erts: Add support for fast erts_is_literal()

10 files changed, 407 insertions, 48 deletions

diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index f0b4be4c3d..a846e96204 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -4431,6 +4431,7 @@ freeze_code(LoaderState* stp)
             if (stp->literals[i].heap_frags) {
                 move_multi_frags(&ptr, &code_off_heap, stp->literals[i].heap_frags,
                                  &stp->literals[i].term, 1);
+                ASSERT(erts_is_literal(ptr_val(stp->literals[i].term)));
             }
             else ASSERT(is_immed(stp->literals[i].term));
 	}
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index bb3998769c..3e300f88ea 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -310,6 +310,25 @@ set_default_literal_alloc_opts(struct au_init *ip)
     ip->init.util.rsbcmt	= 0;
     ip->init.util.rmbcmt	= 0;
     ip->init.util.acul		= 0;
+
+#if defined(ARCH_32)
+# if HAVE_ERTS_MSEG
+    ip->init.util.mseg_alloc   = &erts_alcu_literal_32_mseg_alloc;
+    ip->init.util.mseg_realloc = &erts_alcu_literal_32_mseg_realloc;
+    ip->init.util.mseg_dealloc = &erts_alcu_literal_32_mseg_dealloc;
+# endif
+    ip->init.util.sys_alloc    = &erts_alcu_literal_32_sys_alloc;
+    ip->init.util.sys_realloc  = &erts_alcu_literal_32_sys_realloc;
+    ip->init.util.sys_dealloc  = &erts_alcu_literal_32_sys_dealloc;
+#elif defined(ARCH_64)
+# ifdef ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION
+    ip->init.util.mseg_alloc    = &erts_alcu_literal_64_mseg_alloc;
+    ip->init.util.mseg_realloc  = &erts_alcu_literal_64_mseg_realloc;
+    ip->init.util.mseg_dealloc  = &erts_alcu_literal_64_mseg_dealloc;
+# endif
+#else
+# error Unknown architecture
+#endif
 }
 
 static void
@@ -720,6 +739,9 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
 #if HAVE_ERTS_MSEG
     init.mseg.nos = erts_no_schedulers;
     erts_mseg_init(&init.mseg);
+# if defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
+    erts_mmap_init(&erts_literal_mmapper, &init.mseg.literal_mmap);
+# endif
 #endif
 
     erts_alcu_init(&init.alloc_util);
@@ -991,7 +1013,7 @@ start_au_allocator(ErtsAlcType_t alctr_n,
     }
 
     for (i = 0; i < size; i++) {
-	void *as;
+	Allctr_t *as;
 	atype = init->atype;
 
 	if (!init->thr_spec)
@@ -1028,22 +1050,22 @@ start_au_allocator(ErtsAlcType_t alctr_n,
 
 	switch (atype) {
 	case GOODFIT:
-	    as = (void *) erts_gfalc_start((GFAllctr_t *) as0,
+	    as = erts_gfalc_start((GFAllctr_t *) as0,
 					   &init->init.gf,
 					   &init->init.util);
 	    break;
 	case BESTFIT:
-	    as = (void *) erts_bfalc_start((BFAllctr_t *) as0,
+	    as = erts_bfalc_start((BFAllctr_t *) as0,
 					   &init->init.bf,
 					   &init->init.util);
 	    break;
 	case AFIT:
-	    as = (void *) erts_afalc_start((AFAllctr_t *) as0,
+	    as = erts_afalc_start((AFAllctr_t *) as0,
 					   &init->init.af,
 					   &init->init.util);
 	    break;
     	case AOFIRSTFIT:
-	    as = (void *) erts_aoffalc_start((AOFFAllctr_t *) as0,
+	    as = erts_aoffalc_start((AOFFAllctr_t *) as0,
 					     &init->init.aoff,
 					     &init->init.util);
 	    break;
@@ -1445,25 +1467,25 @@ handle_args(int *argc, char **argv, erts_alc_hndl_args_init_t *init)
 		    }
 		    else if (has_prefix("scs", argv[i]+3)) {
 #if HAVE_ERTS_MSEG
-			init->mseg.mmap.scs =
+			init->mseg.dflt_mmap.scs =
 #endif
 			    get_mb_value(argv[i]+6, argv, &i);
 		    }
 		    else if (has_prefix("sco", argv[i]+3)) {
 #if HAVE_ERTS_MSEG
-			init->mseg.mmap.sco =
+			init->mseg.dflt_mmap.sco =
 #endif
 			    get_bool_value(argv[i]+6, argv, &i);
 		    }
 		    else if (has_prefix("scrpm", argv[i]+3)) {
 #if HAVE_ERTS_MSEG
-			init->mseg.mmap.scrpm =
+			init->mseg.dflt_mmap.scrpm =
 #endif
 			    get_bool_value(argv[i]+8, argv, &i);
 		    }
 		    else if (has_prefix("scrfsd", argv[i]+3)) {
 #if HAVE_ERTS_MSEG
-			init->mseg.mmap.scrfsd =
+			init->mseg.dflt_mmap.scrfsd =
 #endif
 			    get_amount_value(argv[i]+9, argv, &i);
 		    }
diff --git a/erts/emulator/beam/erl_alloc.h b/erts/emulator/beam/erl_alloc.h
index f540bae20d..a0d8561e84 100644
--- a/erts/emulator/beam/erl_alloc.h
+++ b/erts/emulator/beam/erl_alloc.h
@@ -30,6 +30,7 @@
 #ifdef USE_THREADS
 #include "erl_threads.h"
 #endif
+#include "erl_mmap.h"
 
