+    Uint nat_size = 0;
+#endif
 
     *redsp += 1;
 
@@ -1066,6 +1105,20 @@ check_process_code(Process* rp, Module* modp, int *redsp, int fcalls)
 	}
     }
 
+#ifdef HIPE
+    /*
+     * Check all continuation pointers stored on the native stack if the module
+     * has native code.
+     */
+    if (modp->old.hipe_code) {
+	nat_start = modp->old.hipe_code->text_segment;
+	nat_size = modp->old.hipe_code->text_segment_size;
+	if (nat_size && nstack_any_cps_in_segment(rp, nat_start, nat_size)) {
+	    return am_true;
+	}
+    }
+#endif
+
     /* 
      * Check all continuation pointers stored in stackdump
      * and clear exception stackdump if there is a pointer
@@ -1082,8 +1135,16 @@ check_process_code(Process* rp, Module* modp, int *redsp, int fcalls)
 	    rp->ftrace = NIL;
 	} else {
 	    int i;
+	    char *area_start = mod_start;
+	    Uint area_size = mod_size;
+#ifdef HIPE
+	    if (rp->freason & EXF_NATIVE) {
+		area_start = nat_start;
+		area_size = nat_size;
+	    }
+#endif
 	    for (i = 0;  i < s->depth;  i++) {
-		if (ErtsInArea(s->trace[i], mod_start, mod_size)) {
+		if (ErtsInArea(s->trace[i], area_start, area_size)) {
 		    rp->freason = EXC_NULL;
 		    rp->fvalue = NIL;
 		    rp->ftrace = NIL;
@@ -1619,15 +1680,18 @@ BIF_RETTYPE erts_internal_purge_module_2(BIF_ALIST_2)
 		/*
 		 * Unload any NIF library
 		 */
-		if (modp->old.nif != NULL) {
+		if (modp->old.nif != NULL
+		    || IF_HIPE(hipe_purge_need_blocking(modp))) {
 		    /* ToDo: Do unload nif without blocking */
 		    erts_rwunlock_old_code(code_ix);
 		    erts_smp_proc_unlock(BIF_P, ERTS_PROC_LOCK_MAIN);
 		    erts_smp_thr_progress_block();
 		    is_blocking = 1;
 		    erts_rwlock_old_code(code_ix);
-		    erts_unload_nif(modp->old.nif);
-		    modp->old.nif = NULL;
+		    if (modp->old.nif) {
+		      erts_unload_nif(modp->old.nif);
+		      modp->old.nif = NULL;
+		    }
 		}
 
 		/*
@@ -1646,7 +1710,9 @@ BIF_RETTYPE erts_internal_purge_module_2(BIF_ALIST_2)
 		modp->old.code_length = 0;
 		modp->old.catches = BEAM_CATCHES_NIL;
 		erts_remove_from_ranges(code);
-
+#ifdef HIPE
+		hipe_purge_module(modp, is_blocking);
+#endif
 		ERTS_BIF_PREP_RET(ret, am_true);
 	    }
 
@@ -1719,6 +1785,8 @@ delete_code(Module* modp)
 			 (BeamInstr) BeamOp(op_i_generic_breakpoint)) {
 		    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
 		    ASSERT(modp->curr.num_traced_exports > 0);
+		    DBG_TRACE_MFA_P(&ep->info.mfa,
+				  "export trace cleared, code_ix=%d", code_ix);
 		    erts_clear_export_break(modp, &ep->info);
 		}
 		else ASSERT(ep->beam[0] == (BeamInstr) em_call_error_handler
@@ -1727,17 +1795,15 @@ delete_code(Module* modp)
 	    ep->addressv[code_ix] = ep->beam;
 	    ep->beam[0] = (BeamInstr) em_call_error_handler;
 	    ep->beam[1] = 0;
+	    DBG_TRACE_MFA_P(&ep->info.mfa,
+			    "export invalidation, code_ix=%d", code_ix);
 	}
     }
 
     ASSERT(modp->curr.num_breakpoints == 0);
     ASSERT(modp->curr.num_traced_exports == 0);
     modp->old = modp->curr;
-    modp->curr.code_hdr = NULL;
-    modp->curr.code_length = 0;
-    modp->curr.catches = BEAM_CATCHES_NIL;
-    modp->curr.nif = NULL;
-
+    erts_module_instance_init(&modp->curr);
 }
 
 
@@ -1751,9 +1817,11 @@ beam_make_current_old(Process *c_p, ErtsProcLocks c_p_locks, Eterm module)
      * if not, delete old code; error if old code already exists.
      */
 
-    if (modp->curr.code_hdr && modp->old.code_hdr)  {
-	return am_not_purged;
-    } else if (!modp->old.code_hdr) { /* Make the current version old. */
+    if (modp->curr.code_hdr) {
+	if (modp->old.code_hdr)  {
+	    return am_not_purged;
+	}
+	/* Make the current version old. */
 	delete_code(modp);
     }
     return NIL;
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 3b02df8bac..84e2d42af6 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -6121,6 +6121,7 @@ call_error_handler(Process* p, ErtsCodeMFA* mfa, Eterm* reg, Eterm func)
     Uint sz;
     int i;
 
