Merge branch 'sverk/egil/r16/new-alloc-header-scheme/OTP-10273'

OTP-10415

* sverk/egil/r16/new-alloc-header-scheme/OTP-10273: (42 commits)
  erts: Make ll main mbc fit into 2pow size
  erts: Clear entire mseg cache upon request
  erts: Reduce max heap sizes
  tests: Refactor away ?line macro in beam_SUITE
  tests: Fix heap_sizes check
  tests: Refactor away ?line macro in process_SUITE
  tests: Use new correct min_bin_vheap_size in test
  erts: Set super alignment (256kb) and limits for sbct (8Mb) and lmbcs (128Mb)
  erts: Do not cache segments that are misaligned
  erts: Add mseg cache for large sbc segments
  erts: Reintroduce mseg options amcbf and rmcbf
  erts: Optimize erl_alloc_util.c by substitute MBC_BLK_SZ
  erts: Fix bug when allocating size near sbc_threshold
  erts: Make gc sizes fit into MB Carrier blocks
  erts: Force sbmbc to be disabled in a crude way
  erts: Fix new header scheme for win64
  erts: Fix mseg cache. Malplaced NULL pointer
  erts: Remove unused mseg options amcbf and rmcbf
  erts: Use aligned bits as constant in mseg_alloc
  erts: Don't let zero be considered a power of two
  ...

Conflicts:
	erts/emulator/test/process_SUITE.erl

22 files changed, 2344 insertions, 2195 deletions

diff --git a/erts/doc/src/erts_alloc.xml b/erts/doc/src/erts_alloc.xml
index ec5e7d9b74..87d6682328 100644
--- a/erts/doc/src/erts_alloc.xml
+++ b/erts/doc/src/erts_alloc.xml
@@ -341,7 +341,8 @@
        Largest (<c>mseg_alloc</c>) multiblock carrier size (in
        kilobytes).  See <seealso marker="#mseg_mbc_sizes">the description
        on how sizes for mseg_alloc multiblock carriers are decided</seealso>
-       in "the <c>alloc_util</c> framework" section.</item>
+       in "the <c>alloc_util</c> framework" section. On 32-bit Unix style OS
+       this limit can not be set higher than 128 megabyte.</item>
       <tag><marker id="M_mbcgs"><c><![CDATA[+M<S>mbcgs <ratio>]]></c></marker></tag>
       <item>
        (<c>mseg_alloc</c>) multiblock carrier growth stages. See
@@ -413,7 +414,8 @@
        Singleblock carrier threshold. Blocks larger than this
        threshold will be placed in singleblock carriers. Blocks
        smaller than this threshold will be placed in multiblock
-       carriers.</item>
+       carriers. On 32-bit Unix style OS this threshold can not be set higher
+       than 8 megabytes.</item>
       <tag><marker id="M_sbmbcs"><c><![CDATA[+M<S>sbmbcs <size>]]></c></marker></tag>
       <item>
 	Small block multiblock carrier size (in bytes). Memory blocks smaller
diff --git a/erts/emulator/Makefile.in b/erts/emulator/Makefile.in
index 67320697ef..89c948cc00 100644
--- a/erts/emulator/Makefile.in
+++ b/erts/emulator/Makefile.in
@@ -61,7 +61,7 @@ else
 ifeq ($(TYPE),purify)
 PURIFY = purify $(PURIFY_BUILD_OPTIONS)
 TYPEMARKER = .purify
-TYPE_FLAGS = $(DEBUG_CFLAGS) -DPURIFY -DNO_JUMP_TABLE -DERTS_MSEG_FAKE_SEGMENTS
+TYPE_FLAGS = $(DEBUG_CFLAGS) -DPURIFY -DNO_JUMP_TABLE
 ENABLE_ALLOC_TYPE_VARS += purify
 else
 
@@ -92,7 +92,7 @@ else
 ifeq ($(TYPE),valgrind)
 PURIFY = 
 TYPEMARKER = .valgrind
-TYPE_FLAGS = $(DEBUG_CFLAGS) -DVALGRIND -DNO_JUMP_TABLE -DERTS_MSEG_FAKE_SEGMENTS
+TYPE_FLAGS = $(DEBUG_CFLAGS) -DVALGRIND -DNO_JUMP_TABLE
 ENABLE_ALLOC_TYPE_VARS += valgrind
 else
 
diff --git a/erts/emulator/beam/beam_bp.c b/erts/emulator/beam/beam_bp.c
index 2168d410b5..c1e11f6448 100644
--- a/erts/emulator/beam/beam_bp.c
+++ b/erts/emulator/beam/beam_bp.c
@@ -65,10 +65,10 @@
 
 #define ERTS_BPF_ALL               0xFF
 
-extern Eterm beam_return_to_trace[1];   /* OpCode(i_return_to_trace) */
-extern Eterm beam_return_trace[1];      /* OpCode(i_return_trace) */
-extern Eterm beam_exception_trace[1];   /* OpCode(i_exception_trace) */
-extern Eterm beam_return_time_trace[1]; /* OpCode(i_return_time_trace) */
+extern BeamInstr beam_return_to_trace[1];   /* OpCode(i_return_to_trace) */
+extern BeamInstr beam_return_trace[1];      /* OpCode(i_return_trace) */
+extern BeamInstr beam_exception_trace[1];   /* OpCode(i_exception_trace) */
+extern BeamInstr beam_return_time_trace[1]; /* OpCode(i_return_time_trace) */
 
 erts_smp_atomic32_t erts_active_bp_index;
 erts_smp_atomic32_t erts_staging_bp_index;
@@ -161,7 +161,7 @@ erts_bp_match_functions(BpFunctions* f, Eterm mfa[3], int specified)
     for (current = 0; current < num_modules; current++) {
 	BeamInstr** code_base = (BeamInstr **) module[current]->curr.code;
 	BeamInstr* code;
-	Uint num_functions = (Uint) code_base[MI_NUM_FUNCTIONS];
+	Uint num_functions = (Uint)(UWord) code_base[MI_NUM_FUNCTIONS];
 	Uint fi;
 
 	if (specified > 0) {
@@ -172,7 +172,7 @@ erts_bp_match_functions(BpFunctions* f, Eterm mfa[3], int specified)
 	}
 
 	for (fi = 0; fi < num_functions; fi++) {
-	    Eterm* pc;
+	    BeamInstr* pc;
 	    int wi;
 
 	    code = code_base[MI_FUNCTIONS+fi];
@@ -555,7 +555,7 @@ erts_clear_module_break(Module *modp) {
     if (code_base == NULL) {
 	return 0;
     }
-    n = (Uint) code_base[MI_NUM_FUNCTIONS];
+    n = (Uint)(UWord) code_base[MI_NUM_FUNCTIONS];
     for (i = 0; i < n; ++i) {
 	BeamInstr* pc;
 
@@ -919,7 +919,7 @@ do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
 	E -= 1;
 	ASSERT(c_p->htop <= E && E <= c_p->hend);
 	E[0] = make_cp(c_p->cp);
-	c_p->cp = (BeamInstr *) beam_return_to_trace;
+	c_p->cp = beam_return_to_trace;
     }
     if (flags & MATCH_SET_RX_TRACE) {
 	E -= 3;
diff --git a/erts/emulator/beam/erl_afit_alloc.c b/erts/emulator/beam/erl_afit_alloc.c
index 570cc59be2..306b32691c 100644
--- a/erts/emulator/beam/erl_afit_alloc.c
+++ b/erts/emulator/beam/erl_afit_alloc.c
@@ -37,6 +37,12 @@
 #define GET_ERL_AF_ALLOC_IMPL
 #include "erl_afit_alloc.h"
 
+struct AFFreeBlock_t_ {
+    Block_t block_head;
+    AFFreeBlock_t *prev;
+    AFFreeBlock_t *next;
+};
+#define AF_BLK_SZ(B) MBC_FBLK_SZ(&(B)->block_head)
 
 #define MIN_MBC_SZ		(16*1024)
 #define MIN_MBC_FIRST_FREE_SZ	(4*1024)
@@ -80,7 +86,6 @@ erts_afalc_start(AFAllctr_t *afallctr,
 
     init->sbmbct = 0; /* Small mbc not supported by afit */
 
-    allctr->mbc_header_size		= sizeof(Carrier_t);
     allctr->min_mbc_size		= MIN_MBC_SZ;
     allctr->min_mbc_first_free_size	= MIN_MBC_FIRST_FREE_SZ;
     allctr->min_block_size		= sizeof(AFFreeBlock_t);
@@ -118,7 +123,7 @@ get_free_block(Allctr_t *allctr, Uint size, Block_t *cand_blk, Uint cand_size,
 
     ASSERT(!cand_blk || cand_size >= size);
 
-    if (afallctr->free_list && BLK_SZ(afallctr->free_list) >= size) {
+    if (afallctr->free_list && AF_BLK_SZ(afallctr->free_list) >= size) {
 	AFFreeBlock_t *res = afallctr->free_list;	
 	afallctr->free_list = res->next;
 	if (res->next)
@@ -135,7 +140,7 @@ link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
     AFFreeBlock_t *blk = (AFFreeBlock_t *) block;
     AFAllctr_t *afallctr = (AFAllctr_t *) allctr;
 
-    if (afallctr->free_list && BLK_SZ(afallctr->free_list) > BLK_SZ(blk)) {
+    if (afallctr->free_list && AF_BLK_SZ(afallctr->free_list) > AF_BLK_SZ(blk)) {
 	blk->next = afallctr->free_list->next;
 	blk->prev = afallctr->free_list;
 	afallctr->free_list->next = blk;
diff --git a/erts/emulator/beam/erl_afit_alloc.h b/erts/emulator/beam/erl_afit_alloc.h
index ea408a7194..87caccac20 100644
--- a/erts/emulator/beam/erl_afit_alloc.h
+++ b/erts/emulator/beam/erl_afit_alloc.h
@@ -49,11 +49,6 @@ Allctr_t *erts_afalc_start(AFAllctr_t *, AFAllctrInit_t *, AllctrInit_t *);
 #include "erl_alloc_util.h"
 
 typedef struct AFFreeBlock_t_ AFFreeBlock_t;
-struct AFFreeBlock_t_ {
-    Block_t block_head;
-    AFFreeBlock_t *prev;
-    AFFreeBlock_t *next;
-};
 
 struct AFAllctr_t_ {
     Allctr_t		allctr; /* Has to be first! */
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index 04d807b780..5da8c69c03 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -310,9 +310,9 @@ set_default_ll_alloc_opts(struct au_init *ip)
     ip->init.util.name_prefix	= "ll_";
     ip->init.util.alloc_no	= ERTS_ALC_A_LONG_LIVED;
 #ifndef SMALL_MEMORY
-    ip->init.util.mmbcs 	= 2*1024*1024; /* Main carrier size */
+    ip->init.util.mmbcs 	= 2*1024*1024 - 40; /* Main carrier size */
 #else
-    ip->init.util.mmbcs 	= 1*1024*1024; /* Main carrier size */
+    ip->init.util.mmbcs 	= 1*1024*1024 - 40; /* Main carrier size */
 #endif
     ip->init.util.ts 		= ERTS_ALC_MTA_LONG_LIVED;
     ip->init.util.asbcst	= 0;
@@ -1173,6 +1173,11 @@ handle_au_arg(struct au_init *auip,
 	break;
     case 'e':
 	auip->enable = get_bool_value(sub_param+1, argv, ip);
+#if !HAVE_ERTS_SBMBC
+	if (auip->init.util.alloc_no == ERTS_ALC_A_SBMBC) {
+	    auip->enable = 0;
+	}
+#endif
 	break;
     case 'l':
 	if (has_prefix("lmbcs", sub_param)) {
@@ -1233,10 +1238,16 @@ handle_au_arg(struct au_init *auip,
 	    auip->init.util.sbct = get_kb_value(sub_param + 4, argv, ip);
 	}
 	else if (has_prefix("sbmbcs", sub_param)) {
-	    auip->init.util.sbmbcs = get_byte_value(sub_param + 6, argv, ip);
+#if HAVE_ERTS_SBMBC
+	    auip->init.util.sbmbcs =
+#endif
+		get_byte_value(sub_param + 6, argv, ip);
 	}
 	else if (has_prefix("sbmbct", sub_param)) {
-	    auip->init.util.sbmbct = get_byte_value(sub_param + 6, argv, ip);
+#if HAVE_ERTS_SBMBC
+	    auip->init.util.sbmbct =
+#endif
+		get_byte_value(sub_param + 6, argv, ip);
 	}
 	else if (has_prefix("smbcs", sub_param)) {
 	    auip->default_.smbcs = 0;
@@ -1403,6 +1414,9 @@ handle_args(int *argc, char **argv, erts_alc_hndl_args_init_t *init)
 			else if (strcmp("max", arg) == 0) {
 			    for (a = 0; a < aui_sz; a++)
 				aui[a]->enable = 1;
+#if !HAVE_ERTS_SBMBC
+			    init->sbmbc_alloc.enable = 0;
+#endif
 			}
 			else if (strcmp("config", arg) == 0) {
 			    init->erts_alloc_config = 1;
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 97ba306a79..3d6761339b 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -78,10 +78,12 @@ int erts_have_sbmbc_alloc;
 
 #if HAVE_ERTS_MSEG
 
-#define INV_MSEG_UNIT_MASK	((UWord) (mseg_unit_size - 1))
-#define MSEG_UNIT_MASK		(~INV_MSEG_UNIT_MASK)
+#define MSEG_UNIT_SHIFT         MSEG_ALIGN_BITS
+#define MSEG_UNIT_SZ            (1 << MSEG_UNIT_SHIFT)
+#define MSEG_UNIT_MASK		((~(UWord)0) << MSEG_UNIT_SHIFT)
+
 #define MSEG_UNIT_FLOOR(X)	((X) & MSEG_UNIT_MASK)
-#define MSEG_UNIT_CEILING(X)	MSEG_UNIT_FLOOR((X) + INV_MSEG_UNIT_MASK)
+#define MSEG_UNIT_CEILING(X)	MSEG_UNIT_FLOOR((X) + ~MSEG_UNIT_MASK)
 
 #endif
 
@@ -104,7 +106,6 @@ int erts_have_sbmbc_alloc;
 static Uint sys_alloc_carrier_size;
 #if HAVE_ERTS_MSEG
 static Uint max_mseg_carriers;
-static Uint mseg_unit_size;
 #endif
 
 #define ONE_GIGA (1000000000)
@@ -117,16 +118,47 @@ static Uint mseg_unit_size;
 		    ? ((CC).giga_no--, (CC).no = ONE_GIGA - 1)		\
 		    : (CC).no--)
 
-/* ... */
+/* Multi block carrier (MBC) memory layout in R16: 
+
+Empty MBC:
+[Carrier_t|pad|Block_t L0T|fhdr| free... ]
+
+MBC after allocating first block:
+[Carrier_t|pad|Block_t 000|        udata        |pad|Block_t L0T|fhdr| free... ]
+
+MBC after allocating second block:
+[Carrier_t|pad|Block_t 000|        udata        |pad|Block_t 000|   udata   |pad|Block_t L0T|fhdr| free... ]
+
+MBC after deallocating first block:
+[Carrier_t|pad|Block_t 00T|fhdr| free  |FreeBlkFtr_t|Block_t 0P0|   udata   |pad|Block_t L0T|fhdr| free... ]
+
+
+    udata = Allocated user data
+    pad   = Padding to ensure correct alignment for user data
+    fhdr  = Allocator specific header to keep track of free block
+    free  = Unused free memory
+    T     = This block is free (THIS_FREE_BLK_HDR_FLG)
+    P     = Previous block is free (PREV_FREE_BLK_HDR_FLG)
+    L     = Last block in carrier (LAST_BLK_HDR_FLG)
+*/
+
+/* Single block carrier (SBC):
+[Carrier_t|pad|Block_t 111| udata... ]
+*/
+
 
 /* Blocks ... */
 
-#define SBC_BLK_FTR_FLG		(((UWord) 1) << 0)
+#define UNUSED0_BLK_FTR_FLG	(((UWord) 1) << 0)
 #define UNUSED1_BLK_FTR_FLG	(((UWord) 1) << 1)
 #define UNUSED2_BLK_FTR_FLG	(((UWord) 1) << 2)
 
-#define ABLK_HDR_SZ (sizeof(Block_t))
-#define FBLK_FTR_SZ (sizeof(UWord))
+#if MBC_ABLK_OFFSET_BITS
+#  define ABLK_HDR_SZ (offsetof(Block_t,u))
+#else
+#  define ABLK_HDR_SZ (sizeof(Block_t))
+#endif
+#define FBLK_FTR_SZ (sizeof(FreeBlkFtr_t))
 