 #ifdef DEBUG
 #  undef ERTS_ALC_WANT_INLINE
@@ -204,6 +205,7 @@ void erts_free(ErtsAlcType_t type, void *ptr);
 void *erts_alloc_fnf(ErtsAlcType_t type, Uint size);
 void *erts_realloc_fnf(ErtsAlcType_t type, void *ptr, Uint size);
 int erts_is_allctr_wrapper_prelocked(void);
+int erts_is_literal(void* ptr);
 
 #endif /* #if !ERTS_ALC_DO_INLINE */
 
@@ -281,6 +283,28 @@ int erts_is_allctr_wrapper_prelocked(void)
 	&& !!erts_tsd_get(erts_allctr_prelock_tsd_key);  /* by me  */
 }
 
+ERTS_ALC_INLINE
+int erts_is_literal(void* ptr)
+{
+#if defined(ARCH_32)
+    Uint ix = (UWord)ptr >> ERTS_MMAP_SUPERALIGNED_BITS;
+
+    return erts_literal_vspace_map[ix / ERTS_VSPACE_WORD_BITS]
+                  & ((UWord)1 << (ix % ERTS_VSPACE_WORD_BITS));
+
+#elif defined(ARCH_64)
+# if defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
+    extern char* erts_literals_start;
+    extern UWord erts_literals_size;
+    return ErtsInArea(ptr, erts_literals_start, erts_literals_size);
+# else
+#  error Do the tag thing
+# endif
+#else
+# error No ARCH_xx
+#endif
+}
+
 #endif /* #if ERTS_ALC_DO_INLINE || defined(ERTS_ALC_INTERNAL__) */
 
 #define ERTS_ALC_GET_THR_IX() ((int) erts_get_scheduler_id())
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 444d7055c8..f34916f1ab 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -751,12 +751,77 @@ internal_free(void *ptr)
 
 #endif
 
+#ifdef ARCH_32
+
+/*
+ * Bit vector for the entire 32-bit virtual address space
+ * with one bit for each super aligned memory segment.
+ */
+
+#define VSPACE_MAP_BITS  (1 << (32 - ERTS_MMAP_SUPERALIGNED_BITS))
+#define VSPACE_MAP_SZ    (VSPACE_MAP_BITS / ERTS_VSPACE_WORD_BITS)
+
+static ERTS_INLINE void set_bit(UWord* map, Uint ix)
+{
+    ASSERT(ix / ERTS_VSPACE_WORD_BITS < VSPACE_MAP_SZ);
+    map[ix / ERTS_VSPACE_WORD_BITS]
+        |= ((UWord)1 << (ix % ERTS_VSPACE_WORD_BITS));
+}
+
+static ERTS_INLINE void clr_bit(UWord* map, Uint ix)
+{
+    ASSERT(ix / ERTS_VSPACE_WORD_BITS < VSPACE_MAP_SZ);
+    map[ix / ERTS_VSPACE_WORD_BITS]
+        &= ~((UWord)1 << (ix % ERTS_VSPACE_WORD_BITS));
+}
+
+static ERTS_INLINE int is_bit_set(UWord* map, Uint ix)
+{
+    ASSERT(ix / ERTS_VSPACE_WORD_BITS < VSPACE_MAP_SZ);
+    return map[ix / ERTS_VSPACE_WORD_BITS]
+        & ((UWord)1 << (ix % ERTS_VSPACE_WORD_BITS));
+}
+
+UWord erts_literal_vspace_map[VSPACE_MAP_SZ];
+
+static void set_literal_range(void* start, Uint size)
+{
+    Uint ix = (UWord)start >> ERTS_MMAP_SUPERALIGNED_BITS;
+    Uint n = size >> ERTS_MMAP_SUPERALIGNED_BITS;
+
+    ASSERT(!((UWord)start & ERTS_INV_SUPERALIGNED_MASK));
+    ASSERT(!((UWord)size & ERTS_INV_SUPERALIGNED_MASK));
+    ASSERT(n);
+    while (n--) {
+        ASSERT(!is_bit_set(erts_literal_vspace_map, ix));
+        set_bit(erts_literal_vspace_map, ix);
+        ix++;
+    }
+}
+
+static void clear_literal_range(void* start, Uint size)
+{
+    Uint ix = (UWord)start >> ERTS_MMAP_SUPERALIGNED_BITS;
+    Uint n = size >> ERTS_MMAP_SUPERALIGNED_BITS;
+
+    ASSERT(!((UWord)start & ERTS_INV_SUPERALIGNED_MASK));
+    ASSERT(!((UWord)size & ERTS_INV_SUPERALIGNED_MASK));
+    ASSERT(n);
+    while (n--) {
+        ASSERT(is_bit_set(erts_literal_vspace_map, ix));
+        clr_bit(erts_literal_vspace_map, ix);
+        ix++;
+    }
+}
+
+#endif /* ARCH_32 */
+
 /* mseg ... */
 