+    DBG_TRACE_MFA_P(mfa, "call_error_handler");
     /*
      * Search for the error_handler module.
      */
@@ -6676,9 +6677,6 @@ new_fun(Process* p, Eterm* reg, ErlFunEntry* fe, int num_free)
     funp->fe = fe;
     funp->num_free = num_free;
     funp->creator = p->common.id;
-#ifdef HIPE
-    funp->native_address = fe->native_address;
-#endif
     funp->arity = (int)fe->address[-1] - num_free;
     for (i = 0; i < num_free; i++) {
 	*hp++ = reg[i];
diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index ee40c308ce..36cf64e08d 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -45,6 +45,7 @@
 #include "hipe_bif0.h"
 #include "hipe_mode_switch.h"
 #include "hipe_arch.h"
+#include "hipe_load.h"
 #endif
 
 ErlDrvBinary* erts_gzinflate_buffer(char*, int);
@@ -481,9 +482,12 @@ static void free_loader_state(Binary* magic);
 static ErlHeapFragment* new_literal_fragment(Uint size);
 static void free_literal_fragment(ErlHeapFragment*);
 static void loader_state_dtor(Binary* magic);
+#ifdef HIPE
 static Eterm stub_insert_new_code(Process *c_p, ErtsProcLocks c_p_locks,
 				  Eterm group_leader, Eterm module,
-				  BeamCodeHeader* code, Uint size);
+				  BeamCodeHeader* code_hdr, Uint size,
+				  HipeModule *hipe_code);
+#endif
 static int init_iff_file(LoaderState* stp, byte* code, Uint size);
 static int scan_iff_file(LoaderState* stp, Uint* chunk_types,
 			 Uint num_types, Uint num_mandatory);
@@ -538,8 +542,6 @@ static Eterm compilation_info_for_module(Process* p, BeamCodeHeader*);
 static Eterm md5_of_module(Process* p, BeamCodeHeader*);
 static Eterm has_native(BeamCodeHeader*);
 static Eterm native_addresses(Process* p, BeamCodeHeader*);
-int patch_funentries(Eterm Patchlist);
-int patch(Eterm Addresses, Uint fe);
 static int safe_mul(UWord a, UWord b, UWord* resp);
 
 static int must_swap_floats;
@@ -843,9 +845,7 @@ erts_finish_loading(Binary* magic, Process* c_p,
 	    erts_alloc(ERTS_ALC_T_PREPARED_CODE,
 		       sizeof(struct erl_module_instance));
 	inst_p = mod_tab_p->on_load;
-	inst_p->nif = 0;
-        inst_p->num_breakpoints = 0;
-        inst_p->num_traced_exports = 0;
+        erts_module_instance_init(inst_p);
     }
 
     inst_p->code_hdr = stp->hdr;
@@ -944,6 +944,13 @@ erts_module_for_prepared_code(Binary* magic)
     LoaderState* stp;
 
     if (ERTS_MAGIC_BIN_DESTRUCTOR(magic) != loader_state_dtor) {
+#ifdef HIPE
+	HipeLoaderState *hipe_stp;
+	if ((hipe_stp = hipe_get_loader_state(magic))
+	    && hipe_stp->text_segment != 0) {
+	    return hipe_stp->module;
+	}
+#endif
 	return NIL;
     }
     stp = ERTS_MAGIC_BIN_DATA(magic);
@@ -1092,10 +1099,12 @@ loader_state_dtor(Binary* magic)
     ASSERT(stp->genop_blocks == 0);
 }
 
+#ifdef HIPE
 static Eterm
 stub_insert_new_code(Process *c_p, ErtsProcLocks c_p_locks,
 		     Eterm group_leader, Eterm module,
-		     BeamCodeHeader* code_hdr, Uint size)
+		     BeamCodeHeader* code_hdr, Uint size,
+		     HipeModule *hipe_code)
 {
     Module* modp;
     Eterm retval;
@@ -1118,6 +1127,9 @@ stub_insert_new_code(Process *c_p, ErtsProcLocks c_p_locks,
     modp->curr.code_hdr = code_hdr;
     modp->curr.code_length = size;
     modp->curr.catches = BEAM_CATCHES_NIL; /* Will be filled in later. */
+    DBG_TRACE_MFA(make_atom(modp->module), 0, 0, "insert_new_code "
+                  "first_hipe_ref = %p", hipe_code->first_hipe_ref);
+    modp->curr.hipe_code = hipe_code;
 
     /*
      * Update ranges (used for finding a function from a PC value).
@@ -1126,6 +1138,7 @@ stub_insert_new_code(Process *c_p, ErtsProcLocks c_p_locks,
     erts_update_ranges((BeamInstr*)modp->curr.code_hdr, size);
     return NIL;
 }
+#endif
 
 static int
 init_iff_file(LoaderState* stp, byte* code, Uint size)
@@ -4780,9 +4793,7 @@ final_touch(LoaderState* stp, struct erl_module_instance* inst_p)
 	}
 	ep = erts_export_put(stp->module, stp->export[i].function,
 			     stp->export[i].arity);
-	if (!on_load) {
-	    ep->addressv[erts_staging_code_ix()] = address;
-	} else {
+	if (on_load) {
 	    /*
 	     * on_load: Don't make any of the exported functions
 	     * callable yet. Keep any function in the current
@@ -4790,6 +4801,8 @@ final_touch(LoaderState* stp, struct erl_module_instance* inst_p)
 	     */
 	    ep->beam[1] = (BeamInstr) address;
 	}
+        else
+            ep->addressv[erts_staging_code_ix()] = address;
     }
 