 #define UMEMSZ2BLKSZ(AP, SZ)						\
   (ABLK_HDR_SZ + (SZ) <= (AP)->min_block_size				\
@@ -136,88 +168,181 @@ static Uint mseg_unit_size;
 #define UMEM2BLK(P) ((Block_t *) (((char *) (P)) - ABLK_HDR_SZ))
 #define BLK2UMEM(P) ((void *)    (((char *) (P)) + ABLK_HDR_SZ))
 
-#define PREV_BLK_SZ(B) \
-  ((UWord) (*(((UWord *) (B)) - 1) & SZ_MASK))
+#define PREV_BLK_SZ(B) 		((UWord) (((FreeBlkFtr_t *)(B))[-1]))
 
 #define SET_BLK_SZ_FTR(B, SZ) \
-  (*((UWord *) (((char *) (B)) + (SZ) - sizeof(UWord))) = (SZ))
+  (((FreeBlkFtr_t *) (((char *) (B)) + (SZ)))[-1] = (SZ))
 
 #define THIS_FREE_BLK_HDR_FLG 	(((UWord) 1) << 0)
 #define PREV_FREE_BLK_HDR_FLG 	(((UWord) 1) << 1)
 #define LAST_BLK_HDR_FLG 	(((UWord) 1) << 2)
 
-#define SET_BLK_SZ(B, SZ) \
+/* Special flag combo for (allocated) SBC blocks 
+*/
+#define SBC_BLK_HDR_FLG (THIS_FREE_BLK_HDR_FLG | PREV_FREE_BLK_HDR_FLG | LAST_BLK_HDR_FLG)
+
+#define SET_MBC_ABLK_SZ(B, SZ) \
   (ASSERT(((SZ) & FLG_MASK) == 0), \
-   (*((Block_t *) (B)) = ((*((Block_t *) (B)) & FLG_MASK) | (SZ))))
-#define SET_BLK_FREE(B) \
-  (*((Block_t *) (B)) |= THIS_FREE_BLK_HDR_FLG)
-#define SET_BLK_ALLOCED(B) \
-  (*((Block_t *) (B)) &= ~THIS_FREE_BLK_HDR_FLG)
-#define SET_PREV_BLK_FREE(B) \
-  (*((Block_t *) (B)) |= PREV_FREE_BLK_HDR_FLG)
+   (B)->bhdr = (((B)->bhdr) & ~MBC_ABLK_SZ_MASK) | (SZ))
+#define SET_MBC_FBLK_SZ(B, SZ) \
+  (ASSERT(((SZ) & FLG_MASK) == 0), \
+   (B)->bhdr = (((B)->bhdr) & ~MBC_FBLK_SZ_MASK) | (SZ))
+#define SET_SBC_BLK_SZ(B, SZ) \
+  (ASSERT(((SZ) & FLG_MASK) == 0), \
+   (B)->bhdr = (((B)->bhdr) & ~SBC_BLK_SZ_MASK) | (SZ))
+#define SET_PREV_BLK_FREE(AP,B) \
+  (ASSERT(!IS_MBC_FIRST_BLK(AP,B)), \
+   ASSERT(!IS_FREE_BLK(B)), \
+   (B)->bhdr |= PREV_FREE_BLK_HDR_FLG)
 #define SET_PREV_BLK_ALLOCED(B) \
-  (*((Block_t *) (B)) &= ~PREV_FREE_BLK_HDR_FLG)
+  ((B)->bhdr &= ~PREV_FREE_BLK_HDR_FLG)
 #define SET_LAST_BLK(B) \
-  (*((Block_t *) (B)) |= LAST_BLK_HDR_FLG)
+  ((B)->bhdr |= LAST_BLK_HDR_FLG)
 #define SET_NOT_LAST_BLK(B) \
-  (*((Block_t *) (B)) &= ~LAST_BLK_HDR_FLG)
+  ((B)->bhdr &= ~LAST_BLK_HDR_FLG)
 
 #define SBH_THIS_FREE		THIS_FREE_BLK_HDR_FLG
-#define SBH_THIS_ALLOCED	((UWord) 0)
 #define SBH_PREV_FREE		PREV_FREE_BLK_HDR_FLG
-#define SBH_PREV_ALLOCED	((UWord) 0)
 #define SBH_LAST_BLK		LAST_BLK_HDR_FLG
-#define SBH_NOT_LAST_BLK	((UWord) 0)
 
-#define SET_BLK_HDR(B, Sz, F) \
-  (ASSERT(((Sz) & FLG_MASK) == 0), *((Block_t *) (B)) = ((Sz) | (F)))
+
+#if MBC_ABLK_OFFSET_BITS
+
+#  define MBC_SZ_MAX_LIMIT ((((UWord)1 << MBC_ABLK_OFFSET_BITS) - 1) << MSEG_ALIGN_BITS)
+
+#  define BLK_CARRIER_OFFSET(B, C) (((char*)(B) - (char*)(C)) >> MSEG_UNIT_SHIFT)
+
+#  define SET_MBC_ABLK_HDR(B, Sz, F, C) \
+    (ASSERT(((Sz) & ~MBC_ABLK_SZ_MASK) == 0), \
+     ASSERT(!((UWord)(F) & (~FLG_MASK|THIS_FREE_BLK_HDR_FLG))), \
+     (B)->bhdr = ((Sz) | (F) | (BLK_CARRIER_OFFSET(B,C) << MBC_ABLK_OFFSET_SHIFT)))
+
+#  define SET_MBC_FBLK_HDR(B, Sz, F, C) \
+    (ASSERT(((Sz) & ~MBC_FBLK_SZ_MASK) == 0), \
+     ASSERT(((UWord)(F) & (~FLG_MASK|THIS_FREE_BLK_HDR_FLG|PREV_FREE_BLK_HDR_FLG)) == THIS_FREE_BLK_HDR_FLG), \
+     (B)->bhdr = ((Sz) | (F)), \
+     (B)->u.carrier = (C))
+
+#  define ABLK_TO_MBC(B) \
+    (ASSERT(IS_MBC_BLK(B) && IS_ALLOCED_BLK(B)), \
+     (Carrier_t*)((MSEG_UNIT_FLOOR((UWord)(B)) - \
+		  (((B)->bhdr >> MBC_ABLK_OFFSET_SHIFT) << MSEG_UNIT_SHIFT))))
+
+#  define FBLK_TO_MBC(B) \
+    (ASSERT(IS_MBC_BLK(B) && IS_FREE_BLK(B)), \
+     (B)->u.carrier)
+      
+#  define BLK_TO_MBC(B) (IS_FREE_BLK(B) ? FBLK_TO_MBC(B) : ABLK_TO_MBC(B))
+
+#  define IS_MBC_FIRST_ABLK(AP,B) \
+  ((((UWord)(B) & ~MSEG_UNIT_MASK) == MBC_HEADER_SIZE(AP)) \
+   && ((B)->bhdr & MBC_ABLK_OFFSET_MASK) == 0)
+
+#  define IS_MBC_FIRST_FBLK(AP,B) \
+  ((char*)(B) == (char*)((B)->u.carrier) + MBC_HEADER_SIZE(AP))
+
+#  define IS_MBC_FIRST_BLK(AP,B) \
+  (IS_FREE_BLK(B) ? IS_MBC_FIRST_FBLK(AP,B) : IS_MBC_FIRST_ABLK(AP,B))
+
+#  define SET_BLK_FREE(B) \
+  (ASSERT(!IS_PREV_BLK_FREE(B)), \
+   (B)->u.carrier = ABLK_TO_MBC(B), \
+   (B)->bhdr |= THIS_FREE_BLK_HDR_FLG, \
+   (B)->bhdr &= (MBC_ABLK_SZ_MASK|FLG_MASK))
+
+#  define SET_BLK_ALLOCED(B) \
+  (ASSERT(((B)->bhdr & (MBC_ABLK_OFFSET_MASK|THIS_FREE_BLK_HDR_FLG)) == THIS_FREE_BLK_HDR_FLG), \
+   (B)->bhdr &= ~THIS_FREE_BLK_HDR_FLG, \
+   (B)->bhdr |= (BLK_CARRIER_OFFSET(B,(B)->u.carrier) << MBC_ABLK_OFFSET_SHIFT))
+
+#else /* !MBC_ABLK_OFFSET_BITS */
+
+#  define MBC_SZ_MAX_LIMIT ((UWord)~0)
+
+#  define SET_MBC_ABLK_HDR(B, Sz, F, C) \
+    (ASSERT(((Sz) & FLG_MASK) == 0), \
+     ASSERT(!((UWord)(F) & (~FLG_MASK|THIS_FREE_BLK_HDR_FLG))), \
+     ASSERT((UWord)(F) < SBC_BLK_HDR_FLG), \
+     (B)->bhdr = ((Sz) | (F)), \
+     (B)->carrier = (C))
+
+#  define SET_MBC_FBLK_HDR(B, Sz, F, C) \
+    (ASSERT(((Sz) & FLG_MASK) == 0), \
+     ASSERT(((UWord)(F) & (~FLG_MASK|THIS_FREE_BLK_HDR_FLG|PREV_FREE_BLK_HDR_FLG)) == THIS_FREE_BLK_HDR_FLG), \
+     (B)->bhdr = ((Sz) | (F)), \
+     (B)->carrier = (C))
+
+#  define BLK_TO_MBC(B) ((B)->carrier)
+#  define ABLK_TO_MBC(B) BLK_TO_MBC(B)
+#  define FBLK_TO_MBC(B) BLK_TO_MBC(B)
+
+#  define IS_MBC_FIRST_BLK(AP,B) \
+  ((char*)(B) == (char*)((B)->carrier) + MBC_HEADER_SIZE(AP))
+#  define IS_MBC_FIRST_ABLK(AP,B) IS_MBC_FIRST_BLK(AP,B)
+#  define IS_MBC_FIRST_FBLK(AP,B) IS_MBC_FIRST_BLK(AP,B)
+
+#  define SET_BLK_FREE(B) \
+  (ASSERT(!IS_PREV_BLK_FREE(B)), \
+   (B)->bhdr |= THIS_FREE_BLK_HDR_FLG)
+
+#  define SET_BLK_ALLOCED(B) \
+  ((B)->bhdr &= ~THIS_FREE_BLK_HDR_FLG)
+
+#endif /* !MBC_ABLK_OFFSET_BITS */
+
+#define SET_SBC_BLK_HDR(B, Sz) \
+  (ASSERT(((Sz) & FLG_MASK) == 0), (B)->bhdr = ((Sz) | (SBC_BLK_HDR_FLG)))
+
 
 #define BLK_UMEM_SZ(B) \
   (BLK_SZ(B) - (ABLK_HDR_SZ))
 #define IS_PREV_BLK_FREE(B) \
-  (*((Block_t *) (B)) & PREV_FREE_BLK_HDR_FLG)
+  ((B)->bhdr & PREV_FREE_BLK_HDR_FLG)
 #define IS_PREV_BLK_ALLOCED(B) \
   (!IS_PREV_BLK_FREE((B)))
 #define IS_FREE_BLK(B) \
-  (*((Block_t *) (B)) & THIS_FREE_BLK_HDR_FLG)
+  (ASSERT(!IS_SBC_BLK(B)), (B)->bhdr & THIS_FREE_BLK_HDR_FLG)
 #define IS_ALLOCED_BLK(B) \
   (!IS_FREE_BLK((B)))  
 #define IS_LAST_BLK(B) \
-  (*((Block_t *) (B)) & LAST_BLK_HDR_FLG)
+  ((B)->bhdr & LAST_BLK_HDR_FLG)
 #define IS_NOT_LAST_BLK(B) \
   (!IS_LAST_BLK((B)))
 
 #define GET_LAST_BLK_HDR_FLG(B) \
-  (*((Block_t*) (B)) & LAST_BLK_HDR_FLG)
+  ((B)->bhdr & LAST_BLK_HDR_FLG)
 #define GET_THIS_FREE_BLK_HDR_FLG(B) \
-  (*((Block_t*) (B)) & THIS_FREE_BLK_HDR_FLG)
+  ((B)->bhdr & THIS_FREE_BLK_HDR_FLG)
 #define GET_PREV_FREE_BLK_HDR_FLG(B) \
-  (*((Block_t*) (B)) & PREV_FREE_BLK_HDR_FLG)
+  ((B)->bhdr & PREV_FREE_BLK_HDR_FLG)
 #define GET_BLK_HDR_FLGS(B) \
-  (*((Block_t*) (B)) & FLG_MASK)
-
-#define IS_FIRST_BLK(B) \
-  (IS_PREV_BLK_FREE((B)) && (PREV_BLK_SZ((B)) == 0))
-#define IS_NOT_FIRST_BLK(B) \
-  (!IS_FIRST_BLK((B)))
-
-#define SET_SBC_BLK_FTR(FTR) \
-  ((FTR) = (0 | SBC_BLK_FTR_FLG))
-#define SET_MBC_BLK_FTR(FTR) \
-  ((FTR) = 0)
+  ((B)->bhdr & FLG_MASK)
 
 #define IS_SBC_BLK(B) \
-  (IS_PREV_BLK_FREE((B)) && (((UWord *) (B))[-1] & SBC_BLK_FTR_FLG))
+    (((B)->bhdr & FLG_MASK) == SBC_BLK_HDR_FLG)
 #define IS_MBC_BLK(B) \
   (!IS_SBC_BLK((B)))
 
+#define MBC_BLK_SZ(B) (IS_FREE_BLK(B) ? MBC_FBLK_SZ(B) : MBC_ABLK_SZ(B))
+
 #define NXT_BLK(B) \
-  ((Block_t *) (((char *) (B)) + BLK_SZ((B))))
+  (ASSERT(IS_MBC_BLK(B)), \
+   (Block_t *) (((char *) (B)) + MBC_BLK_SZ((B))))
 #define PREV_BLK(B) \
   ((Block_t *) (((char *) (B)) - PREV_BLK_SZ((B))))
 
+#define BLK_AFTER(B,Sz) \
+  ((Block_t *) (((char *) (B)) + (Sz)))
+
+#define BLK_SZ(B) ((B)->bhdr & (((B)->bhdr & THIS_FREE_BLK_HDR_FLG) ? MBC_FBLK_SZ_MASK : MBC_ABLK_SZ_MASK))
+
 /* Carriers ... */
 
+#define SBC_HEADER_SIZE	   (UNIT_CEILING(sizeof(Carrier_t) + ABLK_HDR_SZ) \
+			    - ABLK_HDR_SZ)
+#define MBC_HEADER_SIZE(AP) SBC_HEADER_SIZE
+
+
 #define MSEG_CARRIER_HDR_FLAG		(((UWord) 1) << 0)
 #define SBC_CARRIER_HDR_FLAG		(((UWord) 1) << 1)
 
@@ -226,20 +351,20 @@ static Uint mseg_unit_size;
 #define SCH_MBC				0
 #define SCH_SBC				SBC_CARRIER_HDR_FLAG
 
-#define SET_CARRIER_HDR(C, Sz, F) \
-  (ASSERT(((Sz) & FLG_MASK) == 0), (C)->chdr = ((Sz) | (F)))
+#define SET_CARRIER_HDR(C, Sz, F, AP) \
+  (ASSERT(((Sz) & FLG_MASK) == 0), (C)->chdr = ((Sz) | (F)), (C)->allctr = (AP))
 
-#define BLK2SBC(AP, B) \
-  ((Carrier_t *) (((char *) (B)) - (AP)->sbc_header_size))
-#define FBLK2MBC(AP, B) \
-  ((Carrier_t *) (((char *) (B)) - (AP)->mbc_header_size))
+#define BLK_TO_SBC(B) \
+  ((Carrier_t *) (((char *) (B)) - SBC_HEADER_SIZE))
+#define FIRST_BLK_TO_MBC(AP, B) \
+  ((Carrier_t *) (((char *) (B)) - MBC_HEADER_SIZE(AP)))
 
-#define MBC2FBLK(AP, P) \
-  ((Block_t *) (((char *) (P)) + (AP)->mbc_header_size))
+#define MBC_TO_FIRST_BLK(AP, P) \
+  ((Block_t *) (((char *) (P)) + MBC_HEADER_SIZE(AP)))
 #define SBC2BLK(AP, P) \
-  ((Block_t *) (((char *) (P)) + (AP)->sbc_header_size))
+  ((Block_t *) (((char *) (P)) + SBC_HEADER_SIZE))
 #define SBC2UMEM(AP, P) \
-  ((void *) (((char *) (P)) + ((AP)->sbc_header_size + ABLK_HDR_SZ)))
+  ((void *) (((char *) (P)) + (SBC_HEADER_SIZE + ABLK_HDR_SZ)))
 
 #define IS_MSEG_CARRIER(C) \
   ((C)->chdr & MSEG_CARRIER_HDR_FLAG)
@@ -250,15 +375,6 @@ static Uint mseg_unit_size;
 #define IS_MB_CARRIER(C) \
   (!IS_SB_CARRIER((C)))
 