 #if HAVE_ERTS_MSEG
 
-static ERTS_INLINE void *
-alcu_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
+void*
+erts_alcu_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
 {
     void *res;
     UWord size = (UWord) *size_p;
@@ -766,8 +831,9 @@ alcu_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
     return res;
 }
 
-static ERTS_INLINE void *
-alcu_mseg_realloc(Allctr_t *allctr, void *seg, Uint old_size, Uint *new_size_p)
+void*
+erts_alcu_mseg_realloc(Allctr_t *allctr, void *seg,
+                       Uint old_size, Uint *new_size_p)
 {
     void *res;
     UWord new_size = (UWord) *new_size_p;
@@ -778,17 +844,103 @@ alcu_mseg_realloc(Allctr_t *allctr, void *seg, Uint old_size, Uint *new_size_p)
     return res;
 }
 
-static ERTS_INLINE void
-alcu_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size, Uint flags)
+void
+erts_alcu_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size, Uint flags)
 {
     erts_mseg_dealloc_opt(allctr->alloc_no, seg, (UWord) size, flags, &allctr->mseg_opt);
     INC_CC(allctr->calls.mseg_dealloc);
 }
 
-#endif
 
-static ERTS_INLINE void *
-alcu_sys_alloc(Allctr_t *allctr, Uint size, int superalign)
+#if defined(ARCH_32)
+
+void*
+erts_alcu_literal_32_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
+{
+    void* res;
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL
+                   && !allctr->t && allctr->thread_safe);
+
+    res = erts_alcu_mseg_alloc(allctr, size_p, flags);
+    if (res)
+        set_literal_range(res, *size_p);
+    return res;
+}
+
+void*
+erts_alcu_literal_32_mseg_realloc(Allctr_t *allctr, void *seg,
+                               Uint old_size, Uint *new_size_p)
+{
+    void* res;
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL
+                   && !allctr->t && allctr->thread_safe);
+
+    if (seg && old_size)
+        clear_literal_range(seg, old_size);
+    res = erts_alcu_mseg_realloc(allctr, seg, old_size, new_size_p);
+    if (res)
+        set_literal_range(res, *new_size_p);
+    return res;
+}
+
+void
+erts_alcu_literal_32_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size,
+                               Uint flags)
+{
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL
+                   && !allctr->t && allctr->thread_safe);
+
+    erts_alcu_mseg_dealloc(allctr, seg, size, flags);
+
+    clear_literal_range(seg, size);
+}
+
+#elif defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
+
+void*
+erts_alcu_literal_64_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
+{
+    void* res;
+    UWord size = (UWord) *size_p;
+    Uint32 mmap_flags = ERTS_MMAPFLG_SUPERCARRIER_ONLY;
+    if (flags & ERTS_MSEG_FLG_2POW)
+        mmap_flags |= ERTS_MMAPFLG_SUPERALIGNED;
+
+    res = erts_mmap(&erts_literal_mmapper, mmap_flags, &size);
+    *size_p = (Uint)size;
+    INC_CC(allctr->calls.mseg_alloc);
+    return res;
+}
+
+void*
+erts_alcu_literal_64_mseg_realloc(Allctr_t *allctr, void *seg,
+                               Uint old_size, Uint *new_size_p)
+{
+    void *res;
+    UWord new_size = (UWord) *new_size_p;
+    res = erts_mremap(&erts_literal_mmapper, ERTS_MSEG_FLG_NONE, seg, old_size, &new_size);
+    *new_size_p = (Uint) new_size;
+    INC_CC(allctr->calls.mseg_realloc);
+    return res;
+}
+
+void
+erts_alcu_literal_64_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size,
+                               Uint flags)
+{
+    Uint32 mmap_flags = ERTS_MMAPFLG_SUPERCARRIER_ONLY;
+    if (flags & ERTS_MSEG_FLG_2POW)
+        mmap_flags |= ERTS_MMAPFLG_SUPERALIGNED;
+
+    erts_munmap(&erts_literal_mmapper, mmap_flags, seg, (UWord)size);
+    INC_CC(allctr->calls.mseg_dealloc);
+}
+#endif /* ARCH_64 && ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION */
+
+#endif /* HAVE_ERTS_MSEG */
+
+void*
+erts_alcu_sys_alloc(Allctr_t *allctr, Uint size, int superalign)
 {
     void *res;
 #if ERTS_SA_MB_CARRIERS && ERTS_HAVE_ERTS_SYS_ALIGNED_ALLOC
@@ -803,8 +955,8 @@ alcu_sys_alloc(Allctr_t *allctr, Uint size, int superalign)
     return res;
 }
 