     /*
@@ -5991,23 +6004,19 @@ code_module_md5_1(BIF_ALIST_1)
     return res;
 }
 
+#ifdef HIPE
 #define WORDS_PER_FUNCTION (sizeof(ErtsCodeInfo) / sizeof(UWord) + 1)
 
 static BeamInstr*
 make_stub(ErtsCodeInfo* info, Eterm mod, Eterm func, Uint arity, Uint native, BeamInstr OpCode)
 {
+    DBG_TRACE_MFA(mod,func,arity,"make beam stub at %p", erts_codeinfo_to_code(info));
     ASSERT(WORDS_PER_FUNCTION == 6);
     info->op = (BeamInstr) BeamOp(op_i_func_info_IaaI);
     info->native = native;
     info->mfa.module = mod;
     info->mfa.function = func;
     info->mfa.arity = arity;
-#ifdef HIPE
-    if (native) {
-        erts_codeinfo_to_code(info)[0] = BeamOpCode(op_move_return_n);
-	hipe_mfa_save_orig_beam_op(mod, func, arity, erts_codeinfo_to_code(info));
-    }
-#endif
     erts_codeinfo_to_code(info)[0] = OpCode;
     return erts_codeinfo_to_code(info)+1;
 }
@@ -6072,9 +6081,7 @@ stub_final_touch(LoaderState* stp, ErtsCodeInfo* ci)
 {
     unsigned int i;
     unsigned int n = stp->num_exps;
-#ifdef HIPE
     Lambda* lp;
-#endif
 
     if (is_bif(ci->mfa.module, ci->mfa.function, ci->mfa.arity)) {
 	ci->native = 0;
@@ -6095,6 +6102,8 @@ stub_final_touch(LoaderState* stp, ErtsCodeInfo* ci)
                                          ci->mfa.function,
                                          ci->mfa.arity);
 	    ep->addressv[erts_staging_code_ix()] = erts_codeinfo_to_code(ci);
+	    DBG_TRACE_MFA_P(&ci->mfa,"set beam stub at %p in export at %p (code_ix=%d)",
+			    erts_codeinfo_to_code(ci), ep, erts_staging_code_ix());
 	    return;
 	}
     }
@@ -6104,7 +6113,6 @@ stub_final_touch(LoaderState* stp, ErtsCodeInfo* ci)
      * Search the lambda table to find out which.
      */
     
-#ifdef HIPE
     n = stp->num_lambdas;
     for (i = 0, lp = stp->lambdas; i < n; i++, lp++) {
         ErlFunEntry* fe = stp->lambdas[i].fe;
@@ -6113,7 +6121,6 @@ stub_final_touch(LoaderState* stp, ErtsCodeInfo* ci)
             fe->address = erts_codeinfo_to_code(ci);
 	}
     }
-#endif
     return;
 }
 
@@ -6122,10 +6129,9 @@ stub_final_touch(LoaderState* stp, ErtsCodeInfo* ci)
    [{Adr, Patchtyppe} | Addresses]
    and the address of a fun_entry.
 */
-int 
+static int
 patch(Eterm Addresses, Uint fe) 
  {
-#ifdef HIPE
   Eterm* listp;
   Eterm tuple;
   Eterm* tp;
@@ -6161,15 +6167,13 @@ patch(Eterm Addresses, Uint fe)
 
   }
 
-#endif
   return 1;
 }
 
 
-int
+static int
 patch_funentries(Eterm Patchlist) 
  {
-#ifdef HIPE   
   while (!is_nil(Patchlist)) {
     Eterm Info;
     Eterm MFA;
@@ -6258,23 +6262,23 @@ patch_funentries(Eterm Patchlist)
       return 0;
 
   }
-#endif
   return 1; /* Signal that all went well */
 }
 
-
 /*
  * Do a dummy load of a module. No threaded code will be loaded.
  * Used for loading native code.
  * Will also patch all references to fun_entries to point to 
  * the new fun_entries created.
  */
-
 Eterm
-erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
+erts_make_stub_module(Process* p, Eterm hipe_magic_bin, Eterm Beam, Eterm Info)
 {
     Binary* magic;
+    Binary* hipe_magic;
     LoaderState* stp;
+    HipeLoaderState* hipe_stp;
+    HipeModule *hipe_code;
     BeamInstr Funcs;
     BeamInstr Patchlist;
     Eterm MD5Bin;
@@ -6297,8 +6301,12 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
      */
     magic = erts_alloc_loader_state();
     stp = ERTS_MAGIC_BIN_DATA(magic);
+    hipe_code = erts_alloc(ERTS_ALC_T_HIPE, sizeof(*hipe_code));
 
-    if (is_not_atom(Mod)) {
+    if (!ERTS_TERM_IS_MAGIC_BINARY(hipe_magic_bin) ||
+	!(hipe_magic = ((ProcBin*)binary_val(hipe_magic_bin))->val,
+	  hipe_stp = hipe_get_loader_state(hipe_magic)) ||
+	hipe_stp->module == NIL || hipe_stp->text_segment == 0) {
 	goto error;
     }
     if (is_not_tuple(Info)) {
@@ -6326,7 +6334,7 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
      * Scan the Beam binary and read the interesting sections.
      */
 