-#define SET_MSEG_CARRIER(C) \
-  ((C)->chdr |= MSEG_CARRIER_HDR_FLAG)
-#define SET_SYS_ALLOC_CARRIER(C) \
-  ((C)->chdr &= ~MSEG_CARRIER_HDR_FLAG)
-#define SET_SB_CARRIER(C) \
-  ((C)->chdr |= SBC_CARRIER_HDR_FLAG)
-#define SET_MB_CARRIER(C) \
-  ((C)->chdr &= ~SBC_CARRIER_HDR_FLAG)
-
 #define SET_CARRIER_SZ(C, SZ) \
   (ASSERT(((SZ) & FLG_MASK) == 0), \
    ((C)->chdr = ((C)->chdr & FLG_MASK) | (SZ)))
@@ -506,11 +622,11 @@ static void mbc_free(Allctr_t *allctr, void *p);
 #if HAVE_ERTS_MSEG
 
 static ERTS_INLINE void *
-alcu_mseg_alloc(Allctr_t *allctr, Uint *size_p)
+alcu_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
 {
     void *res;
 
-    res = erts_mseg_alloc_opt(allctr->alloc_no, size_p, &allctr->mseg_opt);
+    res = erts_mseg_alloc_opt(allctr->alloc_no, size_p, flags, &allctr->mseg_opt);
     INC_CC(allctr->calls.mseg_alloc);
     return res;
 }
@@ -521,7 +637,7 @@ alcu_mseg_realloc(Allctr_t *allctr, void *seg, Uint old_size, Uint *new_size_p)
     void *res;
 
     res = erts_mseg_realloc_opt(allctr->alloc_no, seg, old_size, new_size_p,
-				&allctr->mseg_opt);
+				ERTS_MSEG_FLG_NONE, &allctr->mseg_opt);
     INC_CC(allctr->calls.mseg_realloc);
     return res;
 }
@@ -765,13 +881,9 @@ erts_alcu_fix_alloc_shrink(Allctr_t *allctr, erts_aint32_t flgs)
 #define ERTS_ALCU_DD_FIX_TYPE_OFFS \
   ((sizeof(ErtsAllctrDDBlock_t)-1)/sizeof(UWord) + 1)
 
-#define ERTS_AU_PREF_ALLOC_IX_MASK \
-   ((((UWord) 1) << ERTS_AU_PREF_ALLOC_BITS) - 1)
-#define ERTS_AU_PREF_ALLOC_SIZE_MASK \
-   ((((UWord) 1) << (sizeof(UWord)*8 - ERTS_AU_PREF_ALLOC_BITS)) - 1)
 
-static ERTS_INLINE int
-get_pref_allctr(void *extra, Allctr_t **allctr)
+static ERTS_INLINE Allctr_t*
+get_pref_allctr(void *extra)
 {
     ErtsAllocatorThrSpec_t *tspec = (ErtsAllocatorThrSpec_t *) extra;
     int pref_ix;
@@ -781,34 +893,33 @@ get_pref_allctr(void *extra, Allctr_t **allctr)
     ASSERT(sizeof(UWord) == sizeof(Allctr_t *));
     ASSERT(0 <= pref_ix && pref_ix < tspec->size);
 
-    *allctr = tspec->allctr[pref_ix];
-    return pref_ix;
+    return tspec->allctr[pref_ix];
 }
 
-static ERTS_INLINE void *
-get_used_allctr(void *extra, void *p, Allctr_t **allctr, UWord *sizep)
+/* SMP note:
+ * get_used_allctr() must be safe WITHOUT locking the allocator while
+ * concurrent threads may be updating adjacent blocks.
+ * We rely on getting a consistent result (without atomic op) when reading
+ * the block header word even if a concurrent thread is updating
+ * the "PREV_FREE" flag bit. 
+ */
+static ERTS_INLINE Allctr_t*
+get_used_allctr(void *extra, void *p, UWord *sizep)
 {
-    ErtsAllocatorThrSpec_t *tspec = (ErtsAllocatorThrSpec_t *) extra;
-    void *ptr = (void *) (((char *) p) - sizeof(UWord));
-    UWord ainfo = *((UWord *) ptr);
-    int aix = (int) (ainfo & ERTS_AU_PREF_ALLOC_IX_MASK);
-    *allctr = tspec->allctr[aix];
-    if (sizep)
-	*sizep = ((ainfo >> ERTS_AU_PREF_ALLOC_BITS)
-		  & ERTS_AU_PREF_ALLOC_SIZE_MASK);
-    return ptr;
-}
+    Block_t* blk = UMEM2BLK(p);
+    Carrier_t* crr;
 
-static ERTS_INLINE void *
-put_used_allctr(void *p, int ix, UWord size)
-{
-    UWord ainfo = (size >= ERTS_AU_PREF_ALLOC_SIZE_MASK
-		   ? ERTS_AU_PREF_ALLOC_SIZE_MASK
-		   : size);
-    ainfo <<= ERTS_AU_PREF_ALLOC_BITS;
-    ainfo |= (UWord) ix;
-    *((UWord *) p) = ainfo;
-    return (void *) (((char *) p) + sizeof(UWord));
+    if (IS_SBC_BLK(blk)) {
+	crr = BLK_TO_SBC(blk);
+	if (sizep)
+	    *sizep = SBC_BLK_SZ(blk) - ABLK_HDR_SZ;  
+    }
+    else {
+	crr = ABLK_TO_MBC(blk);
+	if (sizep)
+	    *sizep = MBC_ABLK_SZ(blk) - ABLK_HDR_SZ;
+    }
+    return crr->allctr;
 }
 
 static void
@@ -1209,10 +1320,8 @@ mbc_alloc_block(Allctr_t *allctr, Uint size, Uint *blk_szp, Uint32 *alcu_flgsp)
 	if ((*alcu_flgsp) & ERTS_ALCU_FLG_SBMBC) 
 	    blk = create_sbmbc(allctr, get_blk_sz);
 	else {
-#if HALFWORD_HEAP
-	    blk = create_carrier(allctr, get_blk_sz, CFLG_MBC|CFLG_FORCE_MSEG);
-#else
 	    blk = create_carrier(allctr, get_blk_sz, CFLG_MBC);
+#if !HALFWORD_HEAP && !HAVE_SUPER_ALIGNED_MB_CARRIERS
 	    if (!blk) {
 		/* Emergency! We couldn't create the carrier as we wanted.
 		   Try to place it in a sys_alloced sbc. */
@@ -1242,6 +1351,7 @@ mbc_alloc_finalize(Allctr_t *allctr,
 		   Block_t *blk,
 		   Uint org_blk_sz,
 		   UWord flags,
+		   Carrier_t *crr,
 		   Uint want_blk_sz,
 		   int valid_blk_info,
 		   Uint32 alcu_flgs)
@@ -1262,22 +1372,18 @@ mbc_alloc_finalize(Allctr_t *allctr,
 	/* Shrink block... */
 	blk_sz = want_blk_sz;
 	nxt_blk_sz = org_blk_sz - blk_sz;
-	SET_BLK_HDR(blk,
-		    blk_sz,
-		    SBH_THIS_ALLOCED|SBH_NOT_LAST_BLK|prev_free_flg);
+	SET_MBC_ABLK_HDR(blk, blk_sz, prev_free_flg, crr);
 
-	nxt_blk = NXT_BLK(blk);
-	SET_BLK_HDR(nxt_blk,
-		    nxt_blk_sz,
-		    (SBH_THIS_FREE
-		     | SBH_PREV_ALLOCED
-		     | (flags & LAST_BLK_HDR_FLG)));
+	nxt_blk = BLK_AFTER(blk, blk_sz);
+	SET_MBC_FBLK_HDR(nxt_blk, nxt_blk_sz,
+			 SBH_THIS_FREE|(flags & LAST_BLK_HDR_FLG),
+			 crr);
 
 	if (!(flags & LAST_BLK_HDR_FLG)) {
 	    SET_BLK_SZ_FTR(nxt_blk, nxt_blk_sz);
 	    if (!valid_blk_info) {
-		Block_t *nxt_nxt_blk = NXT_BLK(nxt_blk);
-		SET_PREV_BLK_FREE(nxt_nxt_blk);
+		Block_t *nxt_nxt_blk = BLK_AFTER(nxt_blk, nxt_blk_sz);
+		SET_PREV_BLK_FREE(allctr, nxt_nxt_blk);
 	    }
 	}
 	(*allctr->link_free_block)(allctr, nxt_blk, alcu_flgs);
@@ -1291,40 +1397,40 @@ mbc_alloc_finalize(Allctr_t *allctr,
 	       || nxt_blk == PREV_BLK(NXT_BLK(nxt_blk)));
 	ASSERT((flags & LAST_BLK_HDR_FLG)
 	       || IS_PREV_BLK_FREE(NXT_BLK(nxt_blk)));
-	ASSERT(nxt_blk_sz == BLK_SZ(nxt_blk));
+	ASSERT(nxt_blk_sz == MBC_BLK_SZ(nxt_blk));
 	ASSERT(nxt_blk_sz % sizeof(Unit_t) == 0);
 	ASSERT(nxt_blk_sz >= allctr->min_block_size);
+	ASSERT(ABLK_TO_MBC(blk) == crr);
+	ASSERT(FBLK_TO_MBC(nxt_blk) == crr);
     }
     else {
+	ASSERT(org_blk_sz <= MBC_ABLK_SZ_MASK);
 	blk_sz = org_blk_sz;
 	if (flags & LAST_BLK_HDR_FLG) {
 	    if (valid_blk_info)
 		SET_BLK_ALLOCED(blk);
 	    else
-		SET_BLK_HDR(blk,
-			    blk_sz,
-			    SBH_THIS_ALLOCED|SBH_LAST_BLK|prev_free_flg);
+		SET_MBC_ABLK_HDR(blk, blk_sz, SBH_LAST_BLK|prev_free_flg, crr);
 	}
 	else {
 	    if (valid_blk_info)
 		SET_BLK_ALLOCED(blk);
 	    else
-		SET_BLK_HDR(blk,
-			    blk_sz,
-			    SBH_THIS_ALLOCED|SBH_NOT_LAST_BLK|prev_free_flg);
-	    nxt_blk = NXT_BLK(blk);
+		SET_MBC_ABLK_HDR(blk, blk_sz, prev_free_flg, crr);
+	    nxt_blk = BLK_AFTER(blk, blk_sz);
 	    SET_PREV_BLK_ALLOCED(nxt_blk);
 	}
 
 	ASSERT((flags & LAST_BLK_HDR_FLG)
 	       ? IS_LAST_BLK(blk)
 	       : IS_NOT_LAST_BLK(blk));
+	ASSERT(ABLK_TO_MBC(blk) == crr);
     }
 
     STAT_MBC_BLK_ALLOC(allctr, blk_sz, alcu_flgs);
 
     ASSERT(IS_ALLOCED_BLK(blk));
-    ASSERT(blk_sz == BLK_SZ(blk));
+    ASSERT(blk_sz == MBC_BLK_SZ(blk));
     ASSERT(blk_sz % sizeof(Unit_t) == 0);
     ASSERT(blk_sz >= allctr->min_block_size);
     ASSERT(blk_sz >= want_blk_sz);
@@ -1348,8 +1454,9 @@ mbc_alloc(Allctr_t *allctr, Uint size)
     if (IS_MBC_BLK(blk))
 	mbc_alloc_finalize(allctr,
 			   blk,
-			   BLK_SZ(blk),
+			   MBC_FBLK_SZ(blk),
 			   GET_BLK_HDR_FLGS(blk),
+			   FBLK_TO_MBC(blk),
 			   blk_sz,
 			   1,
 			   alcu_flgs);
@@ -1370,7 +1477,7 @@ mbc_free(Allctr_t *allctr, void *p)
     ASSERT(p);
 
     blk = UMEM2BLK(p);
-    blk_sz = BLK_SZ(blk);
+    blk_sz = MBC_ABLK_SZ(blk);
     if (blk_sz < allctr->sbmbc_threshold)
 	alcu_flgs |= ERTS_ALCU_FLG_SBMBC;
 
@@ -1381,17 +1488,18 @@ mbc_free(Allctr_t *allctr, void *p)
 
     STAT_MBC_BLK_FREE(allctr, blk_sz, alcu_flgs);
 
-    is_first_blk = IS_FIRST_BLK(blk);
+    is_first_blk = IS_MBC_FIRST_ABLK(allctr, blk);
     is_last_blk = IS_LAST_BLK(blk);
 
-    if (!is_first_blk && IS_PREV_BLK_FREE(blk)) {
+    if (IS_PREV_BLK_FREE(blk)) {
+	ASSERT(!is_first_blk); 
 	/* Coalesce with previous block... */
 	blk = PREV_BLK(blk);
 	(*allctr->unlink_free_block)(allctr, blk, alcu_flgs);
 
-	blk_sz += BLK_SZ(blk);
-	is_first_blk = IS_FIRST_BLK(blk);
-	SET_BLK_SZ(blk, blk_sz);
+	blk_sz += MBC_FBLK_SZ(blk);
+	is_first_blk = IS_MBC_FIRST_FBLK(allctr, blk);
+	SET_MBC_FBLK_SZ(blk, blk_sz);
     }
     else {
 	SET_BLK_FREE(blk);
@@ -1400,12 +1508,12 @@ mbc_free(Allctr_t *allctr, void *p)
     if (is_last_blk)
 	SET_LAST_BLK(blk);
     else {
-	nxt_blk = NXT_BLK(blk);
+	nxt_blk = BLK_AFTER(blk, blk_sz);
 	if (IS_FREE_BLK(nxt_blk)) {
 	    /* Coalesce with next block... */
 	    (*allctr->unlink_free_block)(allctr, nxt_blk, alcu_flgs);
-	    blk_sz += BLK_SZ(nxt_blk);
-	    SET_BLK_SZ(blk, blk_sz);
+	    blk_sz += MBC_FBLK_SZ(nxt_blk);
+	    SET_MBC_FBLK_SZ(blk, blk_sz);
 
 	    is_last_blk = IS_LAST_BLK(nxt_blk);
 	    if (is_last_blk) 
@@ -1416,26 +1524,26 @@ mbc_free(Allctr_t *allctr, void *p)
 	    }
 	}
 	else {
-	    SET_PREV_BLK_FREE(nxt_blk);
+	    SET_PREV_BLK_FREE(allctr, nxt_blk);
 	    SET_NOT_LAST_BLK(blk);
 	    SET_BLK_SZ_FTR(blk, blk_sz);
 	}
 
     }
 
-    ASSERT(is_last_blk  ? IS_LAST_BLK(blk)  : IS_NOT_LAST_BLK(blk));
-    ASSERT(is_first_blk ? IS_FIRST_BLK(blk) : IS_NOT_FIRST_BLK(blk));
     ASSERT(IS_FREE_BLK(blk));
+    ASSERT(!is_last_blk  == !IS_LAST_BLK(blk));
+    ASSERT(!is_first_blk == !IS_MBC_FIRST_FBLK(allctr, blk));
     ASSERT(is_first_blk || IS_PREV_BLK_ALLOCED(blk));
     ASSERT(is_last_blk  || IS_PREV_BLK_FREE(NXT_BLK(blk)));
-    ASSERT(blk_sz == BLK_SZ(blk));
+    ASSERT(blk_sz == MBC_BLK_SZ(blk));
     ASSERT(is_last_blk || blk == PREV_BLK(NXT_BLK(blk)));
     ASSERT(blk_sz % sizeof(Unit_t) == 0);
     ASSERT(IS_MBC_BLK(blk));
 
     if (is_first_blk
 	&& is_last_blk
-	&& allctr->main_carrier != FBLK2MBC(allctr, blk)) {
+	&& allctr->main_carrier != FIRST_BLK_TO_MBC(allctr, blk)) {
 	if (alcu_flgs & ERTS_ALCU_FLG_SBMBC)
 	    destroy_sbmbc(allctr, blk);
 	else
@@ -1472,7 +1580,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
     ASSERT(size < allctr->sbc_threshold);
 
     blk = (Block_t *) UMEM2BLK(p);
-    old_blk_sz = BLK_SZ(blk);
+    old_blk_sz = MBC_ABLK_SZ(blk);
 
     ASSERT(old_blk_sz >= allctr->min_block_size);
 