-static ERTS_INLINE void *
-alcu_sys_realloc(Allctr_t *allctr, void *ptr, Uint size, Uint old_size, int superalign)
+void*
+erts_alcu_sys_realloc(Allctr_t *allctr, void *ptr, Uint size, Uint old_size, int superalign)
 {
     void *res;
 
@@ -824,8 +976,8 @@ alcu_sys_realloc(Allctr_t *allctr, void *ptr, Uint size, Uint old_size, int supe
     return res;
 }
 
-static ERTS_INLINE void
-alcu_sys_free(Allctr_t *allctr, void *ptr, int superalign)
+void
+erts_alcu_sys_dealloc(Allctr_t *allctr, void *ptr, Uint size, int superalign)
 {
 #if ERTS_SA_MB_CARRIERS && ERTS_HAVE_ERTS_SYS_ALIGNED_ALLOC
     if (superalign)
@@ -838,6 +990,49 @@ alcu_sys_free(Allctr_t *allctr, void *ptr, int superalign)
 	erts_mtrace_crr_free(allctr->alloc_no, ERTS_ALC_A_SYSTEM, ptr);
 }
 
+#ifdef ARCH_32
+
+void*
+erts_alcu_literal_32_sys_alloc(Allctr_t *allctr, Uint size, int superalign)
+{
+    void* res;
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL
+                   && !allctr->t && allctr->thread_safe);
+
+    res = erts_alcu_sys_alloc(allctr, size, 1);
+    if (res)
+        set_literal_range(res, size);
+    return res;
+}
+
+void*
+erts_alcu_literal_32_sys_realloc(Allctr_t *allctr, void *ptr, Uint size, Uint old_size, int superalign)
+{
+    void* res;
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL
+                   && !allctr->t && allctr->thread_safe);
+
+    if (ptr && old_size)
+        clear_literal_range(ptr, old_size);
+    res = erts_alcu_sys_realloc(allctr, ptr, size, old_size, 1);
+    if (res)
+        set_literal_range(res, size);
+    return res;
+}
+
+void
+erts_alcu_literal_32_sys_dealloc(Allctr_t *allctr, void *ptr, Uint size, int superalign)
+{
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL
+                   && !allctr->t && allctr->thread_safe);
+
+    erts_alcu_sys_dealloc(allctr, ptr, size, 1);
+
+    clear_literal_range(ptr, size);
+}
+
+#endif /* ARCH_32 */
+
 static Uint
 get_next_mbc_size(Allctr_t *allctr)
 {
@@ -3519,6 +3714,8 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 	return NULL;
 #endif
     }
+    flags |= allctr->crr_set_flgs;
+    flags &= ~allctr->crr_clr_flgs;
 
     ASSERT((flags & CFLG_SBC && !(flags & CFLG_MBC))
 	   || (flags & CFLG_MBC && !(flags & CFLG_SBC)));
@@ -3602,7 +3799,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
         mseg_flags = ERTS_MSEG_FLG_2POW;
     }
 
-    crr = (Carrier_t *) alcu_mseg_alloc(allctr, &crr_sz, mseg_flags);
+    crr = (Carrier_t *) allctr->mseg_alloc(allctr, &crr_sz, mseg_flags);
     if (!crr) {
 	have_tried_mseg = 1;
 	if (!(have_tried_sys_alloc || flags & CFLG_FORCE_MSEG))
@@ -3644,12 +3841,12 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 	      ? UNIT_CEILING(bcrr_sz)
 	      : SYS_ALLOC_CARRIER_CEILING(bcrr_sz));
 