-    stp->module = Mod;
+    stp->module = hipe_stp->module;
     stp->group_leader = p->group_leader;
     stp->num_functions = n;
     if (!init_iff_file(stp, bytes, size)) {
@@ -6435,12 +6443,8 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
 	 * as the body until we know what kind of trap we should put there.
 	 */
 	code_hdr->functions[i] = (ErtsCodeInfo*)fp;
-#ifdef HIPE
 	op = (Eterm) BeamOpCode(op_hipe_trap_call); /* Might be changed later. */
-#else
-	op = (Eterm) BeamOpCode(op_move_return_n);
-#endif
-	fp = make_stub((ErtsCodeInfo*)fp, Mod, func, arity,
+	fp = make_stub((ErtsCodeInfo*)fp, hipe_stp->module, func, arity,
                        (Uint)native_address, op);
     }
 
@@ -6481,11 +6485,20 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
     }
 
     /*
+     * Initialise HiPE module
+     */
+    hipe_code->text_segment = hipe_stp->text_segment;
+    hipe_code->text_segment_size = hipe_stp->text_segment_size;
+    hipe_code->data_segment = hipe_stp->data_segment;
+    hipe_code->first_hipe_ref = hipe_stp->new_hipe_refs;
+    hipe_code->first_hipe_sdesc = hipe_stp->new_hipe_sdesc;
+
+    /*
      * Insert the module in the module table.
      */
 
-    rval = stub_insert_new_code(p, 0, p->group_leader, Mod,
-				code_hdr, code_size);
+    rval = stub_insert_new_code(p, 0, p->group_leader, hipe_stp->module,
+				code_hdr, code_size, hipe_code);
     if (rval != NIL) {
 	goto error;
     }
@@ -6501,18 +6514,69 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
     }
 
     if (patch_funentries(Patchlist)) {
+	Eterm mod = hipe_stp->module;
+	/* Prevent code from being freed */
+	hipe_stp->text_segment = 0;
+	hipe_stp->data_segment = 0;
+        hipe_stp->new_hipe_refs = NULL;
+        hipe_stp->new_hipe_sdesc = NULL;
+
 	erts_free_aligned_binary_bytes(temp_alloc);
 	free_loader_state(magic);
-	return Mod;
+	hipe_free_loader_state(hipe_stp);
+
+	return mod;
     }
 
  error:
+    erts_free(ERTS_ALC_T_HIPE, hipe_code);
     erts_free_aligned_binary_bytes(temp_alloc);
     free_loader_state(magic);
     BIF_ERROR(p, BADARG);
 }
 
+int erts_commit_hipe_patch_load(Eterm hipe_magic_bin)
+{
+    Binary* hipe_magic;
+    HipeLoaderState* hipe_stp;
+    HipeModule *hipe_code;
+    Module* modp;
+
+    if (!ERTS_TERM_IS_MAGIC_BINARY(hipe_magic_bin) ||
+	!(hipe_magic = ((ProcBin*)binary_val(hipe_magic_bin))->val,
+	  hipe_stp = hipe_get_loader_state(hipe_magic)) ||
+	hipe_stp->module == NIL || hipe_stp->text_segment == 0) {
+	return 0;
+    }
+
+    modp = erts_get_module(hipe_stp->module, erts_active_code_ix());
+    if (!modp)
+	return 0;
+
+    /*
+     * Initialise HiPE module
+     */
+    hipe_code = erts_alloc(ERTS_ALC_T_HIPE, sizeof(*hipe_code));
+    hipe_code->text_segment = hipe_stp->text_segment;
+    hipe_code->text_segment_size = hipe_stp->text_segment_size;
+    hipe_code->data_segment = hipe_stp->data_segment;
+    hipe_code->first_hipe_ref = hipe_stp->new_hipe_refs;
+    hipe_code->first_hipe_sdesc = hipe_stp->new_hipe_sdesc;
+
+    modp->curr.hipe_code = hipe_code;
+
+    /* Prevent code from being freed */
+    hipe_stp->text_segment = 0;
+    hipe_stp->data_segment = 0;
+    hipe_stp->new_hipe_refs = NULL;
+    hipe_stp->new_hipe_sdesc = NULL;
+
+    return 1;
+}
+
 #undef WORDS_PER_FUNCTION
+#endif /* HIPE */
+
 
 static int safe_mul(UWord a, UWord b, UWord* resp)
 {
@@ -6526,3 +6590,46 @@ static int safe_mul(UWord a, UWord b, UWord* resp)
     }
 }
 