@@ -1497,6 +1605,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 	return p;
     else if (blk_sz < old_blk_sz) {
 	/* Shrink block... */
+	Carrier_t* crr;
 	Block_t *nxt_nxt_blk;
 	Uint diff_sz_val = old_blk_sz - blk_sz;
 	Uint old_blk_sz_val = old_blk_sz;
@@ -1516,16 +1625,18 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 		return NULL;
 
 	    cand_blk_sz = old_blk_sz;
-	    if (!IS_PREV_BLK_FREE(blk) || IS_FIRST_BLK(blk))
+	    if (!IS_PREV_BLK_FREE(blk)) {
 		cand_blk = blk;
+	    }
 	    else {
+		ASSERT(!IS_MBC_FIRST_ABLK(allctr, blk));
 		cand_blk = PREV_BLK(blk);
 		cand_blk_sz += PREV_BLK_SZ(blk);
 	    }
 	    if (!is_last_blk) {
-		nxt_blk = NXT_BLK(blk);
+		nxt_blk = BLK_AFTER(blk, old_blk_sz);
 		if (IS_FREE_BLK(nxt_blk))
-		    cand_blk_sz += BLK_SZ(nxt_blk);
+		    cand_blk_sz += MBC_FBLK_SZ(nxt_blk);
 	    }
 
 	    new_blk = (*allctr->get_free_block)(allctr,
@@ -1541,52 +1652,48 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 
 	nxt_blk_sz = old_blk_sz - blk_sz;
 
-	if ((is_last_blk || IS_ALLOCED_BLK(NXT_BLK(blk)))
+	if ((is_last_blk || IS_ALLOCED_BLK(BLK_AFTER(blk,old_blk_sz)))
 	    && (nxt_blk_sz < allctr->min_block_size))
 	    return p;
 
 	HARD_CHECK_BLK_CARRIER(allctr, blk);
 
-	SET_BLK_SZ(blk, blk_sz);
+	nxt_nxt_blk = BLK_AFTER(blk, old_blk_sz);
+
+	SET_MBC_ABLK_SZ(blk, blk_sz);
 	SET_NOT_LAST_BLK(blk);
 
-	nxt_blk = NXT_BLK(blk);
-	SET_BLK_HDR(nxt_blk,
-		    nxt_blk_sz,
-		    SBH_THIS_FREE|SBH_PREV_ALLOCED|SBH_NOT_LAST_BLK);
+	nxt_blk = BLK_AFTER(blk, blk_sz);
 
 	STAT_MBC_BLK_FREE(allctr, old_blk_sz, alcu_flgs);
 	STAT_MBC_BLK_ALLOC(allctr, blk_sz, alcu_flgs);
 
-	ASSERT(BLK_SZ(blk) >= allctr->min_block_size);
+	ASSERT(MBC_BLK_SZ(blk) >= allctr->min_block_size);
 
-	if (is_last_blk)
-	    SET_LAST_BLK(nxt_blk);
-	else {
-	    nxt_nxt_blk = NXT_BLK(nxt_blk);
+	if (!is_last_blk) {
 	    if (IS_FREE_BLK(nxt_nxt_blk)) {
 		/* Coalesce with next free block... */
-		nxt_blk_sz += BLK_SZ(nxt_nxt_blk);
+		nxt_blk_sz += MBC_FBLK_SZ(nxt_nxt_blk);
 		(*allctr->unlink_free_block)(allctr, nxt_nxt_blk, alcu_flgs);
-		SET_BLK_SZ(nxt_blk, nxt_blk_sz);
 
-		is_last_blk = IS_LAST_BLK(nxt_nxt_blk);
-		if (is_last_blk)
-		    SET_LAST_BLK(nxt_blk);
-		else
-		    SET_BLK_SZ_FTR(nxt_blk, nxt_blk_sz);
+		is_last_blk = GET_LAST_BLK_HDR_FLG(nxt_nxt_blk);
 	    }
 	    else {
-		SET_BLK_SZ_FTR(nxt_blk, nxt_blk_sz);
-		SET_PREV_BLK_FREE(nxt_nxt_blk);
+		SET_PREV_BLK_FREE(allctr, nxt_nxt_blk);
 	    }
+	    SET_BLK_SZ_FTR(nxt_blk, nxt_blk_sz);
 	}
 
+	crr = ABLK_TO_MBC(blk);
+	SET_MBC_FBLK_HDR(nxt_blk, nxt_blk_sz,
+			SBH_THIS_FREE | (is_last_blk ? SBH_LAST_BLK : 0),
+			crr);
+
 	(*allctr->link_free_block)(allctr, nxt_blk, alcu_flgs);
 
 
 	ASSERT(IS_ALLOCED_BLK(blk));
-	ASSERT(blk_sz == BLK_SZ(blk));
+	ASSERT(blk_sz == MBC_BLK_SZ(blk));
 	ASSERT(blk_sz % sizeof(Unit_t) == 0);
 	ASSERT(blk_sz >= allctr->min_block_size);
 	ASSERT(blk_sz >= size + ABLK_HDR_SZ);
@@ -1594,14 +1701,15 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
     
 	ASSERT(IS_FREE_BLK(nxt_blk));
 	ASSERT(IS_PREV_BLK_ALLOCED(nxt_blk));
-	ASSERT(nxt_blk_sz == BLK_SZ(nxt_blk));
+	ASSERT(nxt_blk_sz == MBC_BLK_SZ(nxt_blk));
 	ASSERT(nxt_blk_sz % sizeof(Unit_t) == 0);
 	ASSERT(nxt_blk_sz >= allctr->min_block_size);
 	ASSERT(IS_MBC_BLK(nxt_blk));
 	ASSERT(is_last_blk ? IS_LAST_BLK(nxt_blk) : IS_NOT_LAST_BLK(nxt_blk));
 	ASSERT(is_last_blk || nxt_blk == PREV_BLK(NXT_BLK(nxt_blk)));
 	ASSERT(is_last_blk || IS_PREV_BLK_FREE(NXT_BLK(nxt_blk)));
-
+	ASSERT(FBLK_TO_MBC(nxt_blk) == crr);
+		    
 	HARD_CHECK_BLK_CARRIER(allctr, blk);
 
 	return p;
@@ -1610,8 +1718,8 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
     /* Need larger block... */
 
     if (!is_last_blk) {
-	nxt_blk = NXT_BLK(blk);
-	nxt_blk_sz = BLK_SZ(nxt_blk);
+	nxt_blk = BLK_AFTER(blk, old_blk_sz);
+	nxt_blk_sz = MBC_BLK_SZ(nxt_blk);
 	if (IS_FREE_BLK(nxt_blk) && get_blk_sz <= old_blk_sz + nxt_blk_sz) {
 	    /* Grow into next block... */
 
@@ -1624,7 +1732,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 	    if (nxt_blk_sz < allctr->min_block_size) {
 		blk_sz += nxt_blk_sz;
 
-		SET_BLK_SZ(blk, blk_sz);
+		SET_MBC_ABLK_SZ(blk, blk_sz);
 
 		if (is_last_blk) {
 		    SET_LAST_BLK(blk);
@@ -1633,21 +1741,20 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 #endif
 		}
 		else {
-		    nxt_blk = NXT_BLK(blk);
+		    nxt_blk = BLK_AFTER(blk, blk_sz);
 		    SET_PREV_BLK_ALLOCED(nxt_blk);
 #ifdef DEBUG
 		    is_last_blk = IS_LAST_BLK(nxt_blk);
-		    nxt_blk_sz = BLK_SZ(nxt_blk);
+		    nxt_blk_sz = MBC_BLK_SZ(nxt_blk);
 #endif
 		}
 	    }
 	    else {
-		SET_BLK_SZ(blk, blk_sz);
+		Carrier_t* crr = ABLK_TO_MBC(blk);
+		SET_MBC_ABLK_SZ(blk, blk_sz);
 
-		nxt_blk = NXT_BLK(blk);
-		SET_BLK_HDR(nxt_blk,
-			    nxt_blk_sz,
-			    SBH_THIS_FREE|SBH_PREV_ALLOCED|SBH_NOT_LAST_BLK);
+		nxt_blk = BLK_AFTER(blk, blk_sz);
+		SET_MBC_FBLK_HDR(nxt_blk, nxt_blk_sz, SBH_THIS_FREE, crr);
 
 		if (is_last_blk)
 		    SET_LAST_BLK(nxt_blk);
@@ -1657,6 +1764,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 		(*allctr->link_free_block)(allctr, nxt_blk, alcu_flgs);
 
 		ASSERT(IS_FREE_BLK(nxt_blk));
+		ASSERT(FBLK_TO_MBC(nxt_blk) == crr);
 	    }
 
 	    STAT_MBC_BLK_FREE(allctr, old_blk_sz, alcu_flgs);
@@ -1664,14 +1772,14 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 
 
 	    ASSERT(IS_ALLOCED_BLK(blk));
-	    ASSERT(blk_sz == BLK_SZ(blk));
+	    ASSERT(blk_sz == MBC_BLK_SZ(blk));
 	    ASSERT(blk_sz % sizeof(Unit_t) == 0);
 	    ASSERT(blk_sz >= allctr->min_block_size);
 	    ASSERT(blk_sz >= size + ABLK_HDR_SZ);
 	    ASSERT(IS_MBC_BLK(blk));
 
 	    ASSERT(!nxt_blk || IS_PREV_BLK_ALLOCED(nxt_blk));
-	    ASSERT(!nxt_blk || nxt_blk_sz == BLK_SZ(nxt_blk));
+	    ASSERT(!nxt_blk || nxt_blk_sz == MBC_BLK_SZ(nxt_blk));
 	    ASSERT(!nxt_blk || nxt_blk_sz % sizeof(Unit_t) == 0);
 	    ASSERT(!nxt_blk || nxt_blk_sz >= allctr->min_block_size);
 	    ASSERT(!nxt_blk || IS_MBC_BLK(nxt_blk));
@@ -1696,18 +1804,19 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 
     /* Need to grow in another block */
 
-    if (!IS_PREV_BLK_FREE(blk) || IS_FIRST_BLK(blk)) {
+    if (!IS_PREV_BLK_FREE(blk)) {
 	cand_blk = NULL;
 	cand_blk_sz = 0;
     }
     else {
+	ASSERT(!IS_MBC_FIRST_ABLK(allctr, blk));
 	cand_blk = PREV_BLK(blk);
 	cand_blk_sz = old_blk_sz + PREV_BLK_SZ(blk);
 
 	if (!is_last_blk) {
-	    nxt_blk = NXT_BLK(blk);
+	    nxt_blk = BLK_AFTER(blk, old_blk_sz);
 	    if (IS_FREE_BLK(nxt_blk))
-		cand_blk_sz += BLK_SZ(nxt_blk);
+		cand_blk_sz += MBC_FBLK_SZ(nxt_blk);
 	}
     }
 
@@ -1743,8 +1852,9 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 	if (new_blk) {
 	    mbc_alloc_finalize(allctr,
 			       new_blk,
-			       BLK_SZ(new_blk),
+			       MBC_FBLK_SZ(new_blk),
 			       GET_BLK_HDR_FLGS(new_blk),
+			       FBLK_TO_MBC(new_blk),
 			       blk_sz,
 			       1,
 			       alcu_flgs);
@@ -1754,6 +1864,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 	    return new_p;
 	}
 	else {
+	    Carrier_t* crr;
 	    Uint new_blk_sz;
 	    UWord new_blk_flgs;
 	    Uint prev_blk_sz;
@@ -1774,10 +1885,10 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 	    if (is_last_blk) 
 		new_blk_flgs |= LAST_BLK_HDR_FLG;
 	    else {
-		nxt_blk = NXT_BLK(blk);
+		nxt_blk = BLK_AFTER(blk, old_blk_sz);
 		if (IS_FREE_BLK(nxt_blk)) {
 		    new_blk_flgs |= GET_LAST_BLK_HDR_FLG(nxt_blk);
-		    new_blk_sz += BLK_SZ(nxt_blk);
+		    new_blk_sz += MBC_FBLK_SZ(nxt_blk);
 		    (*allctr->unlink_free_block)(allctr, nxt_blk, alcu_flgs);
 		}
 	    }
@@ -1790,6 +1901,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 
 	    new_p = BLK2UMEM(new_blk);
 	    blk_cpy_sz = MIN(blk_sz, old_blk_sz);
+	    crr = FBLK_TO_MBC(new_blk);
 
 	    if (prev_blk_sz >= blk_cpy_sz)
 		sys_memcpy(new_p, p, blk_cpy_sz - ABLK_HDR_SZ);
@@ -1800,6 +1912,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 			       new_blk,
 			       new_blk_sz,
 			       new_blk_flgs,
+			       crr,
 			       blk_sz,
 			       0,
 			       alcu_flgs);
@@ -1815,35 +1928,40 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs)
 #ifdef DEBUG
 
 #if HAVE_ERTS_MSEG
-#define ASSERT_MSEG_UNIT_SIZE_MULTIPLE(CSZ) ASSERT((CSZ) % mseg_unit_size == 0)
+#define ASSERT_MSEG_UNIT_SIZE_MULTIPLE(CSZ) ASSERT((CSZ) % MSEG_UNIT_SZ == 0)
 #else
 #define ASSERT_MSEG_UNIT_SIZE_MULTIPLE(CSZ)
 #endif
 
-#define CHECK_1BLK_CARRIER(A, SBC, MSEGED, C, CSZ, B, BSZ)		\
-do {									\
-    ASSERT(IS_FIRST_BLK((B)));						\
-    ASSERT(IS_LAST_BLK((B)));						\
-    ASSERT((CSZ) == CARRIER_SZ((C)));					\
-    ASSERT((BSZ) == BLK_SZ((B)));					\
-    ASSERT((BSZ) % sizeof(Unit_t) == 0);				\
-    if ((SBC)) {							\
-	ASSERT(IS_SBC_BLK((B)));					\
-	ASSERT(IS_SB_CARRIER((C)));					\
-    }									\
-    else {								\
-	ASSERT(IS_MBC_BLK((B)));					\
-	ASSERT(IS_MB_CARRIER((C)));					\
-    }									\
-    if ((MSEGED)) {							\
-	ASSERT(IS_MSEG_CARRIER((C)));					\
-	ASSERT_MSEG_UNIT_SIZE_MULTIPLE((CSZ));				\
-    }									\
-    else {								\
-	ASSERT(IS_SYS_ALLOC_CARRIER((C)));					\
-	ASSERT((CSZ) % sizeof(Unit_t) == 0);				\
-    }									\
-} while (0)
+static void CHECK_1BLK_CARRIER(Allctr_t* A, int SBC, int MSEGED, Carrier_t* C,
+			       UWord CSZ, Block_t* B, UWord BSZ) 
+{
+    ASSERT(IS_LAST_BLK((B)));
+    ASSERT((CSZ) == CARRIER_SZ((C)));
+    ASSERT((BSZ) % sizeof(Unit_t) == 0);
+    if ((SBC)) {
+	ASSERT((BSZ) == SBC_BLK_SZ((B)));
+	ASSERT((char*)B == (char*)C + SBC_HEADER_SIZE);
+	ASSERT(IS_SBC_BLK((B)));
+	ASSERT(IS_SB_CARRIER((C)));
+    }
+    else {
+	ASSERT(IS_FREE_BLK(B));
+	ASSERT((BSZ) == MBC_FBLK_SZ((B)));
+	ASSERT(IS_MBC_FIRST_FBLK(A, (B)));
+	ASSERT(IS_MBC_BLK((B)));
+	ASSERT(IS_MB_CARRIER((C)));
+	ASSERT(FBLK_TO_MBC(B) == (C));
+    }
+    if ((MSEGED)) {
+	ASSERT(IS_MSEG_CARRIER((C)));
+	ASSERT_MSEG_UNIT_SIZE_MULTIPLE((CSZ));
+    }
+    else {
+	ASSERT(IS_SYS_ALLOC_CARRIER((C)));
+	ASSERT((CSZ) % sizeof(Unit_t) == 0);
+    }
+}
 
 #else
 #define CHECK_1BLK_CARRIER(A, SBC, MSEGED, C, CSZ, B, BSZ)
@@ -1865,37 +1983,18 @@ create_sbmbc(Allctr_t *allctr, Uint umem_sz)
 	crr = erts_alloc(ERTS_ALC_T_SBMBC, crr_sz);
 
     INC_CC(allctr->calls.sbmbc_alloc);
-    SET_CARRIER_HDR(crr, crr_sz, SCH_SYS_ALLOC|SCH_MBC);
-
-    blk = MBC2FBLK(allctr, crr);
+    SET_CARRIER_HDR(crr, crr_sz, SCH_SYS_ALLOC|SCH_MBC, allctr);
 
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-    if (allctr->mbc_header_size % sizeof(Unit_t) == 0)
-	crr_sz -= sizeof(UWord);
-#endif
+    blk = MBC_TO_FIRST_BLK(allctr, crr);
 
-    blk_sz = UNIT_FLOOR(crr_sz - allctr->mbc_header_size);
+    blk_sz = UNIT_FLOOR(crr_sz - MBC_HEADER_SIZE(allctr));
 
-    SET_MBC_BLK_FTR(((UWord *) blk)[-1]);
-    SET_BLK_HDR(blk, blk_sz, SBH_THIS_FREE|SBH_PREV_FREE|SBH_LAST_BLK);
-
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-    *((Carrier_t **) NXT_BLK(blk)) = crr;
-#endif
+    SET_MBC_FBLK_HDR(blk, blk_sz, SBH_THIS_FREE|SBH_LAST_BLK, crr);
 
     link_carrier(&allctr->sbmbc_list, crr);
 