-    crr = (Carrier_t *) alcu_sys_alloc(allctr, crr_sz, flags & CFLG_MBC);
+    crr = (Carrier_t *) allctr->sys_alloc(allctr, crr_sz, flags & CFLG_MBC);
 	
     if (!crr) {
 	if (crr_sz > UNIT_CEILING(bcrr_sz)) {
 	    crr_sz = UNIT_CEILING(bcrr_sz);
-	    crr = (Carrier_t *) alcu_sys_alloc(allctr, crr_sz, flags & CFLG_MBC);
+	    crr = (Carrier_t *) allctr->sys_alloc(allctr, crr_sz, flags & CFLG_MBC);
 	}
 	if (!crr) {
 #if HAVE_ERTS_MSEG
@@ -3748,7 +3945,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 
 	    new_crr_sz = new_blk_sz + SBC_HEADER_SIZE;
 	    new_crr_sz = ERTS_SACRR_UNIT_CEILING(new_crr_sz);
-	    new_crr = (Carrier_t *) alcu_mseg_realloc(allctr,
+	    new_crr = (Carrier_t *) allctr->mseg_realloc(allctr,
 						      old_crr,
 						      old_crr_sz,
 						      &new_crr_sz);
@@ -3773,7 +3970,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 		       (void *) BLK2UMEM(old_blk),
 		       MIN(new_blk_sz, old_blk_sz) - ABLK_HDR_SZ);
 	    unlink_carrier(&allctr->sbc_list, old_crr);
-	    alcu_mseg_dealloc(allctr, old_crr, old_crr_sz, ERTS_MSEG_FLG_NONE);
+            allctr->mseg_dealloc(allctr, old_crr, old_crr_sz, ERTS_MSEG_FLG_NONE);
 	}
 	else {
 	    /* Old carrier unchanged; restore stat */
@@ -3790,7 +3987,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 			  ? UNIT_CEILING(new_bcrr_sz)
 			  : SYS_ALLOC_CARRIER_CEILING(new_bcrr_sz));
 
-	    new_crr = (Carrier_t *) alcu_sys_realloc(allctr,
+	    new_crr = (Carrier_t *) allctr->sys_realloc(allctr,
 						     (void *) old_crr,
 						     new_crr_sz,
 						     old_crr_sz,
@@ -3811,7 +4008,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 	    else if (new_crr_sz > UNIT_CEILING(new_bcrr_sz)) {
 		new_crr_sz = new_blk_sz + SBC_HEADER_SIZE;
 		new_crr_sz = UNIT_CEILING(new_crr_sz);
-		new_crr = (Carrier_t *) alcu_sys_realloc(allctr,
+		new_crr = (Carrier_t *) allctr->sys_realloc(allctr,
 							 (void *) old_crr,
 							 new_crr_sz,
 							 old_crr_sz,
@@ -3834,7 +4031,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 		       (void *) BLK2UMEM(old_blk),
 		       MIN(new_blk_sz, old_blk_sz) - ABLK_HDR_SZ);
 	    unlink_carrier(&allctr->sbc_list, old_crr);
-	    alcu_sys_free(allctr, old_crr, 0);
+	    allctr->sys_dealloc(allctr, old_crr, CARRIER_SZ(old_crr), 0);
 	}
 	else {
 	    /* Old carrier unchanged; restore... */
@@ -3850,13 +4047,13 @@ dealloc_carrier(Allctr_t *allctr, Carrier_t *crr, int superaligned)
 {
 #if HAVE_ERTS_MSEG
     if (IS_MSEG_CARRIER(crr))
-	alcu_mseg_dealloc(allctr, crr, CARRIER_SZ(crr),
+	allctr->mseg_dealloc(allctr, crr, CARRIER_SZ(crr),
 			  (superaligned
 			   ? ERTS_MSEG_FLG_2POW
 			   : ERTS_MSEG_FLG_NONE));
     else
 #endif
-	alcu_sys_free(allctr, crr, superaligned);
+	allctr->sys_dealloc(allctr, crr, CARRIER_SZ(crr), superaligned);
 }
 
 static void
@@ -5854,17 +6051,52 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 					    + ABLK_HDR_SZ)
 			       - ABLK_HDR_SZ);
 
+    if (init->sys_alloc) {
+        ASSERT(init->sys_realloc && init->sys_dealloc);
+        allctr->sys_alloc   = init->sys_alloc;
+        allctr->sys_realloc = init->sys_realloc;
+        allctr->sys_dealloc = init->sys_dealloc;
+    }
+    else {
+        ASSERT(!init->sys_realloc && !init->sys_dealloc);
+        allctr->sys_alloc   = &erts_alcu_sys_alloc;
+        allctr->sys_realloc = &erts_alcu_sys_realloc;
+        allctr->sys_dealloc = &erts_alcu_sys_dealloc;
+    }
+#if HAVE_ERTS_MSEG
+    if (init->mseg_alloc) {
+        ASSERT(init->mseg_realloc && init->mseg_dealloc);
+        allctr->mseg_alloc   = init->mseg_alloc;
+        allctr->mseg_realloc = init->mseg_realloc;
+        allctr->mseg_dealloc = init->mseg_dealloc;
+    }
+    else {
+        ASSERT(!init->mseg_realloc && !init->mseg_dealloc);
+        allctr->mseg_alloc   = &erts_alcu_mseg_alloc;
+        allctr->mseg_realloc = &erts_alcu_mseg_realloc;
+        allctr->mseg_dealloc = &erts_alcu_mseg_dealloc;
+    }
+    /* If a custom carrier alloc function is specified, make sure it's used */
+    if (init->mseg_alloc && !init->sys_alloc) {
+        allctr->crr_set_flgs = CFLG_FORCE_MSEG;
+        allctr->crr_clr_flgs = CFLG_FORCE_SYS_ALLOC;
+    }
+    else if (!init->mseg_alloc && init->sys_alloc) {
+        allctr->crr_set_flgs = CFLG_FORCE_SYS_ALLOC;
+        allctr->crr_clr_flgs = CFLG_FORCE_MSEG;
+    }
+#endif
+
     if (allctr->main_carrier_size) {
 	Block_t *blk;
 