+#ifdef ENABLE_DBG_TRACE_MFA
+
+#define MFA_MAX 10
+Eterm dbg_trace_m[MFA_MAX];
+Eterm dbg_trace_f[MFA_MAX];
+Uint  dbg_trace_a[MFA_MAX];
+unsigned int dbg_trace_ix = 0;
+
+void dbg_set_traced_mfa(const char* m, const char* f, Uint a)
+{
+    unsigned i = dbg_trace_ix++;
+    ASSERT(i < MFA_MAX);
+    dbg_trace_m[i] = am_atom_put(m, strlen(m));
+    dbg_trace_f[i] = am_atom_put(f, strlen(f));
+    dbg_trace_a[i] = a;
+}
+
+int dbg_is_traced_mfa(Eterm m, Eterm f, Uint a)
+{
+    unsigned int i;
+    for (i = 0; i < dbg_trace_ix; ++i) {
+        if (m == dbg_trace_m[i] &&
+            (!f || (f == dbg_trace_f[i] && a == dbg_trace_a[i]))) {
+
+            return i+1;
+        }
+    }
+    return 0;
+}
+
+void dbg_vtrace_mfa(unsigned ix, const char* format, ...)
+{
+    va_list arglist;
+    va_start(arglist, format);
+    ASSERT(--ix < MFA_MAX);
+    erts_fprintf(stderr, "MFA TRACE %T:%T/%u: ",
+                 dbg_trace_m[ix], dbg_trace_f[ix], (int)dbg_trace_a[ix]);
+
+    erts_vfprintf(stderr, format, arglist);
+    va_end(arglist);
+}
+
+#endif /* ENABLE_DBG_TRACE_MFA */
diff --git a/erts/emulator/beam/beam_load.h b/erts/emulator/beam/beam_load.h
index f67dfd7a81..6a3110d0f5 100644
--- a/erts/emulator/beam/beam_load.h
+++ b/erts/emulator/beam/beam_load.h
@@ -143,4 +143,34 @@ struct BeamCodeLineTab_ {
 #define LOC_FILE(Loc) ((Loc) >> 24)
 #define LOC_LINE(Loc) ((Loc) & ((1 << 24)-1))
 
+
+/*
+ * MFA event debug "tracing" usage:
+ *
+ * #define ENABLE_DBG_TRACE_MFA
+ * call dbg_set_traced_mfa("mymod","myfunc",arity)
+ * for the function(s) to trace, in some init function.
+ *
+ * Run and get stderr printouts when interesting things happen to your MFA.
+ */
+#ifdef  ENABLE_DBG_TRACE_MFA
+
+void dbg_set_traced_mfa(const char* m, const char* f, Uint a);
+int dbg_is_traced_mfa(Eterm m, Eterm f, Uint a);
+void dbg_vtrace_mfa(unsigned ix, const char* format, ...);
+#define DBG_TRACE_MFA(M,F,A,FMT, ...) do {\
+    unsigned ix;\
+    if ((ix=dbg_is_traced_mfa(M,F,A))) \
+        dbg_vtrace_mfa(ix, FMT"\n", ##__VA_ARGS__);\
+  }while(0)
+
+#define DBG_TRACE_MFA_P(MFA, FMT, ...) \
+        DBG_TRACE_MFA((MFA)->module, (MFA)->function, (MFA)->arity, FMT, ##__VA_ARGS__)
+
+#else
+#  define dbg_set_traced_mfa(M,F,A)
+#  define DBG_TRACE_MFA(M,F,A,FMT, ...)
+#  define DBG_TRACE_MFA_P(MFA,FMT, ...)
+#endif /* ENABLE_DBG_TRACE_MFA */
+
 #endif /* _BEAM_LOAD_H */
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index 3c2c9def3b..56aa78a08a 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -373,10 +373,16 @@ set_default_exec_alloc_opts(struct au_init *ip)
     ip->init.util.rmbcmt	= 0;
     ip->init.util.acul		= 0;
 
+# ifdef ERTS_HAVE_EXEC_MMAPPER
     ip->init.util.mseg_alloc    = &erts_alcu_mmapper_mseg_alloc;
     ip->init.util.mseg_realloc  = &erts_alcu_mmapper_mseg_realloc;
     ip->init.util.mseg_dealloc  = &erts_alcu_mmapper_mseg_dealloc;
     ip->init.util.mseg_mmapper  = &erts_exec_mmapper;
+# else
+    ip->init.util.mseg_alloc    = &erts_alcu_exec_mseg_alloc;
+    ip->init.util.mseg_realloc  = &erts_alcu_exec_mseg_realloc;
+    ip->init.util.mseg_dealloc  = &erts_alcu_exec_mseg_dealloc;
+# endif
 }
 #endif /* ERTS_ALC_A_EXEC */
 