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-    if (allctr->mbc_header_size % sizeof(Unit_t) == 0)
-	crr_sz += sizeof(UWord);
-#endif
-
     STAT_SBMBC_ALLOC(allctr, crr_sz);
     CHECK_1BLK_CARRIER(allctr, 0, 0, crr, crr_sz, blk, blk_sz);
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-    if (allctr->mbc_header_size % sizeof(Unit_t) == 0)
-	crr_sz -= sizeof(UWord);
-#endif
     if (allctr->creating_mbc)
 	(*allctr->creating_mbc)(allctr, crr, ERTS_ALCU_FLG_SBMBC);
 
@@ -1909,11 +2008,10 @@ destroy_sbmbc(Allctr_t *allctr, Block_t *blk)
     Uint crr_sz;
     Carrier_t *crr;
 
-    ASSERT(IS_FIRST_BLK(blk));
-
     ASSERT(IS_MBC_BLK(blk));
+    ASSERT(IS_MBC_FIRST_FBLK(allctr, blk));
 
-    crr = FBLK2MBC(allctr, blk);
+    crr = FIRST_BLK_TO_MBC(allctr, blk);
     crr_sz = CARRIER_SZ(crr);
 
 #ifdef DEBUG
@@ -1952,14 +2050,25 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
     Uint blk_sz, bcrr_sz, crr_sz;
 #if HAVE_ERTS_MSEG
     int have_tried_sys_alloc = 0, have_tried_mseg = 0;
+    Uint mseg_flags;
 #endif
 #ifdef DEBUG
     int is_mseg = 0;
 #endif
 
+#if HALFWORD_HEAP
+    flags |= CFLG_FORCE_MSEG;
+#elif HAVE_SUPER_ALIGNED_MB_CARRIERS
+    if (flags & CFLG_MBC) {
+	flags |= CFLG_FORCE_MSEG;
+    }
+#endif
+
     ASSERT((flags & CFLG_SBC && !(flags & CFLG_MBC))
 	   || (flags & CFLG_MBC && !(flags & CFLG_SBC)));
 
+    ASSERT(!(flags & CFLG_FORCE_MSEG && flags & CFLG_FORCE_SYS_ALLOC));
+
     blk_sz = UMEMSZ2BLKSZ(allctr, umem_sz);
 
 #if HAVE_ERTS_MSEG
@@ -1974,29 +2083,27 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 	if (allctr->sbcs.curr.norm.mseg.no >= allctr->max_mseg_sbcs)
 	    goto try_sys_alloc;
     }
+#if !HAVE_SUPER_ALIGNED_MB_CARRIERS
     else {
 	if (allctr->mbcs.curr.norm.mseg.no >= allctr->max_mseg_mbcs)
 	    goto try_sys_alloc;
     }
+#endif
 
  try_mseg:
 
     if (flags & CFLG_SBC) {
-	crr_sz = blk_sz + allctr->sbc_header_size;
+	crr_sz = blk_sz + SBC_HEADER_SIZE;
+	mseg_flags = ERTS_MSEG_FLG_NONE;
     }
     else {
 	crr_sz = (*allctr->get_next_mbc_size)(allctr);
-	if (crr_sz < allctr->mbc_header_size + blk_sz)
-	    crr_sz = allctr->mbc_header_size + blk_sz;
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (allctr->mbc_header_size % sizeof(Unit_t) == 0)
-	    crr_sz += sizeof(UWord);
-#endif
+	if (crr_sz < MBC_HEADER_SIZE(allctr) + blk_sz)
+	    crr_sz = MBC_HEADER_SIZE(allctr) + blk_sz;
+	mseg_flags = ERTS_MSEG_FLG_2POW;
     }
-    crr_sz = MSEG_UNIT_CEILING(crr_sz);
-    ASSERT(crr_sz % mseg_unit_size == 0);
 
-    crr = (Carrier_t *) alcu_mseg_alloc(allctr, &crr_sz);
+    crr = (Carrier_t *) alcu_mseg_alloc(allctr, &crr_sz, mseg_flags);
     if (!crr) {
 	have_tried_mseg = 1;
 	if (!(have_tried_sys_alloc || flags & CFLG_FORCE_MSEG))
@@ -2008,32 +2115,31 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
     is_mseg = 1;
 #endif
     if (flags & CFLG_SBC) {
-	SET_CARRIER_HDR(crr, crr_sz, SCH_MSEG|SCH_SBC);
+	SET_CARRIER_HDR(crr, crr_sz, SCH_MSEG|SCH_SBC, allctr);
 	STAT_MSEG_SBC_ALLOC(allctr, crr_sz, blk_sz);
 	goto sbc_final_touch;
     }
     else {
-	SET_CARRIER_HDR(crr, crr_sz, SCH_MSEG|SCH_MBC);
+#ifndef ARCH_64
+	ASSERT(crr_sz <= MBC_SZ_MAX_LIMIT);
+#endif
+	SET_CARRIER_HDR(crr, crr_sz, SCH_MSEG|SCH_MBC, allctr);
 	STAT_MSEG_MBC_ALLOC(allctr, crr_sz);
 	goto mbc_final_touch;
     }
 
  try_sys_alloc:
+
 #endif /* #if HAVE_ERTS_MSEG */
 
     if (flags & CFLG_SBC) {
-	bcrr_sz = blk_sz + allctr->sbc_header_size;
+	bcrr_sz = blk_sz + SBC_HEADER_SIZE;
     }
     else {
-	bcrr_sz = allctr->mbc_header_size + blk_sz;
+	bcrr_sz = MBC_HEADER_SIZE(allctr) + blk_sz;
 	if (!(flags & CFLG_MAIN_CARRIER)
 	    && bcrr_sz < allctr->smallest_mbc_size)
 	    bcrr_sz = allctr->smallest_mbc_size;
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (allctr->mbc_header_size % sizeof(Unit_t) == 0)
-	    bcrr_sz += sizeof(UWord);
-#endif
-
     }
 
     crr_sz = (flags & CFLG_FORCE_SIZE
@@ -2057,7 +2163,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 	}
     }
     if (flags & CFLG_SBC) {
-	SET_CARRIER_HDR(crr, crr_sz, SCH_SYS_ALLOC|SCH_SBC);
+	SET_CARRIER_HDR(crr, crr_sz, SCH_SYS_ALLOC|SCH_SBC, allctr);
 	STAT_SYS_ALLOC_SBC_ALLOC(allctr, crr_sz, blk_sz);
 
 #if HAVE_ERTS_MSEG
@@ -2066,8 +2172,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 
 	blk = SBC2BLK(allctr, crr);
 
-	SET_SBC_BLK_FTR(((UWord *) blk)[-1]);
-	SET_BLK_HDR(blk, blk_sz, SBH_THIS_ALLOCED|SBH_PREV_FREE|SBH_LAST_BLK);
+	SET_SBC_BLK_HDR(blk, blk_sz);
 
 	link_carrier(&allctr->sbc_list, crr);
 
@@ -2075,28 +2180,18 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 
     }
     else {
-	SET_CARRIER_HDR(crr, crr_sz, SCH_SYS_ALLOC|SCH_MBC);
+	SET_CARRIER_HDR(crr, crr_sz, SCH_SYS_ALLOC|SCH_MBC, allctr);
 	STAT_SYS_ALLOC_MBC_ALLOC(allctr, crr_sz);
 
 #if HAVE_ERTS_MSEG
     mbc_final_touch:
 #endif
 
-	blk = MBC2FBLK(allctr, crr);
-
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (allctr->mbc_header_size % sizeof(Unit_t) == 0)
-	    crr_sz -= sizeof(UWord);
-#endif
-
-	blk_sz = UNIT_FLOOR(crr_sz - allctr->mbc_header_size);
+	blk = MBC_TO_FIRST_BLK(allctr, crr);
 
-	SET_MBC_BLK_FTR(((UWord *) blk)[-1]);
-	SET_BLK_HDR(blk, blk_sz, SBH_THIS_FREE|SBH_PREV_FREE|SBH_LAST_BLK);
+	blk_sz = UNIT_FLOOR(crr_sz - MBC_HEADER_SIZE(allctr));
 
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	*((Carrier_t **) NXT_BLK(blk)) = crr;
-#endif
+	SET_MBC_FBLK_HDR(blk, blk_sz, SBH_THIS_FREE|SBH_LAST_BLK, crr);
 
 	if (flags & CFLG_MAIN_CARRIER) {
 	    ASSERT(!allctr->main_carrier);
@@ -2105,15 +2200,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 
 	link_carrier(&allctr->mbc_list, crr);
 
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (allctr->mbc_header_size % sizeof(Unit_t) == 0)
-	    crr_sz += sizeof(UWord);
-#endif
 	CHECK_1BLK_CARRIER(allctr, 0, is_mseg, crr, crr_sz, blk, blk_sz);
-#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (allctr->mbc_header_size % sizeof(Unit_t) == 0)
-	    crr_sz -= sizeof(UWord);
-#endif
 	if (allctr->creating_mbc)
 	    (*allctr->creating_mbc)(allctr, crr, 0);
 
@@ -2142,8 +2229,8 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 
     HARD_CHECK_BLK_CARRIER(allctr, old_blk);
 
-    old_blk_sz = BLK_SZ(old_blk);
-    old_crr = BLK2SBC(allctr, old_blk);
+    old_blk_sz = SBC_BLK_SZ(old_blk);
+    old_crr = BLK_TO_SBC(old_blk);
     old_crr_sz = CARRIER_SZ(old_crr);
     ASSERT(IS_SB_CARRIER(old_crr));
     ASSERT(IS_SBC_BLK(old_blk));
@@ -2157,7 +2244,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 
 	if (!(flags & CFLG_FORCE_SYS_ALLOC)) {
 
-	    new_crr_sz = new_blk_sz + allctr->sbc_header_size;
+	    new_crr_sz = new_blk_sz + SBC_HEADER_SIZE;
 	    new_crr_sz = MSEG_UNIT_CEILING(new_crr_sz);
 	    new_crr = (Carrier_t *) alcu_mseg_realloc(allctr,
 						      old_crr,
@@ -2166,7 +2253,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 	    if (new_crr) {
 		SET_CARRIER_SZ(new_crr, new_crr_sz);
 		new_blk = SBC2BLK(allctr, new_crr);
-		SET_BLK_SZ(new_blk, new_blk_sz);
+		SET_SBC_BLK_SZ(new_blk, new_blk_sz);
 		STAT_MSEG_SBC_ALLOC(allctr, new_crr_sz, new_blk_sz);
 		relink_carrier(&allctr->sbc_list, new_crr);
 		CHECK_1BLK_CARRIER(allctr, 1, 1, new_crr, new_crr_sz,
@@ -2174,6 +2261,11 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 		DEBUG_SAVE_ALIGNMENT(new_crr);
 		return new_blk;
 	    }
+#if HALFWORD_HEAP
+	    /* Old carrier unchanged; restore stat */
+	    STAT_MSEG_SBC_ALLOC(allctr, old_crr_sz, old_blk_sz);
+	    return NULL;
+#endif	    
 	    create_flags |= CFLG_FORCE_SYS_ALLOC; /* since mseg_realloc()
 						     failed */
 	}
@@ -2196,7 +2288,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
     else {
 	if (!(flags & CFLG_FORCE_MSEG)) {
 #endif /* #if HAVE_ERTS_MSEG */
-	    new_bcrr_sz = new_blk_sz + allctr->sbc_header_size;
+	    new_bcrr_sz = new_blk_sz + SBC_HEADER_SIZE;
 	    new_crr_sz = (flags & CFLG_FORCE_SIZE
 			  ? UNIT_CEILING(new_bcrr_sz)
 			  : SYS_ALLOC_CARRIER_CEILING(new_bcrr_sz));
@@ -2208,7 +2300,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 	    sys_realloc_success:
 		SET_CARRIER_SZ(new_crr, new_crr_sz);
 		new_blk = SBC2BLK(allctr, new_crr);
-		SET_BLK_SZ(new_blk, new_blk_sz);
+		SET_SBC_BLK_SZ(new_blk, new_blk_sz);
 		STAT_SYS_ALLOC_SBC_FREE(allctr, old_crr_sz, old_blk_sz);
 		STAT_SYS_ALLOC_SBC_ALLOC(allctr, new_crr_sz, new_blk_sz);
 		relink_carrier(&allctr->sbc_list, new_crr);
@@ -2218,7 +2310,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 		return new_blk;
 	    }
 	    else if (new_crr_sz > UNIT_CEILING(new_bcrr_sz)) {
-		new_crr_sz = new_blk_sz + allctr->sbc_header_size;
+		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,
 							 (void *) old_crr,
@@ -2262,11 +2354,9 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk)
     Uint is_mseg = 0;
 #endif
 
-    ASSERT(IS_FIRST_BLK(blk));
-
     if (IS_SBC_BLK(blk)) {
-	Uint blk_sz = BLK_SZ(blk);
-	crr = BLK2SBC(allctr, blk);
+	Uint blk_sz = SBC_BLK_SZ(blk);
+	crr = BLK_TO_SBC(blk);
 	crr_sz = CARRIER_SZ(crr);
 
 	ASSERT(IS_LAST_BLK(blk));
@@ -2276,7 +2366,7 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk)
 #if HAVE_ERTS_MSEG
 	if (IS_MSEG_CARRIER(crr)) {
 	    is_mseg++;
-	    ASSERT(crr_sz % mseg_unit_size == 0);
+	    ASSERT(crr_sz % MSEG_UNIT_SZ == 0);
 	    STAT_MSEG_SBC_FREE(allctr, crr_sz, blk_sz);
 	}
 	else
@@ -2287,7 +2377,8 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk)
 
     }
     else {
-	crr = FBLK2MBC(allctr, blk);
+	ASSERT(IS_MBC_FIRST_FBLK(allctr, blk));
+	crr = FIRST_BLK_TO_MBC(allctr, blk);
 	crr_sz = CARRIER_SZ(crr);
 
 #ifdef DEBUG
@@ -2305,7 +2396,7 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk)
 #if HAVE_ERTS_MSEG
 	if (IS_MSEG_CARRIER(crr)) {
 	    is_mseg++;
-	    ASSERT(crr_sz % mseg_unit_size == 0);
+	    ASSERT(crr_sz % MSEG_UNIT_SZ == 0);
 	    STAT_MSEG_MBC_FREE(allctr, crr_sz);
 	}
 	else
@@ -3430,12 +3521,7 @@ do_erts_alcu_alloc(ErtsAlcType_t type, void *extra, Uint size)
 	if (allctr->dd.use)
 	    ERTS_ALCU_HANDLE_DD_IN_OP(allctr, 1);
 #endif
-#if HALFWORD_HEAP
-	blk = create_carrier(allctr, size,
-			     CFLG_SBC | CFLG_FORCE_MSEG);
-#else
 	blk = create_carrier(allctr, size, CFLG_SBC);
-#endif
 	res = blk ? BLK2UMEM(blk) : NULL;
     }
     else
@@ -3506,24 +3592,20 @@ erts_alcu_alloc_thr_spec(ErtsAlcType_t type, void *extra, Uint size)
 void *
 erts_alcu_alloc_thr_pref(ErtsAlcType_t type, void *extra, Uint size)
 {
-    int pref_ix;
     Allctr_t *pref_allctr;
     void *res;
 
-    pref_ix = get_pref_allctr(extra, &pref_allctr);
+    pref_allctr = get_pref_allctr(extra);
 
     if (pref_allctr->thread_safe)
 	erts_mtx_lock(&pref_allctr->mutex);
 
     ERTS_ALCU_DBG_CHK_THR_ACCESS(pref_allctr);
 
-    res = do_erts_alcu_alloc(type, pref_allctr, size + sizeof(UWord));
+    res = do_erts_alcu_alloc(type, pref_allctr, size);
     if (pref_allctr->thread_safe)
 	erts_mtx_unlock(&pref_allctr->mutex);
 
-    if (res)
-	res = put_used_allctr(res, pref_ix, size);
-
     DEBUG_CHECK_ALIGNMENT(res);
 
 
@@ -3644,21 +3726,20 @@ erts_alcu_free_thr_pref(ErtsAlcType_t type, void *extra, void *p)
 {
     if (p) {
 	Allctr_t *pref_allctr, *used_allctr;
-	void *ptr;
 