 	blk = create_carrier(allctr,
 			     allctr->main_carrier_size,
-			     CFLG_MBC
+                             (ERTS_SUPER_ALIGNED_MSEG_ONLY
+                              ? CFLG_FORCE_MSEG : CFLG_FORCE_SYS_ALLOC)
+                             | CFLG_MBC
 			     | CFLG_FORCE_SIZE
 			     | CFLG_NO_CPOOL
-#if !ERTS_SUPER_ALIGNED_MSEG_ONLY
-			     | CFLG_FORCE_SYS_ALLOC
-#endif
 			     | CFLG_MAIN_CARRIER);
 	if (!blk) {
 #ifdef USE_THREADS
diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h
index f314eab4c2..16ad673d26 100644
--- a/erts/emulator/beam/erl_alloc_util.h
+++ b/erts/emulator/beam/erl_alloc_util.h
@@ -24,6 +24,12 @@
 #define ERTS_ALCU_VSN_STR "3.0"
 
 #include "erl_alloc_types.h"
+#ifdef USE_THREADS
+#define ERL_THREADS_EMU_INTERNAL__
+#include "erl_threads.h"
+#endif
+
+#include "erl_mseg.h"
 
 #define ERTS_AU_PREF_ALLOC_BITS 11
 #define ERTS_AU_MAX_PREF_ALLOC_INSTANCES (1 << ERTS_AU_PREF_ALLOC_BITS)
@@ -60,6 +66,15 @@ typedef struct {
 
     void *fix;
     size_t *fix_type_size;
+
+#if HAVE_ERTS_MSEG
+    void* (*mseg_alloc)(Allctr_t*, Uint *size_p, Uint flags);
+    void* (*mseg_realloc)(Allctr_t*, void *seg, Uint old_size, Uint *new_size_p);
+    void  (*mseg_dealloc)(Allctr_t*, void *seg, Uint size, Uint flags);
+#endif
+    void* (*sys_alloc)(Allctr_t *allctr, Uint size, int superalign);
+    void* (*sys_realloc)(Allctr_t *allctr, void *ptr, Uint size, Uint old_size, int superalign);
+    void  (*sys_dealloc)(Allctr_t *allctr, void *ptr, Uint size, int superalign);
 } AllctrInit_t;
 
 typedef struct {
@@ -173,20 +188,44 @@ void    erts_alcu_check_delayed_dealloc(Allctr_t *, int, int *, ErtsThrPrgrVal *
 #endif
 erts_aint32_t erts_alcu_fix_alloc_shrink(Allctr_t *, erts_aint32_t);
 
+#ifdef ARCH_32
+extern UWord erts_literal_vspace_map[];
+# define ERTS_VSPACE_WORD_BITS (sizeof(UWord)*8)
+#endif
+
+void* erts_alcu_mseg_alloc(Allctr_t*, Uint *size_p, Uint flags);
+void* erts_alcu_mseg_realloc(Allctr_t*, void *seg, Uint old_size, Uint *new_size_p);
+void  erts_alcu_mseg_dealloc(Allctr_t*, void *seg, Uint size, Uint flags);
+
+#if HAVE_ERTS_MSEG
+# if defined(ARCH_32)
+void* erts_alcu_literal_32_mseg_alloc(Allctr_t*, Uint *size_p, Uint flags);
+void* erts_alcu_literal_32_mseg_realloc(Allctr_t*, void *seg, Uint old_size, Uint *new_size_p);
+void  erts_alcu_literal_32_mseg_dealloc(Allctr_t*, void *seg, Uint size, Uint flags);
+
+# elif defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
+void* erts_alcu_literal_64_mseg_alloc(Allctr_t*, Uint *size_p, Uint flags);
+void* erts_alcu_literal_64_mseg_realloc(Allctr_t*, void *seg, Uint old_size, Uint *new_size_p);
+void  erts_alcu_literal_64_mseg_dealloc(Allctr_t*, void *seg, Uint size, Uint flags);
+# endif
+#endif /* HAVE_ERTS_MSEG */
+
+void* erts_alcu_sys_alloc(Allctr_t*, Uint size, int superalign);
+void* erts_alcu_sys_realloc(Allctr_t*, void *ptr, Uint size, Uint old_size, int superalign);
+void  erts_alcu_sys_dealloc(Allctr_t*, void *ptr, Uint size, int superalign);
+#ifdef ARCH_32
+void* erts_alcu_literal_32_sys_alloc(Allctr_t*, Uint size, int superalign);
+void* erts_alcu_literal_32_sys_realloc(Allctr_t*, void *ptr, Uint size, Uint old_size, int superalign);
+void  erts_alcu_literal_32_sys_dealloc(Allctr_t*, void *ptr, Uint size, int superalign);
 #endif
 