@@ -1571,7 +1577,7 @@ handle_args(int *argc, char **argv, erts_alc_hndl_args_init_t *init)
 		    break;
                 case 'X':
                     if (has_prefix("scs", argv[i]+3)) {
-#ifdef ERTS_ALC_A_EXEC
+#ifdef ERTS_HAVE_EXEC_MMAPPER
                         init->mseg.exec_mmap.scs =
 #endif
                             get_mb_value(argv[i]+6, argv, &i);
@@ -2852,7 +2858,7 @@ erts_allocator_info(int to, void *arg)
         erts_print(to, arg, "=allocator:erts_mmap.literal_mmap\n");
         erts_mmap_info(&erts_literal_mmapper, &to, arg, NULL, NULL, &emis);
 #endif
-#ifdef ERTS_ALC_A_EXEC
+#ifdef ERTS_HAVE_EXEC_MMAPPER
         erts_print(to, arg, "=allocator:erts_mmap.exec_mmap\n");
         erts_mmap_info(&erts_exec_mmapper, &to, arg, NULL, NULL, &emis);
 #endif
@@ -3010,7 +3016,7 @@ erts_allocator_options(void *proc)
 #if defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
     terms[length++] = ERTS_MAKE_AM("literal_mmap");
 #endif
-#ifdef ERTS_ALC_A_EXEC
+#ifdef ERTS_HAVE_EXEC_MMAPPER
     terms[length++] = ERTS_MAKE_AM("exec_mmap");
 #endif
     features = length ? erts_bld_list(hpp, szp, length, terms) : NIL;
@@ -3102,7 +3108,7 @@ reply_alloc_info(void *vair)
 # if defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
     struct erts_mmap_info_struct mmap_info_literal;
 # endif
-# ifdef ERTS_ALC_A_EXEC
+# ifdef ERTS_HAVE_EXEC_MMAPPER
     struct erts_mmap_info_struct mmap_info_exec;
 # endif
 #endif
@@ -3232,7 +3238,7 @@ reply_alloc_info(void *vair)
                                             erts_bld_atom(hpp,szp,"literal_mmap"),
                                             ainfo);
 #  endif
-#  ifdef ERTS_ALC_A_EXEC
+#  ifdef ERTS_HAVE_EXEC_MMAPPER
                     ai_list = erts_bld_cons(hpp, szp,
                                             ainfo, ai_list);
                     ainfo = (air->only_sz ? NIL :
@@ -4129,12 +4135,20 @@ debug_free(ErtsAlcType_t n, void *extra, void *ptr)
     ErtsAllocatorFunctions_t *real_af = (ErtsAllocatorFunctions_t *) extra;
     void *dptr;
     Uint size;
+    int free_pattern = n;
 
     ASSERT(ERTS_ALC_N_MIN <= n && n <= ERTS_ALC_N_MAX);
 
     dptr = check_memory_fence(ptr, &size, n, ERTS_ALC_O_FREE);
 
-    sys_memset((void *) dptr, n, size + FENCE_SZ);
+#ifdef ERTS_ALC_A_EXEC
+# if defined(__i386__) || defined(__x86_64__)
+    if (ERTS_ALC_T2A(ERTS_ALC_N2T(n)) == ERTS_ALC_A_EXEC) {
+        free_pattern = 0x0f; /* Illegal instruction */
+    }
+# endif
+#endif
+    sys_memset((void *) dptr, free_pattern, size + FENCE_SZ);
 
     (*real_af->free)(n, real_af->extra, dptr);
 
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index dd8ca5fac1..163244237d 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -907,7 +907,9 @@ erts_alcu_literal_32_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size,
 
 #elif defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
 
-/* Used by literal allocator that has its own mmapper (super carrier) */
+/* For allocators that have their own mmapper (super carrier),
+ * like literal_alloc and exec_alloc on amd64
+ */
 void*
 erts_alcu_mmapper_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
 {
@@ -948,6 +950,50 @@ erts_alcu_mmapper_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size,
 }
 #endif /* ARCH_64 && ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION */
 
+#if defined(ERTS_ALC_A_EXEC) && !defined(ERTS_HAVE_EXEC_MMAPPER)
+
+/*
+ * For exec_alloc on non-amd64 that just need memory with PROT_EXEC
+ */
+void*
+erts_alcu_exec_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
+{
+    void* res = erts_alcu_mseg_alloc(allctr, size_p, flags);
+
+    if (res) {
+        int r = mprotect(res, *size_p, PROT_EXEC | PROT_READ | PROT_WRITE);
+        ASSERT(r == 0); (void)r;
+    }
+    return res;
+}
+
+void*
+erts_alcu_exec_mseg_realloc(Allctr_t *allctr, void *seg,
+                            Uint old_size, Uint *new_size_p)
+{
+    void *res;
+
+    if (seg && old_size) {
+        int r = mprotect(seg, old_size, PROT_READ | PROT_WRITE);
+        ASSERT(r == 0); (void)r;
+    }
+    res = erts_alcu_mseg_realloc(allctr, seg, old_size, new_size_p);
+    if (res) {
+        int r = mprotect(res, *new_size_p, PROT_EXEC | PROT_READ | PROT_WRITE);
+        ASSERT(r == 0); (void)r;
+    }
+    return res;
+}
+
+void
+erts_alcu_exec_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size, Uint flags)
+{
+    int r = mprotect(seg, size, PROT_READ | PROT_WRITE);
+    ASSERT(r == 0); (void)r;
+    erts_alcu_mseg_dealloc(allctr, seg, size, flags);
+}
+#endif /* ERTS_ALC_A_EXEC && !ERTS_HAVE_EXEC_MMAPPER */
+
 #endif /* HAVE_ERTS_MSEG */
 