-	get_pref_allctr(extra, &pref_allctr);
-	ptr = get_used_allctr(extra, p, &used_allctr, NULL);
+	pref_allctr = get_pref_allctr(extra);
+	used_allctr = get_used_allctr(extra, p, NULL);
 	if (pref_allctr != used_allctr)
 	    enqueue_dealloc_other_instance(type,
 					   used_allctr,
-					   ptr,
+					   p,
 					   (used_allctr->dd.ix
 					    - pref_allctr->dd.ix));
 	else {
 	    if (used_allctr->thread_safe)
 		erts_mtx_lock(&used_allctr->mutex);
 	    ERTS_ALCU_DBG_CHK_THR_ACCESS(used_allctr);
-	    do_erts_alcu_free(type, used_allctr, ptr);
+	    do_erts_alcu_free(type, used_allctr, p);
 	    if (used_allctr->thread_safe)
 		erts_mtx_unlock(&used_allctr->mutex);
 	}
@@ -3739,13 +3820,13 @@ do_erts_alcu_realloc(ErtsAlcType_t type,
 	if (IS_MBC_BLK(blk))
 	    res = mbc_realloc(allctr, p, size, alcu_flgs);
 	else {
-	    Uint used_sz = allctr->sbc_header_size + ABLK_HDR_SZ + size;
+	    Uint used_sz = SBC_HEADER_SIZE + ABLK_HDR_SZ + size;
 	    Uint crr_sz;
 	    Uint diff_sz_val;
 	    Uint crr_sz_val;
 
 #if HAVE_ERTS_MSEG
-	    if (IS_SYS_ALLOC_CARRIER(BLK2SBC(allctr, blk)))
+	    if (IS_SYS_ALLOC_CARRIER(BLK_TO_SBC(blk)))
 #endif
 		crr_sz = SYS_ALLOC_CARRIER_CEILING(used_sz);
 #if HAVE_ERTS_MSEG
@@ -3775,7 +3856,7 @@ do_erts_alcu_realloc(ErtsAlcType_t type,
 	    if (res) {
 		sys_memcpy((void*) res,
 			   (void*) p,
-			   MIN(BLK_SZ(blk) - ABLK_HDR_SZ, size));
+			   MIN(SBC_BLK_SZ(blk) - ABLK_HDR_SZ, size));
 		destroy_carrier(allctr, blk);
 	    }
 	}
@@ -3798,16 +3879,12 @@ do_erts_alcu_realloc(ErtsAlcType_t type,
 	else if (alcu_flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE)
 	    return NULL;
 	else {
-#if HALFWORD_HEAP
-	    new_blk = create_carrier(allctr, size, CFLG_SBC | CFLG_FORCE_MSEG);
-#else
 	    new_blk = create_carrier(allctr, size, CFLG_SBC);
-#endif
 	    if (new_blk) {
 		res = BLK2UMEM(new_blk);
 		sys_memcpy((void *) res,
 			   (void *) p,
-			   MIN(BLK_SZ(blk) - ABLK_HDR_SZ, size));
+			   MIN(MBC_ABLK_SZ(blk) - ABLK_HDR_SZ, size));
 		mbc_free(allctr, p);
 	    }
 	    else
@@ -3966,16 +4043,15 @@ static ERTS_INLINE void *
 realloc_thr_pref(ErtsAlcType_t type, void *extra, void *p, Uint size,
 		 int force_move)
 {
-    int pref_ix;
-    void *ptr, *res;
+    void *res;
     Allctr_t *pref_allctr, *used_allctr;
     UWord old_user_size;
 
     if (!p)
 	return erts_alcu_alloc_thr_pref(type, extra, size);
 
-    pref_ix = get_pref_allctr(extra, &pref_allctr);
-    ptr = get_used_allctr(extra, p, &used_allctr, &old_user_size);
+    pref_allctr = get_pref_allctr(extra);
+    used_allctr = get_used_allctr(extra, p, &old_user_size);
 
     ASSERT(used_allctr && pref_allctr);
 
@@ -3985,56 +4061,33 @@ realloc_thr_pref(ErtsAlcType_t type, void *extra, void *p, Uint size,
 	ERTS_ALCU_DBG_CHK_THR_ACCESS(used_allctr);
 	res = do_erts_alcu_realloc(type,
 				   used_allctr,
-				   ptr,
-				   size + sizeof(UWord),
+				   p,
+				   size,
 				   0);
 	if (used_allctr->thread_safe)
 	    erts_mtx_unlock(&used_allctr->mutex);
-	if (res)
-	    res = put_used_allctr(res, pref_ix, size);
     }
     else {
 	if (pref_allctr->thread_safe)
 	    erts_mtx_lock(&pref_allctr->mutex);
-	res = do_erts_alcu_alloc(type, pref_allctr, size + sizeof(UWord));
-	if (pref_allctr->thread_safe && (!force_move
-					 || used_allctr != pref_allctr))
+	res = do_erts_alcu_alloc(type, pref_allctr, size);
+	if (pref_allctr->thread_safe && used_allctr != pref_allctr) {
 	    erts_mtx_unlock(&pref_allctr->mutex);
+	}
 	if (res) {
-	    Block_t *blk;
-	    size_t cpy_size;
-
-	    res = put_used_allctr(res, pref_ix, size);
-
 	    DEBUG_CHECK_ALIGNMENT(res);
 
-	    blk = UMEM2BLK(ptr);
-	    if (old_user_size != ERTS_AU_PREF_ALLOC_SIZE_MASK)
-		cpy_size = old_user_size;
-	    else {
-		if (used_allctr->thread_safe && (!force_move
-						 || used_allctr != pref_allctr))
-		    erts_mtx_lock(&used_allctr->mutex);
-		ERTS_SMP_LC_ASSERT(!used_allctr->thread_safe ||
-				   erts_lc_mtx_is_locked(&used_allctr->mutex));
-		cpy_size = BLK_SZ(blk);
-		if (used_allctr->thread_safe && (!force_move
-						 || used_allctr != pref_allctr))
-		    erts_mtx_unlock(&used_allctr->mutex);
-		cpy_size -= ABLK_HDR_SZ + sizeof(UWord);
-	    }
-	    if (cpy_size > size)
-		cpy_size = size;
-	    sys_memcpy(res, p, cpy_size);
+	    sys_memcpy(res, p, MIN(size,old_user_size));
 
-	    if (!force_move || used_allctr != pref_allctr)
+	    if (used_allctr != pref_allctr) {
 		enqueue_dealloc_other_instance(type,
 					       used_allctr,
-					       ptr,
+					       p,
 					       (used_allctr->dd.ix
 						- pref_allctr->dd.ix));
+	    }
 	    else {
-		do_erts_alcu_free(type, used_allctr, ptr);
+		do_erts_alcu_free(type, used_allctr, p);
 		ASSERT(pref_allctr == used_allctr);
 		if (pref_allctr->thread_safe)
 		    erts_mtx_unlock(&pref_allctr->mutex);
@@ -4111,7 +4164,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 
     allctr->ramv			= init->ramv;
     allctr->main_carrier_size		= init->mmbcs;
-    allctr->sbc_threshold		= init->sbct;
+
 #if HAVE_ERTS_MSEG
     allctr->mseg_opt.abs_shrink_th	= init->asbcst;
     allctr->mseg_opt.rel_shrink_th	= init->rsbcst;
@@ -4120,20 +4173,29 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
     allctr->mbc_move_threshold		= init->rmbcmt;
 #if HAVE_ERTS_MSEG
     allctr->max_mseg_sbcs		= init->mmsbc;
+# if HAVE_SUPER_ALIGNED_MB_CARRIERS
+    allctr->max_mseg_mbcs		= ~(Uint)0;
+# else
     allctr->max_mseg_mbcs		= init->mmmbc;
+# endif
 #endif
 
     allctr->largest_mbc_size		= MAX(init->lmbcs, init->smbcs);
+#ifndef ARCH_64
+    if (allctr->largest_mbc_size > MBC_SZ_MAX_LIMIT) {
+	allctr->largest_mbc_size = MBC_SZ_MAX_LIMIT;
+    }
+#endif
     allctr->smallest_mbc_size		= init->smbcs;
     allctr->mbc_growth_stages		= MAX(1, init->mbcgs);
 
     if (allctr->min_block_size < ABLK_HDR_SZ)
 	goto error;
     allctr->min_block_size		= UNIT_CEILING(allctr->min_block_size
-						       + sizeof(UWord));
+						       + sizeof(FreeBlkFtr_t));
 #if ERTS_SMP
     if (init->tpref) {
-	Uint sz = sizeof(Block_t);
+	Uint sz = ABLK_HDR_SZ;
 	sz += ERTS_ALCU_DD_FIX_TYPE_OFFS*sizeof(UWord);
 	if (init->fix)
 	    sz += sizeof(UWord);
@@ -4143,6 +4205,23 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
     }
 #endif
 
+    allctr->sbc_threshold		= init->sbct;
+#ifndef ARCH_64
+    if (allctr->sbc_threshold > 0) {
+	Uint max_mbc_block_sz = UNIT_CEILING(allctr->sbc_threshold - 1 + ABLK_HDR_SZ); 
+	if (max_mbc_block_sz + UNIT_FLOOR(allctr->min_block_size - 1) > MBC_ABLK_SZ_MASK
+	    || max_mbc_block_sz < allctr->sbc_threshold) { /* wrap around */
+	    /* 
+	     * By limiting sbc_threshold to (hard limit - min_block_size)
+	     * we avoid having to split off free "residue blocks"
+	     * smaller than min_block_size.
+	     */
+	    max_mbc_block_sz = MBC_ABLK_SZ_MASK - UNIT_FLOOR(allctr->min_block_size - 1);
+	    allctr->sbc_threshold = max_mbc_block_sz - ABLK_HDR_SZ + 1;
+	}
+    }
+#endif
+
 
 
     allctr->sbmbc_threshold		= init->sbmbct;
@@ -4158,7 +4237,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 	allctr->sbmbc_size = init->sbmbcs;
 	min_size = allctr->sbmbc_threshold;
 	min_size += allctr->min_block_size;
-	min_size += allctr->mbc_header_size;
+	min_size += MBC_HEADER_SIZE(allctr);
 	if (allctr->sbmbc_size < min_size)
 	    allctr->sbmbc_size = min_size;
     }
@@ -4203,59 +4282,26 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
     if (!allctr->get_next_mbc_size)
 	allctr->get_next_mbc_size = get_next_mbc_size;
 
-    if (allctr->mbc_header_size < sizeof(Carrier_t))
-	goto error;
 #ifdef ERTS_SMP
     allctr->dd.use = 0;
     if (init->tpref) {
-	allctr->mbc_header_size = (UNIT_CEILING(allctr->mbc_header_size
-						+ FBLK_FTR_SZ
-						+ ABLK_HDR_SZ
-						+ sizeof(UWord))
-				   - ABLK_HDR_SZ
-				   - sizeof(UWord));
-	allctr->sbc_header_size = (UNIT_CEILING(sizeof(Carrier_t)
-						+ FBLK_FTR_SZ
-						+ ABLK_HDR_SZ
-						+ sizeof(UWord))
-				   - ABLK_HDR_SZ
-				   - sizeof(UWord));
-
 	allctr->dd.use = 1;
 	init_dd_queue(&allctr->dd.q);
 	allctr->dd.ix = init->ix;
     }
-    else
 #endif
-    {
-	allctr->mbc_header_size = (UNIT_CEILING(allctr->mbc_header_size
-						+ FBLK_FTR_SZ
-						+ ABLK_HDR_SZ)
-				   - ABLK_HDR_SZ);
-	allctr->sbc_header_size = (UNIT_CEILING(sizeof(Carrier_t)
-						+ FBLK_FTR_SZ
-						+ ABLK_HDR_SZ)
-				   - ABLK_HDR_SZ);
-    }
 
     if (allctr->main_carrier_size) {
 	Block_t *blk;
 
-#if HALFWORD_HEAP
-	blk = create_carrier(allctr,
-			     allctr->main_carrier_size,
-			     CFLG_MBC
-			     | CFLG_FORCE_SIZE
-			     | CFLG_FORCE_MSEG
-			     | CFLG_MAIN_CARRIER);
-#else
 	blk = create_carrier(allctr,
 			     allctr->main_carrier_size,
 			     CFLG_MBC
 			     | CFLG_FORCE_SIZE
+#if !HALFWORD_HEAP && !HAVE_SUPER_ALIGNED_MB_CARRIERS
 			     | CFLG_FORCE_SYS_ALLOC
-			     | CFLG_MAIN_CARRIER);
 #endif
+			     | CFLG_MAIN_CARRIER);
 	if (!blk)
 	    goto error;
 
@@ -4303,9 +4349,9 @@ erts_alcu_stop(Allctr_t *allctr)
     while (allctr->sbc_list.first)
 	destroy_carrier(allctr, SBC2BLK(allctr, allctr->sbc_list.first));
     while (allctr->mbc_list.first)
-	destroy_carrier(allctr, MBC2FBLK(allctr, allctr->mbc_list.first));
+	destroy_carrier(allctr, MBC_TO_FIRST_BLK(allctr, allctr->mbc_list.first));
     while (allctr->sbmbc_list.first)
-	destroy_sbmbc(allctr, MBC2FBLK(allctr, allctr->sbmbc_list.first));
+	destroy_sbmbc(allctr, MBC_TO_FIRST_BLK(allctr, allctr->sbmbc_list.first));
 
 #ifdef USE_THREADS
     if (allctr->thread_safe)
@@ -4319,17 +4365,9 @@ erts_alcu_stop(Allctr_t *allctr)
 void
 erts_alcu_init(AlcUInit_t *init)
 {
-
+    ASSERT(SBC_BLK_SZ_MASK == MBC_FBLK_SZ_MASK); /* see BLK_SZ */
 #if HAVE_ERTS_MSEG
-    mseg_unit_size = erts_mseg_unit_size();
-
-    if (mseg_unit_size % sizeof(Unit_t)) /* A little paranoid... */
-	erl_exit(-1,
-		 "Mseg unit size (%d) not evenly divideble by "
-		 "internal unit size of alloc_util (%d)\n",
-		 mseg_unit_size,
-		 sizeof(Unit_t));
-
+    ASSERT(erts_mseg_unit_size() == MSEG_UNIT_SZ);
     max_mseg_carriers = init->mmc;
     sys_alloc_carrier_size = MSEG_UNIT_CEILING(init->ycs);
 #else /* #if HAVE_ERTS_MSEG */
@@ -4372,11 +4410,10 @@ erts_alcu_test(unsigned long op, unsigned long a1, unsigned long a2)
     case 0x00b:	return (unsigned long) CARRIER_SZ((Carrier_t *) a1);
     case 0x00c:	return (unsigned long) SBC2BLK((Allctr_t *) a1,
 					       (Carrier_t *) a2);
-    case 0x00d:	return (unsigned long) BLK2SBC((Allctr_t *) a1,
-					       (Block_t *) a2);
-    case 0x00e:	return (unsigned long) MBC2FBLK((Allctr_t *) a1,
+    case 0x00d:	return (unsigned long) BLK_TO_SBC((Block_t *) a2);
+    case 0x00e:	return (unsigned long) MBC_TO_FIRST_BLK((Allctr_t *) a1,
 						(Carrier_t *) a2);
-    case 0x00f:	return (unsigned long) FBLK2MBC((Allctr_t *) a1,
+    case 0x00f:	return (unsigned long) FIRST_BLK_TO_MBC((Allctr_t *) a1,
 						(Block_t *) a2);
     case 0x010:	return (unsigned long) ((Allctr_t *) a1)->mbc_list.first;
     case 0x011:	return (unsigned long) ((Allctr_t *) a1)->mbc_list.last;
@@ -4388,7 +4425,7 @@ erts_alcu_test(unsigned long op, unsigned long a1, unsigned long a2)
     case 0x017:	return (unsigned long) ((Allctr_t *) a1)->min_block_size;
     case 0x018:	return (unsigned long) NXT_BLK((Block_t *) a1);
     case 0x019:	return (unsigned long) PREV_BLK((Block_t *) a1);
-    case 0x01a: return (unsigned long) IS_FIRST_BLK((Block_t *) a1);
+    case 0x01a: return (unsigned long) IS_MBC_FIRST_BLK((Allctr_t*)a1, (Block_t *) a2);
     case 0x01b: return (unsigned long) sizeof(Unit_t);
     default:	ASSERT(0); return ~((unsigned long) 0);
     }
@@ -4432,6 +4469,13 @@ erts_alcu_verify_unused_ts(Allctr_t *allctr)
 #endif
 }
 
+#ifdef DEBUG
+int is_sbc_blk(Block_t* blk)
+{
+    return IS_SBC_BLK(blk);
+}
+#endif
+
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
 
 static void
@@ -4441,34 +4485,37 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk)
     CarrierList_t *cl;
 
     if (IS_SBC_BLK(iblk)) {
-	Carrier_t *sbc = BLK2SBC(allctr, iblk);
+	Carrier_t *sbc = BLK_TO_SBC(iblk);
 