+#endif /* !ERL_ALLOC_UTIL__ */
+
 #if defined(GET_ERL_ALLOC_UTIL_IMPL) && !defined(ERL_ALLOC_UTIL_IMPL__)
 #define ERL_ALLOC_UTIL_IMPL__
 
 #define ERTS_ALCU_FLG_FAIL_REALLOC_MOVE		(((Uint32) 1) << 0)
 
-#ifdef USE_THREADS
-#define ERL_THREADS_EMU_INTERNAL__
-#include "erl_threads.h"
-#endif
-
-#include "erl_mseg.h"
-
 #undef ERTS_ALLOC_UTIL_HARD_DEBUG
 #ifdef DEBUG
 #  if 0
@@ -498,6 +537,8 @@ struct Allctr_t_ {
     Uint		min_mbc_size;
     Uint		min_mbc_first_free_size;
     Uint		min_block_size;
+    UWord               crr_set_flgs;
+    UWord               crr_clr_flgs;
 
     /* Carriers */
     CarrierList_t	mbc_list;
@@ -543,6 +584,15 @@ struct Allctr_t_ {
     void		(*remove_mbc)	        (Allctr_t *, Carrier_t *);
     UWord		(*largest_fblk_in_mbc)  (Allctr_t *, Carrier_t *);
 
+#if HAVE_ERTS_MSEG
+    void*               (*mseg_alloc)(Allctr_t*, Uint *size_p, Uint flags);
+    void*               (*mseg_realloc)(Allctr_t*, void *seg, Uint old_size, Uint *new_size_p);
+    void                (*mseg_dealloc)(Allctr_t*, void *seg, Uint size, Uint flags);
+#endif
+    void*               (*sys_alloc)(Allctr_t *allctr, Uint size, int superalign);
+    void*               (*sys_realloc)(Allctr_t *allctr, void *ptr, Uint size, Uint old_size, int superalign);
+    void                (*sys_dealloc)(Allctr_t *allctr, void *ptr, Uint size, int superalign);
+
     void		(*init_atoms)		(void);
 
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
diff --git a/erts/emulator/beam/sys.h b/erts/emulator/beam/sys.h
index 34011147d9..46577bd07f 100644
--- a/erts/emulator/beam/sys.h
+++ b/erts/emulator/beam/sys.h
@@ -1045,6 +1045,9 @@ extern int erts_use_kernel_poll;
 #define put_int8(i, s) do {((unsigned char*)(s))[0] = (i) & 0xff;} while (0)
 
 
+#define ErtsInArea(PTR,START,NBYTES) \
+    ((UWord)((char*)(PTR) - (char*)(START)) < (NBYTES))
+
 /*
  * Use DEBUGF as you would use printf, but use double parentheses:
  *
diff --git a/erts/emulator/sys/common/erl_mmap.c b/erts/emulator/sys/common/erl_mmap.c
index e8ff2bf752..03ca080c14 100644
--- a/erts/emulator/sys/common/erl_mmap.c
+++ b/erts/emulator/sys/common/erl_mmap.c
@@ -349,6 +349,12 @@ struct ErtsMemMapper_ {
 
 ErtsMemMapper erts_dflt_mmapper;
 
+#if defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
+ErtsMemMapper erts_literal_mmapper;
+char* erts_literals_start;
+UWord erts_literals_size;
+#endif
+
 #define ERTS_MMAP_SIZE_SC_SA_INC(SZ) 						\
     do {									\
 	mm->size.supercarrier.used.total += (SZ);			\
@@ -2108,6 +2114,7 @@ static void hard_dbg_mseg_init(void);
 void
 erts_mmap_init(ErtsMemMapper* mm, ErtsMMapInit *init)
 {
+    static int is_first_call = 1;
     int virtual_map = 0;
     char *start = NULL, *end = NULL;
     UWord pagesize;
@@ -2145,7 +2152,9 @@ erts_mmap_init(ErtsMemMapper* mm, ErtsMMapInit *init)
 #endif
 
     erts_smp_mtx_init(&mm->mtx, "erts_mmap");
-    erts_mtx_init(&am.init_mutex, "mmap_init_atoms");
+    if (is_first_call) {
+        erts_mtx_init(&am.init_mutex, "mmap_init_atoms");
+    }
 