 void*
diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h
index f50f09907a..2958bdf8d1 100644
--- a/erts/emulator/beam/erl_alloc_util.h
+++ b/erts/emulator/beam/erl_alloc_util.h
@@ -210,6 +210,12 @@ void* erts_alcu_mmapper_mseg_alloc(Allctr_t*, Uint *size_p, Uint flags);
 void* erts_alcu_mmapper_mseg_realloc(Allctr_t*, void *seg, Uint old_size, Uint *new_size_p);
 void  erts_alcu_mmapper_mseg_dealloc(Allctr_t*, void *seg, Uint size, Uint flags);
 # endif
+
+# if defined(ERTS_ALC_A_EXEC) && !defined(ERTS_HAVE_EXEC_MMAPPER)
+void* erts_alcu_exec_mseg_alloc(Allctr_t*, Uint *size_p, Uint flags);
+void* erts_alcu_exec_mseg_realloc(Allctr_t*, void *seg, Uint old_size, Uint *new_size_p);
+void  erts_alcu_exec_mseg_dealloc(Allctr_t*, void *seg, Uint size, Uint flags);
+# endif
 #endif /* HAVE_ERTS_MSEG */
 
 void* erts_alcu_sys_alloc(Allctr_t*, Uint *size_p, int superalign);
diff --git a/erts/emulator/beam/erl_fun.h b/erts/emulator/beam/erl_fun.h
index 73c3e19c1c..6ba055d92c 100644
--- a/erts/emulator/beam/erl_fun.h
+++ b/erts/emulator/beam/erl_fun.h
@@ -57,9 +57,6 @@ typedef struct erl_fun_thing {
     Eterm thing_word;		/* Subtag FUN_SUBTAG. */
     ErlFunEntry* fe;		/* Pointer to fun entry. */
     struct erl_off_heap_header* next;
-#ifdef HIPE
-    UWord* native_address;	/* Native code for the fun. */
-#endif
     Uint arity;			/* The arity of the fun. */
     Uint num_free;		/* Number of free variables (in env). */
   /* -- The following may be compound Erlang terms ---------------------- */
diff --git a/erts/emulator/beam/export.c b/erts/emulator/beam/export.c
index 060fde074c..94568e79a9 100644
--- a/erts/emulator/beam/export.c
+++ b/erts/emulator/beam/export.c
@@ -146,6 +146,8 @@ export_alloc(struct export_entry* tmpl_e)
 	    blob->entryv[ix].ep = &blob->exp;
 	}
 	ix = 0;
+
+	DBG_TRACE_MFA_P(&obj->info.mfa, "export allocation at %p", obj);
     }
     else { /* Existing entry in another table, use free entry in blob */
 	blob = entry_to_blob(tmpl_e);
@@ -164,9 +166,12 @@ export_free(struct export_entry* obj)
     obj->slot.index = -1;
     for (i=0; i < ERTS_NUM_CODE_IX; i++) {
 	if (blob->entryv[i].slot.index >= 0) {
+	    DBG_TRACE_MFA_P(&blob->exp.info.mfa, "export entry slot %u freed for %p",
+			  (obj - blob->entryv), &blob->exp);
 	    return;
 	}
     }
+    DBG_TRACE_MFA_P(&blob->exp.info.mfa, "export blob deallocation at %p", &blob->exp);
     erts_free(ERTS_ALC_T_EXPORT, blob);
     erts_smp_atomic_add_nob(&total_entries_bytes, -sizeof(*blob));
 }
diff --git a/erts/emulator/beam/external.c b/erts/emulator/beam/external.c
index c297bb26b3..98df8a0726 100644
--- a/erts/emulator/beam/external.c
+++ b/erts/emulator/beam/external.c
@@ -3750,7 +3750,6 @@ dec_term_atom_common:
 		if (funp->fe->native_address == NULL) {
 		    hipe_set_closure_stub(funp->fe, num_free);
 		}
-		funp->native_address = funp->fe->native_address;
 #endif
 		hp = factory->hp;
 
@@ -3822,9 +3821,6 @@ dec_term_atom_common:
 
 		funp->fe = erts_put_fun_entry(module, old_uniq, old_index);
 		funp->arity = funp->fe->address[-1] - num_free;
-#ifdef HIPE
-		funp->native_address = funp->fe->native_address;
-#endif
 		hp = factory->hp;
 
 		/* Environment */
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 7ee046ac12..537aaf3177 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -1048,6 +1048,7 @@ void erts_set_current_function(FunctionInfo* fi, ErtsCodeMFA* mfa);
 Eterm erts_module_info_0(Process* p, Eterm module);
 Eterm erts_module_info_1(Process* p, Eterm module, Eterm what);
 Eterm erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info);
+int erts_commit_hipe_patch_load(Eterm hipe_magic_bin);
 
 /* beam_ranges.c */
 void erts_init_ranges(void);
diff --git a/erts/emulator/beam/module.c b/erts/emulator/beam/module.c
index 4f36377450..705dd1d43f 100644
--- a/erts/emulator/beam/module.c
+++ b/erts/emulator/beam/module.c
@@ -26,6 +26,7 @@
 #include "erl_vm.h"
 #include "global.h"
 #include "module.h"
+#include "beam_catches.h"
 
 #ifdef DEBUG
 #  define IF_DEBUG(x) x
@@ -67,6 +68,18 @@ static int module_cmp(Module* tmpl, Module* obj)
     return tmpl->module != obj->module;
 }
 