 	ASSERT(SBC2BLK(allctr, sbc) == iblk);
-	ASSERT(IS_ALLOCED_BLK(iblk));
-	ASSERT(IS_FIRST_BLK(iblk));
-	ASSERT(IS_LAST_BLK(iblk));
-	ASSERT(CARRIER_SZ(sbc) - allctr->sbc_header_size >= BLK_SZ(iblk));
+	ASSERT(CARRIER_SZ(sbc) - SBC_HEADER_SIZE >= SBC_BLK_SZ(iblk));
 #if HAVE_ERTS_MSEG
 	if (IS_MSEG_CARRIER(sbc)) {
-	    ASSERT(CARRIER_SZ(sbc) % mseg_unit_size == 0);
+	    ASSERT(CARRIER_SZ(sbc) % MSEG_UNIT_SZ == 0);
 	}
 #endif
 	crr = sbc;
 	cl = &allctr->sbc_list;
     }
     else {
-	Carrier_t *mbc = NULL;
 	Block_t *prev_blk = NULL;
 	Block_t *blk;
 	char *carrier_end;
 	Uint is_free_blk;
 	Uint tot_blk_sz;
 	Uint blk_sz;
+	int has_wrapped_around = 0;
 
 	blk = iblk;
 	tot_blk_sz = 0;
+	crr = BLK_TO_MBC(blk);
+	ASSERT(IS_MB_CARRIER(crr));
 
+	/* Step around the carrier one whole lap starting at 'iblk'
+	 */
 	while (1) {
+	    ASSERT(IS_MBC_BLK(blk));
+	    ASSERT(BLK_TO_MBC(blk) == crr);
 
 	    if (prev_blk) {
 		ASSERT(NXT_BLK(prev_blk) == blk);
@@ -4481,18 +4528,16 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk)
 		}
 	    }
 
-	    if (mbc) {
+	    if (has_wrapped_around) {
+		ASSERT(((Block_t *) crr) < blk);
 		if (blk == iblk)
 		    break;
-		ASSERT(((Block_t *) mbc) < blk && blk < iblk);
+		ASSERT(blk < iblk);
 	    }
 	    else
 		ASSERT(blk >= iblk);
 
-
-	    ASSERT(IS_MBC_BLK(blk));
-
-	    blk_sz = BLK_SZ(blk);
+	    blk_sz = MBC_BLK_SZ(blk);
 
 	    ASSERT(blk_sz % sizeof(Unit_t) == 0);
 	    ASSERT(blk_sz >= allctr->min_block_size);
@@ -4500,44 +4545,40 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk)
 	    tot_blk_sz += blk_sz;
 
 	    is_free_blk = (int) IS_FREE_BLK(blk);
-	    if(is_free_blk) {
-		if (IS_NOT_LAST_BLK(blk))
-		    ASSERT(*((UWord *) (((char *) blk)+blk_sz-sizeof(UWord)))
-			   == blk_sz);
-	    }
+	    ASSERT(!is_free_blk
+		   || IS_LAST_BLK(blk)
+		   || PREV_BLK_SZ(((char *) blk)+blk_sz) == blk_sz);
 
 	    if (allctr->check_block)
 		(*allctr->check_block)(allctr, blk, (int) is_free_blk);
 
 	    if (IS_LAST_BLK(blk)) {
 		carrier_end = ((char *) NXT_BLK(blk));
-		mbc = *((Carrier_t **) NXT_BLK(blk));
+		has_wrapped_around = 1;
 		prev_blk = NULL;
-		blk = MBC2FBLK(allctr, mbc);
-		ASSERT(IS_FIRST_BLK(blk));
+		blk = MBC_TO_FIRST_BLK(allctr, crr);
+		ASSERT(IS_MBC_FIRST_BLK(allctr,blk));
 	    }
 	    else {
 		prev_blk = blk;
 		blk = NXT_BLK(blk);
 	    }
 	}
-
-	ASSERT(IS_MB_CARRIER(mbc));
-	ASSERT((((char *) mbc)
-		+ allctr->mbc_header_size
+	
+	ASSERT((((char *) crr)
+		+ MBC_HEADER_SIZE(allctr)
 		+ tot_blk_sz) == carrier_end);
-	ASSERT(((char *) mbc) + CARRIER_SZ(mbc) - sizeof(Unit_t) <= carrier_end
-	       && carrier_end <= ((char *) mbc) + CARRIER_SZ(mbc));
+	ASSERT(((char *) crr) + CARRIER_SZ(crr) - sizeof(Unit_t) <= carrier_end
+	       && carrier_end <= ((char *) crr) + CARRIER_SZ(crr));
 
 	if (allctr->check_mbc)
-	    (*allctr->check_mbc)(allctr, mbc);
+	    (*allctr->check_mbc)(allctr, crr);
 
 #if HAVE_ERTS_MSEG
-	if (IS_MSEG_CARRIER(mbc)) {
-	    ASSERT(CARRIER_SZ(mbc) % mseg_unit_size == 0);
+	if (IS_MSEG_CARRIER(crr)) {
+	    ASSERT(CARRIER_SZ(crr) % MSEG_UNIT_SZ == 0);
 	}
 #endif
-	crr = mbc;
 	cl = &allctr->mbc_list;
     }
 
@@ -4559,4 +4600,5 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk)
 #endif
 }
 
-#endif
+#endif /* ERTS_ALLOC_UTIL_HARD_DEBUG */
+
diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h
index cedf4ccf85..e0754e7f69 100644
--- a/erts/emulator/beam/erl_alloc_util.h
+++ b/erts/emulator/beam/erl_alloc_util.h
@@ -216,16 +216,40 @@ erts_aint32_t erts_alcu_fix_alloc_shrink(Allctr_t *, erts_aint32_t);
 #define UNIT_FLOOR(X)	((X) & UNIT_MASK)
 #define UNIT_CEILING(X)	UNIT_FLOOR((X) + INV_UNIT_MASK)
 
+#define FLG_MASK		INV_UNIT_MASK
+#define SBC_BLK_SZ_MASK         UNIT_MASK
+#define MBC_FBLK_SZ_MASK        UNIT_MASK
+#define CARRIER_SZ_MASK         UNIT_MASK
 
-#define SZ_MASK			(~((UWord) 0) << 3)
-#define FLG_MASK		(~(SZ_MASK))
 
+#if HAVE_ERTS_MSEG
+#  ifdef ARCH_64 
+#    define MBC_ABLK_OFFSET_BITS   24
+#  elif HAVE_SUPER_ALIGNED_MB_CARRIERS
+#    define MBC_ABLK_OFFSET_BITS   9
+     /* Affects hard limits for sbct and lmbcs documented in erts_alloc.xml */
+#  endif
+#endif
+#ifndef MBC_ABLK_OFFSET_BITS
+#  define MBC_ABLK_OFFSET_BITS   0 /* no carrier offset in block header */
+#endif
+
+#if MBC_ABLK_OFFSET_BITS
+#  define MBC_ABLK_OFFSET_SHIFT  (sizeof(UWord)*8 - MBC_ABLK_OFFSET_BITS)
+#  define MBC_ABLK_OFFSET_MASK   (~((UWord)0) << MBC_ABLK_OFFSET_SHIFT)
+#  define MBC_ABLK_SZ_MASK	(~MBC_ABLK_OFFSET_MASK & ~FLG_MASK)
+#  define HAVE_ERTS_SBMBC 0
+#else
+#  define MBC_ABLK_SZ_MASK	(~FLG_MASK)
+#  define HAVE_ERTS_SBMBC 1
+#endif
 
-#define BLK_SZ(B) \
-  (*((Block_t *) (B)) & SZ_MASK)
+#define MBC_ABLK_SZ(B) (ASSERT_EXPR(!is_sbc_blk(B)), (B)->bhdr & MBC_ABLK_SZ_MASK)
+#define MBC_FBLK_SZ(B) (ASSERT_EXPR(!is_sbc_blk(B)), (B)->bhdr & MBC_FBLK_SZ_MASK)
+#define SBC_BLK_SZ(B) (ASSERT_EXPR(is_sbc_blk(B)), (B)->bhdr & SBC_BLK_SZ_MASK)
 
 #define CARRIER_SZ(C) \
-  ((C)->chdr & SZ_MASK)
+  ((C)->chdr & CARRIER_SZ_MASK)
 
 extern int erts_have_sbmbc_alloc;
 
@@ -236,6 +260,7 @@ struct Carrier_t_ {
     UWord chdr;
     Carrier_t *next;
     Carrier_t *prev;
+    Allctr_t  *allctr;
 };
 
 typedef struct {
@@ -243,8 +268,19 @@ typedef struct {
     Carrier_t *last;
 } CarrierList_t;
 
-typedef UWord Block_t;
-typedef UWord FreeBlkFtr_t;
+typedef struct {
+    UWord bhdr;
+#if !MBC_ABLK_OFFSET_BITS
+    Carrier_t *carrier;
+#else
+    union {
+	Carrier_t *carrier;     /* if free */
+	char       udata__[1];  /* if allocated */
+    }u;
+#endif
+} Block_t;
+
+typedef UWord FreeBlkFtr_t; /* Footer of a free block */
 
 typedef struct {
     UWord giga_no;
@@ -381,8 +417,6 @@ struct Allctr_t_ {
 #endif
 
     /* */
-    Uint		mbc_header_size;
-    Uint		sbc_header_size;
     Uint		min_mbc_size;
     Uint		min_mbc_first_free_size;
     Uint		min_block_size;
@@ -469,6 +503,9 @@ void	erts_alcu_verify_unused_ts(Allctr_t *allctr);
 
 unsigned long	erts_alcu_test(unsigned long, unsigned long, unsigned long);
 
+#ifdef DEBUG
+int is_sbc_blk(Block_t*);
+#endif
 
 
 #endif /* #if defined(GET_ERL_ALLOC_UTIL_IMPL)
diff --git a/erts/emulator/beam/erl_ao_firstfit_alloc.c b/erts/emulator/beam/erl_ao_firstfit_alloc.c
index 5bdb752d3a..86b4696d8f 100644
--- a/erts/emulator/beam/erl_ao_firstfit_alloc.c
+++ b/erts/emulator/beam/erl_ao_firstfit_alloc.c
@@ -91,6 +91,7 @@ struct AOFF_RBTree_t_ {
     AOFF_RBTree_t *right;
     Uint max_sz;  /* of all blocks in this sub-tree */
 };
+#define AOFF_BLK_SZ(B) MBC_FBLK_SZ(&(B)->hdr)
 
 #ifdef HARD_DEBUG
 static AOFF_RBTree_t * check_tree(AOFF_RBTree_t* root, Uint);
@@ -102,7 +103,7 @@ static AOFF_RBTree_t * check_tree(AOFF_RBTree_t* root, Uint);
  */
 static ERTS_INLINE Uint node_max_size(AOFF_RBTree_t *x)
 {
-    Uint sz = BLK_SZ(x);
+    Uint sz = AOFF_BLK_SZ(x);
     if (x->left && x->left->max_sz > sz) {
 	sz = x->left->max_sz;
     }
@@ -183,7 +184,6 @@ erts_aoffalc_start(AOFFAllctr_t *alc,
 
     sys_memcpy((void *) alc, (void *) &zero.allctr, sizeof(AOFFAllctr_t));
 
-    allctr->mbc_header_size		= sizeof(Carrier_t);
     allctr->min_mbc_size		= MIN_MBC_SZ;
     allctr->min_mbc_first_free_size	= MIN_MBC_FIRST_FREE_SZ;
     allctr->min_block_size		= sizeof(AOFF_RBTree_t);
@@ -587,7 +587,7 @@ aoff_link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
     AOFF_RBTree_t *blk = (AOFF_RBTree_t *) block;
     AOFF_RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC)
 			    ? &alc->sbmbc_root : &alc->mbc_root);
-    Uint blk_sz = BLK_SZ(blk);
+    Uint blk_sz = AOFF_BLK_SZ(blk);
 
 #ifdef HARD_DEBUG
     check_tree(*root, 0);
@@ -659,7 +659,7 @@ aoff_get_free_block(Allctr_t *allctr, Uint size,
 	if (x->left && x->left->max_sz >= size) {
 	    x = x->left;
 	}
-	else if (BLK_SZ(x) >= size) {
+	else if (AOFF_BLK_SZ(x) >= size) {
 	    blk = x;
 	    break;
 	}
@@ -910,12 +910,12 @@ check_tree(AOFF_RBTree_t* root, Uint size)
 	    ASSERT(x->right > x);
 	    ASSERT(x->right->max_sz <= x->max_sz);	    
 	}
-	ASSERT(x->max_sz >= BLK_SZ(x));
-	ASSERT(x->max_sz == BLK_SZ(x)
+	ASSERT(x->max_sz >= AOFF_BLK_SZ(x));
+	ASSERT(x->max_sz == AOFF_BLK_SZ(x)
 	       || x->max_sz == (x->left ? x->left->max_sz : 0)
 	       || x->max_sz == (x->right ? x->right->max_sz : 0));
 
-	if (size && BLK_SZ(x) >= size) {
+	if (size && AOFF_BLK_SZ(x) >= size) {
 	    if (!res || x < res) {
 		res = x;
 	    }
@@ -956,7 +956,7 @@ print_tree_aux(AOFF_RBTree_t *x, int indent)
 	}
 	fprintf(stderr, "%s: sz=%lu addr=0x%lx max_size=%lu\r\n",
 		IS_BLACK(x) ? "BLACK" : "RED",
-		BLK_SZ(x), (Uint)x, x->max_sz);
+		AOFF_BLK_SZ(x), (Uint)x, x->max_sz);
 	print_tree_aux(x->left,  indent + INDENT_STEP);
     }
 }
diff --git a/erts/emulator/beam/erl_bestfit_alloc.c b/erts/emulator/beam/erl_bestfit_alloc.c
index c50fdeb4e8..743cbd93c9 100644
--- a/erts/emulator/beam/erl_bestfit_alloc.c
+++ b/erts/emulator/beam/erl_bestfit_alloc.c
@@ -73,6 +73,8 @@
 #define SET_RED(N)		(((RBTree_t *) (N))->flags |= RED_FLG)
 #define SET_BLACK(N)		(((RBTree_t *) (N))->flags &= ~RED_FLG)
 
+#define BF_BLK_SZ(B)            MBC_FBLK_SZ(&(B)->hdr)
+
 #undef ASSERT
 #define ASSERT ASSERT_EXPR
 
@@ -177,7 +179,6 @@ erts_bfalc_start(BFAllctr_t *bfallctr,
     bfallctr->address_order		= bfinit->ao;
 
 
-    allctr->mbc_header_size		= sizeof(Carrier_t);
     allctr->min_mbc_size		= MIN_MBC_SZ;
     allctr->min_mbc_first_free_size	= MIN_MBC_FIRST_FREE_SZ;
     allctr->min_block_size		= (bfinit->ao
@@ -592,7 +593,7 @@ aobf_link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
 		       ? &bfallctr->sbmbc_root
 		       : &bfallctr->mbc_root);
     RBTree_t *blk = (RBTree_t *) block;
-    Uint blk_sz = BLK_SZ(blk);
+    Uint blk_sz = BF_BLK_SZ(blk);
 
     
 
@@ -610,7 +611,7 @@ aobf_link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
 	while (1) {
 	    Uint size;
 
-	    size = BLK_SZ(x);
+	    size = BF_BLK_SZ(x);
 
 	    if (blk_sz < size || (blk_sz == size && blk < x)) {
 		if (!x->left) {
@@ -668,7 +669,7 @@ aobf_get_free_block(Allctr_t *allctr, Uint size,
     ASSERT(!cand_blk || cand_size >= size);
 
     while (x) {
-	blk_sz = BLK_SZ(x);
+	blk_sz = BF_BLK_SZ(x);
 	if (blk_sz < size) {
 	    x = x->right;
 	}
@@ -686,7 +687,7 @@ aobf_get_free_block(Allctr_t *allctr, Uint size,
 #endif
 
     if (cand_blk) {
-	blk_sz = BLK_SZ(blk);
+	blk_sz = BF_BLK_SZ(blk);
 	if (cand_size < blk_sz)
 	    return NULL; /* cand_blk was better */
 	if (cand_size == blk_sz && ((void *) cand_blk) < ((void *) blk))
@@ -711,7 +712,7 @@ bf_link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
 		       ? &bfallctr->sbmbc_root
 		       : &bfallctr->mbc_root);
     RBTree_t *blk = (RBTree_t *) block;
-    Uint blk_sz = BLK_SZ(blk);
+    Uint blk_sz = BF_BLK_SZ(blk);
 
     SET_TREE_NODE(blk);
 
@@ -730,7 +731,7 @@ bf_link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
 	while (1) {
 	    Uint size;
 
-	    size = BLK_SZ(x);
+	    size = BF_BLK_SZ(x);
 
 	    if (blk_sz == size) {
 
@@ -796,7 +797,7 @@ bf_unlink_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
     else if (LIST_NEXT(x)) {
 	/* Replace tree node by next element in list... */
 
-	ASSERT(BLK_SZ(LIST_NEXT(x)) == BLK_SZ(x));
+	ASSERT(BF_BLK_SZ(LIST_NEXT(x)) == BF_BLK_SZ(x));
 	ASSERT(IS_TREE_NODE(x));
 	ASSERT(IS_LIST_ELEM(LIST_NEXT(x)));
 