 #ifdef ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION
     if (init->virtual_range.start) {
@@ -2302,6 +2311,14 @@ erts_mmap_init(ErtsMemMapper* mm, ErtsMMapInit *init)
 #ifdef HARD_DEBUG_MSEG
     hard_dbg_mseg_init();
 #endif
+
+#if defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
+   if (mm == &erts_literal_mmapper) {
+       erts_literals_start = erts_literal_mmapper.sa.bot;
+       erts_literals_size  = erts_literal_mmapper.sua.top - erts_literals_start;
+    }
+#endif
+    is_first_call = 0;
 }
 
 
diff --git a/erts/emulator/sys/common/erl_mmap.h b/erts/emulator/sys/common/erl_mmap.h
index 8707c23527..61d912fd28 100644
--- a/erts/emulator/sys/common/erl_mmap.h
+++ b/erts/emulator/sys/common/erl_mmap.h
@@ -50,8 +50,11 @@ typedef struct {
 #define ERTS_MMAP_INIT_DEFAULT_INITER \
     {{NULL, NULL}, {NULL, NULL}, 0, 1, (1 << 16), 1}
 
+#define ERTS_MMAP_INIT_LITERAL_INITER \
+    {{NULL, NULL}, {NULL, NULL}, 1024*1024*1024, 1, (1 << 16), 0}
+
 typedef struct ErtsMemMapper_ ErtsMemMapper;
-extern ErtsMemMapper erts_dflt_mmapper;
+
 void *erts_mmap(ErtsMemMapper*, Uint32 flags, UWord *sizep);
 void erts_munmap(ErtsMemMapper*, Uint32 flags, void *ptr, UWord size);
 void *erts_mremap(ErtsMemMapper*, Uint32 flags, void *ptr, UWord old_size, UWord *sizep);
@@ -121,6 +124,11 @@ Eterm erts_mmap_debug_info(ErtsMemMapper*, struct process*);
 #  define ERTS_HAVE_OS_MMAP 1
 #endif
 
+extern ErtsMemMapper erts_dflt_mmapper;
+#if defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
+extern ErtsMemMapper erts_literal_mmapper;
+#endif
+
 /*#define HARD_DEBUG_MSEG*/
 #ifdef HARD_DEBUG_MSEG
 #  define HARD_DBG_INSERT_MSEG hard_dbg_insert_mseg
diff --git a/erts/emulator/sys/common/erl_mseg.c b/erts/emulator/sys/common/erl_mseg.c
index aadcc755c1..20695899eb 100644
--- a/erts/emulator/sys/common/erl_mseg.c
+++ b/erts/emulator/sys/common/erl_mseg.c
@@ -1402,7 +1402,7 @@ erts_mseg_init(ErtsMsegInit_t *init)
 
     erts_mtx_init(&init_atoms_mutex, "mseg_init_atoms");
 
-    erts_mmap_init(&erts_dflt_mmapper, &init->mmap);
+    erts_mmap_init(&erts_dflt_mmapper, &init->dflt_mmap);
 
     if (!IS_2POW(GET_PAGE_SIZE))
 	erl_exit(ERTS_ABORT_EXIT, "erts_mseg: Unexpected page_size %beu\n", GET_PAGE_SIZE);
diff --git a/erts/emulator/sys/common/erl_mseg.h b/erts/emulator/sys/common/erl_mseg.h
index 677f8ea4ed..2acd8f8505 100644
--- a/erts/emulator/sys/common/erl_mseg.h
+++ b/erts/emulator/sys/common/erl_mseg.h
@@ -59,7 +59,8 @@ typedef struct {
     Uint rmcbf;
     Uint mcs;
     Uint nos;
-    ErtsMMapInit mmap;
+    ErtsMMapInit dflt_mmap;
+    ErtsMMapInit literal_mmap;
 } ErtsMsegInit_t;
 
 #define ERTS_MSEG_INIT_DEFAULT_INITIALIZER				\
@@ -68,7 +69,8 @@ typedef struct {
     20,			/* rmcbf: Relative max cache bad fit	*/	\
     10,			/* mcs:   Max cache size		*/	\
     1000,		/* cci:   Cache check interval		*/	\
-    ERTS_MMAP_INIT_DEFAULT_INITER					\
+    ERTS_MMAP_INIT_DEFAULT_INITER,					\
+    ERTS_MMAP_INIT_LITERAL_INITER                                       \
 }
 
 typedef struct {