+void erts_module_instance_init(struct erl_module_instance* modi)
+{
+    modi->code_hdr = 0;
+    modi->code_length = 0;
+    modi->catches = BEAM_CATCHES_NIL;
+    modi->nif = NULL;
+    modi->num_breakpoints = 0;
+    modi->num_traced_exports = 0;
+#ifdef HIPE
+    modi->hipe_code = NULL;
+#endif
+}
 
 static Module* module_alloc(Module* tmpl)
 {
@@ -74,18 +87,11 @@ static Module* module_alloc(Module* tmpl)
     erts_smp_atomic_add_nob(&tot_module_bytes, sizeof(Module));
 
     obj->module = tmpl->module;
-    obj->curr.code_hdr = 0;
-    obj->old.code_hdr = 0;
-    obj->curr.code_length = 0;
-    obj->old.code_length = 0;
     obj->slot.index = -1;
-    obj->curr.nif = NULL;
-    obj->old.nif = NULL;
-    obj->curr.num_breakpoints = 0;
-    obj->old.num_breakpoints  = 0;
-    obj->curr.num_traced_exports = 0;
-    obj->old.num_traced_exports = 0;
+    erts_module_instance_init(&obj->curr);
+    erts_module_instance_init(&obj->old);
     obj->on_load = 0;
+    DBG_TRACE_MFA(make_atom(obj->module), 0, 0, "module_alloc");
     return obj;
 }
 
@@ -119,6 +125,7 @@ void init_module_table(void)
     erts_smp_atomic_init_nob(&tot_module_bytes, 0);
 }
 
+
 Module*
 erts_get_module(Eterm mod, ErtsCodeIndex code_ix)
 {
@@ -139,19 +146,14 @@ erts_get_module(Eterm mod, ErtsCodeIndex code_ix)
     }
 }
 
-Module*
-erts_put_module(Eterm mod)
+
+static Module* put_module(Eterm mod, IndexTable* mod_tab)
 {
     Module e;
-    IndexTable* mod_tab;
     int oldsz, newsz;
     Module* res;
 
     ASSERT(is_atom(mod));
-    ERTS_SMP_LC_ASSERT(erts_initialized == 0
-		       || erts_has_code_write_permission());
-
-    mod_tab = &module_tables[erts_staging_code_ix()];
     e.module = atom_val(mod);
     oldsz = index_table_sz(mod_tab);
     res = (Module*) index_put_entry(mod_tab, (void*) &e);
@@ -160,6 +162,15 @@ erts_put_module(Eterm mod)
     return res;
 }
 
+Module*
+erts_put_module(Eterm mod)
+{
+    ERTS_SMP_LC_ASSERT(erts_initialized == 0
+		       || erts_has_code_write_permission());
+
+    return put_module(mod, &module_tables[erts_staging_code_ix()]);
+}
+
 Module *module_code(int i, ErtsCodeIndex code_ix)
 {
     return (Module*) erts_index_lookup(&module_tables[code_ix], i);
@@ -181,6 +192,13 @@ static ErtsCodeIndex dbg_load_code_ix = 0;
 
 static int entries_at_start_staging = 0;
 
+static ERTS_INLINE void copy_module(Module* dst_mod, Module* src_mod)
+{
+    dst_mod->curr = src_mod->curr;
+    dst_mod->old = src_mod->old;
+    dst_mod->on_load = src_mod->on_load;
+}
+
 void module_start_staging(void)
 {
     IndexTable* src = &module_tables[erts_active_code_ix()];
@@ -199,10 +217,7 @@ void module_start_staging(void)
 	src_mod = (Module*) erts_index_lookup(src, i);
 	dst_mod = (Module*) erts_index_lookup(dst, i);
 	ASSERT(src_mod->module == dst_mod->module);
-
-	dst_mod->curr = src_mod->curr;
-	dst_mod->old = src_mod->old;
-	dst_mod->on_load = src_mod->on_load;
+        copy_module(dst_mod, src_mod);
     }
 
     /*
@@ -214,9 +229,7 @@ void module_start_staging(void)
 	dst_mod = (Module*) index_put_entry(dst, src_mod);
 	ASSERT(dst_mod != src_mod);
 
-	dst_mod->curr = src_mod->curr;
-	dst_mod->old = src_mod->old;
-	dst_mod->on_load = src_mod->on_load;
+        copy_module(dst_mod, src_mod);
     }
     newsz = index_table_sz(dst);
     erts_smp_atomic_add_nob(&tot_module_bytes, (newsz - oldsz));
diff --git a/erts/emulator/beam/module.h b/erts/emulator/beam/module.h
index 1c1afc8461..8f44e11f3d 100644
--- a/erts/emulator/beam/module.h
+++ b/erts/emulator/beam/module.h
@@ -23,6 +23,10 @@
 
 #include "index.h"
 
+#ifdef HIPE
+#include "hipe_module.h"
+#endif
+
 struct erl_module_instance {
     BeamCodeHeader* code_hdr;
     int code_length;		/* Length of loaded code in bytes. */
@@ -30,6 +34,9 @@ struct erl_module_instance {
     struct erl_module_nif* nif;
     int num_breakpoints;
     int num_traced_exports;
+#ifdef HIPE
+    HipeModule *hipe_code;
+#endif
 };
 
 typedef struct erl_module {
@@ -42,6 +49,7 @@ typedef struct erl_module {
     struct erl_module_instance* on_load;
 } Module; 
 
+void erts_module_instance_init(struct erl_module_instance* modi);
 Module* erts_get_module(Eterm mod, ErtsCodeIndex code_ix);
 Module* erts_put_module(Eterm mod);