@@ -834,7 +835,7 @@ bf_get_free_block(Allctr_t *allctr, Uint size,
     ASSERT(!cand_blk || cand_size >= size);
 
     while (x) {
-	blk_sz = BLK_SZ(x);
+	blk_sz = BF_BLK_SZ(x);
 	if (blk_sz < size) {
 	    x = x->right;
 	}
@@ -855,11 +856,11 @@ bf_get_free_block(Allctr_t *allctr, Uint size,
 #ifdef HARD_DEBUG
     {
 	RBTree_t *ct_blk = check_tree(root, 0, size);
-	ASSERT(BLK_SZ(ct_blk) == BLK_SZ(blk));
+	ASSERT(BF_BLK_SZ(ct_blk) == BF_BLK_SZ(blk));
     }
 #endif
 
-    if (cand_blk && cand_size <= BLK_SZ(blk))
+    if (cand_blk && cand_size <= BF_BLK_SZ(blk))
 	return NULL; /* cand_blk was better */
 
     /* Use next block if it exist in order to avoid replacing
@@ -1093,36 +1094,36 @@ check_tree(RBTree_t *root, int ao, Uint size)
 	if (x->left) {
 	    ASSERT(x->left->parent == x);
 	    if (ao) {
-		ASSERT(BLK_SZ(x->left) < BLK_SZ(x)
-		       || (BLK_SZ(x->left) == BLK_SZ(x) && x->left < x));
+		ASSERT(BF_BLK_SZ(x->left) < BF_BLK_SZ(x)
+		       || (BF_BLK_SZ(x->left) == BF_BLK_SZ(x) && x->left < x));
 	    }
 	    else {
 		ASSERT(IS_TREE_NODE(x->left));
-		ASSERT(BLK_SZ(x->left) < BLK_SZ(x));
+		ASSERT(BF_BLK_SZ(x->left) < BF_BLK_SZ(x));
 	    }
 	}
 
 	if (x->right) {
 	    ASSERT(x->right->parent == x);
 	    if (ao) {
-		ASSERT(BLK_SZ(x->right) > BLK_SZ(x)
-		       || (BLK_SZ(x->right) == BLK_SZ(x) && x->right > x));
+		ASSERT(BF_BLK_SZ(x->right) > BF_BLK_SZ(x)
+		       || (BF_BLK_SZ(x->right) == BF_BLK_SZ(x) && x->right > x));
 	    }
 	    else {
 		ASSERT(IS_TREE_NODE(x->right));
-		ASSERT(BLK_SZ(x->right) > BLK_SZ(x));
+		ASSERT(BF_BLK_SZ(x->right) > BF_BLK_SZ(x));
 	    }
 	}
 
-	if (size && BLK_SZ(x) >= size) {
+	if (size && BF_BLK_SZ(x) >= size) {
 	    if (ao) {
 		if (!res
-		    || BLK_SZ(x) < BLK_SZ(res)
-		    || (BLK_SZ(x) == BLK_SZ(res) && x < res))
+		    || BF_BLK_SZ(x) < BF_BLK_SZ(res)
+		    || (BF_BLK_SZ(x) == BF_BLK_SZ(res) && x < res))
 		    res = x;
 	    }
 	    else {
-		if (!res || BLK_SZ(x) < BLK_SZ(res))
+		if (!res || BF_BLK_SZ(x) < BF_BLK_SZ(res))
 		    res = x;
 	    }
 	}
@@ -1168,7 +1169,7 @@ print_tree_aux(RBTree_t *x, int indent)
 	}
 	fprintf(stderr, "%s: sz=%lu addr=0x%lx\r\n",
 		IS_BLACK(x) ? "BLACK" : "RED",
-		BLK_SZ(x),
+		BF_BLK_SZ(x),
 		(Uint) x);
 	print_tree_aux(x->left,  indent + INDENT_STEP);
     }
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index d377ba8f31..a33085315a 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -129,7 +129,7 @@ static void disallow_heap_frag_ref(Process* p, Eterm* n_htop, Eterm* objv, int n
 #if defined(ARCH_64) && !HALFWORD_HEAP
 # define MAX_HEAP_SIZES 154
 #else
-# define MAX_HEAP_SIZES 55
+# define MAX_HEAP_SIZES 59
 #endif
 
 static Sint heap_sizes[MAX_HEAP_SIZES];	/* Suitable heap sizes. */
@@ -144,6 +144,7 @@ void
 erts_init_gc(void)
 {
     int i = 0;
+    Sint max_heap_size = 0;
 
     ASSERT(offsetof(ProcBin,thing_word) == offsetof(struct erl_off_heap_header,thing_word));
     ASSERT(offsetof(ProcBin,thing_word) == offsetof(ErlFunThing,thing_word));
@@ -168,16 +169,30 @@ erts_init_gc(void)
      * we really don't want that growth when the heaps are that big.
      */
 	    
-    heap_sizes[0] = 34;
-    heap_sizes[1] = 55;
-    for (i = 2; i < 23; i++) {
-	heap_sizes[i] = heap_sizes[i-1] + heap_sizes[i-2];
+    /* Growth stage 1 - Fibonacci + 1*/
+    /* 12,38 will hit size 233, the old default */
+
+    heap_sizes[0] = 12;
+    heap_sizes[1] = 38;
+
+    for(i = 2; i < 23; i++) {
+        /* one extra word for block header */
+        heap_sizes[i] = heap_sizes[i-1] + heap_sizes[i-2] + 1;
     }
 
+
+    /* for 32 bit we want max_heap_size to be MAX(32bit) / 4 [words] (and halfword)
+     * for 64 bit we want max_heap_size to be MAX(52bit) / 8 [words]
+     */
+
+    max_heap_size = sizeof(Eterm) < 8 ? (Sint)((~(Uint)0)/(sizeof(Eterm))) : 
+					(Sint)(((Uint64)1 << 53)/sizeof(Eterm));
+
+    /* Growth stage 2 - 20% growth */
     /* At 1.3 mega words heap, we start to slow down. */
     for (i = 23; i < ALENGTH(heap_sizes); i++) {
-	heap_sizes[i] = 5*(heap_sizes[i-1]/4);
-	if (heap_sizes[i] < 0) {
+	heap_sizes[i] = heap_sizes[i-1] + heap_sizes[i-1]/5;
+	if ((heap_sizes[i] < 0) || heap_sizes[i] > max_heap_size) {
 	    /* Size turned negative. Discard this last size. */
 	    i--;
 	    break;
@@ -868,14 +883,12 @@ minor_collection(Process* p, int need, Eterm* objv, int nobj, Uint *recl)
 		}
 	    }
 
-            if (wanted < MIN_HEAP_SIZE(p)) {
-                wanted = MIN_HEAP_SIZE(p);
-            } else {
-                wanted = next_heap_size(p, wanted, 0);
-            }
+	    wanted = wanted < MIN_HEAP_SIZE(p) ? MIN_HEAP_SIZE(p)
+					       : next_heap_size(p, wanted, 0);
             if (wanted < HEAP_SIZE(p)) {
                 shrink_new_heap(p, wanted, objv, nobj);
             }
+
             ASSERT(HEAP_SIZE(p) == next_heap_size(p, HEAP_SIZE(p), 0));
 	    return 1;		/* We are done. */
         }
@@ -1434,11 +1447,10 @@ adjust_after_fullsweep(Process *p, Uint size_before, int need, Eterm *objv, int
            I think this is better as fullsweep is used mainly on
            small memory systems, but I could be wrong... */
         wanted = 2 * need_after;
-        if (wanted < p->min_heap_size) {
-            sz = p->min_heap_size;
-        } else {
-            sz = next_heap_size(p, wanted, 0);
-        }
+	
+	sz = wanted < p->min_heap_size ? p->min_heap_size
+				       : next_heap_size(p, wanted, 0);
+
         if (sz < HEAP_SIZE(p)) {
             shrink_new_heap(p, sz, objv, nobj);
         }
diff --git a/erts/emulator/beam/erl_goodfit_alloc.c b/erts/emulator/beam/erl_goodfit_alloc.c
index e7d4ac2b67..f98d377fd6 100644
--- a/erts/emulator/beam/erl_goodfit_alloc.c
+++ b/erts/emulator/beam/erl_goodfit_alloc.c
@@ -205,7 +205,6 @@ erts_gfalc_start(GFAllctr_t *gfallctr,
 
     init->sbmbct = 0; /* Small mbc not yet supported by goodfit */
 
-    allctr->mbc_header_size		= sizeof(Carrier_t);
     allctr->min_mbc_size		= MIN_MBC_SZ;
     allctr->min_mbc_first_free_size	= MIN_MBC_FIRST_FREE_SZ;
     allctr->min_block_size		= sizeof(GFFreeBlock_t);
@@ -363,7 +362,7 @@ search_bucket(Allctr_t *allctr, int ix, Uint size)
 	 blk && i < max_blk_search;
 	 blk = blk->next, i++) {
 
-	blk_sz = BLK_SZ(blk);
+	blk_sz = MBC_FBLK_SZ(&blk->block_head);
 	blk_on_lambc = (((char *) blk) < gfallctr->last_aux_mbc_end
 			&& gfallctr->last_aux_mbc_start <= ((char *) blk));
 
@@ -402,7 +401,7 @@ get_free_block(Allctr_t *allctr, Uint size,
     if (min_bi == unsafe_bi) {
 	blk = search_bucket(allctr, min_bi, size);
 	if (blk) {
-	    if (cand_blk && cand_size <= BLK_SZ(blk))
+	    if (cand_blk && cand_size <= MBC_FBLK_SZ(blk))
 		return NULL; /* cand_blk was better */
 	    unlink_free_block(allctr, blk, flags);
 	    return blk;
@@ -422,7 +421,7 @@ get_free_block(Allctr_t *allctr, Uint size,
     /* We are guaranteed to find a block that fits in this bucket */
     blk = search_bucket(allctr, min_bi, size);
     ASSERT(blk);
-    if (cand_blk && cand_size <= BLK_SZ(blk))
+    if (cand_blk && cand_size <= MBC_FBLK_SZ(blk))
 	return NULL; /* cand_blk was better */
     unlink_free_block(allctr, blk, flags);
     return blk;
@@ -435,7 +434,7 @@ link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
 {
     GFAllctr_t *gfallctr = (GFAllctr_t *) allctr;
     GFFreeBlock_t *blk = (GFFreeBlock_t *) block;
-    Uint sz = BLK_SZ(blk);
+    Uint sz = MBC_FBLK_SZ(&blk->block_head);
     int i = BKT_IX(gfallctr, sz);
 
     ASSERT(sz >= MIN_BLK_SZ);
@@ -456,7 +455,7 @@ unlink_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags)
 {
     GFAllctr_t *gfallctr = (GFAllctr_t *) allctr;
     GFFreeBlock_t *blk = (GFFreeBlock_t *) block;
-    Uint sz = BLK_SZ(blk);
+    Uint sz = MBC_FBLK_SZ(&blk->block_head);
     int i = BKT_IX(gfallctr, sz);
 
     if (!blk->prev) {
@@ -618,7 +617,7 @@ check_block(Allctr_t *allctr, Block_t * blk, int free_block)
     GFFreeBlock_t *fblk;
 
     if(free_block) {
-	Uint blk_sz = BLK_SZ(blk);
+	Uint blk_sz = is_sbc_blk(blk) ? SBC_BLK_SZ(blk) : MBC_BLK_SZ(blk);
 	bi = BKT_IX(gfallctr, blk_sz);
 
 	ASSERT(gfallctr->bucket_mask.main & (((UWord) 1) << IX2SMIX(bi)));
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index 41f33efa79..848877d43e 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -44,9 +44,9 @@
 #undef DEBUG_PRINTOUTS
 #endif
 
-extern Eterm beam_return_to_trace[1];   /* OpCode(i_return_to_trace) */
-extern Eterm beam_return_trace[1];      /* OpCode(i_return_trace) */
-extern Eterm beam_return_time_trace[1]; /* OpCode(i_return_time_trace) */
+extern BeamInstr beam_return_to_trace[1];   /* OpCode(i_return_to_trace) */
+extern BeamInstr beam_return_trace[1];      /* OpCode(i_return_trace) */
+extern BeamInstr beam_return_time_trace[1]; /* OpCode(i_return_time_trace) */
 
 /* Pseudo export entries. Never filled in with data, only used to
    yield unique pointers of the correct type. */
diff --git a/erts/emulator/sys/common/erl_mseg.c b/erts/emulator/sys/common/erl_mseg.c
index db2854fa40..94f9f76a20 100644
--- a/erts/emulator/sys/common/erl_mseg.c
+++ b/erts/emulator/sys/common/erl_mseg.c
@@ -57,26 +57,48 @@
      /* Implement some other way to get the real page size if needed! */
 #endif
 
-#define MAX_CACHE_SIZE 30
-
 #undef MIN
 #define MIN(X, Y) ((X) < (Y) ? (X) : (Y))
 #undef MAX
 #define MAX(X, Y) ((X) > (Y) ? (X) : (Y))
 
-#undef  PAGE_MASK
-#define INV_PAGE_MASK	((Uint) (page_size - 1))
-#define PAGE_MASK	(~INV_PAGE_MASK)
-#define PAGE_FLOOR(X)	((X) & PAGE_MASK)
-#define PAGE_CEILING(X)	PAGE_FLOOR((X) + INV_PAGE_MASK)
-#define PAGES(X)	((X) >> page_shift)
+#define INV_ALIGNED_MASK    ((UWord) ((MSEG_ALIGNED_SIZE) - 1))
+#define ALIGNED_MASK        (~INV_ALIGNED_MASK)
+#define ALIGNED_FLOOR(X)    (((UWord)(X)) & ALIGNED_MASK)
+#define ALIGNED_CEILING(X)  ALIGNED_FLOOR((X) + INV_ALIGNED_MASK)
+#define MAP_IS_ALIGNED(X)   (((UWord)(X) & (MSEG_ALIGNED_SIZE - 1)) == 0)
+
+#define IS_2POW(X)          ((X) && !((X) & ((X) - 1)))
+static ERTS_INLINE Uint ceil_2pow(Uint x) {
+    int i = 1 << (4 + (sizeof(Uint) != 4 ? 1 : 0));
+    x--;
+    do { x |= x >> i; } while(i >>= 1);
+    return x + 1;
+}
+static const int debruijn[32] = {
+     0,  1, 28,  2, 29, 14, 24,  3, 30, 22, 20, 15, 25, 17,  4,  8, 
+    31, 27, 13, 23, 21, 19, 16,  7, 26, 12, 18,  6, 11,  5, 10,  9
+};
+
+#define LOG2(X) (debruijn[((Uint32)(((X) & -(X)) * 0x077CB531U)) >> 27])
+
+#define CACHE_AREAS      (32 - MSEG_ALIGN_BITS)
+
+#define SIZE_TO_CACHE_AREA_IDX(S)   (LOG2((S)) - MSEG_ALIGN_BITS)
+#define MAX_CACHE_SIZE   (30)
+
+#define MSEG_FLG_IS_2POW(X)    ((X) & ERTS_MSEG_FLG_2POW)
+
+#ifdef DEBUG
+#define DBG(F,...) fprintf(stderr, (F), __VA_ARGS__ )
+#else
+#define DBG(F,...) do{}while(0)
+#endif
 
 static int atoms_initialized;
 
 typedef struct mem_kind_t MemKind;
 
-static void mseg_clear_cache(MemKind*);
-
 #if HALFWORD_HEAP
 static int initialize_pmmap(void);
 static void *pmmap(size_t size);
@@ -116,15 +138,6 @@ static int mmap_fd;
 #error "Not supported"
 #endif /* #if HAVE_MMAP */
 
-#if defined(ERTS_MSEG_FAKE_SEGMENTS) && HALFWORD_HEAP
-#  warning "ERTS_MSEG_FAKE_SEGMENTS will only be used for high memory segments" 
-#endif 
-
-#if defined(ERTS_MSEG_FAKE_SEGMENTS)
-#undef CAN_PARTLY_DESTROY
-#define CAN_PARTLY_DESTROY 0
-#endif
-
 const ErtsMsegOpt_t erts_mseg_default_opt = {
     1,			/* Use cache		     */
     1,			/* Preserv data		     */
@@ -137,26 +150,17 @@ const ErtsMsegOpt_t erts_mseg_default_opt = {
 };
 
 
-typedef struct cache_desc_t_ {
-    void *seg;
-    Uint size;
-    struct cache_desc_t_ *next;
-    struct cache_desc_t_ *prev;
-} cache_desc_t;
-
 typedef struct {
     Uint32 giga_no;
     Uint32 no;
 } CallCounter;
 
-static Uint page_size;
-static Uint page_shift;
-
 typedef struct {
     CallCounter alloc;
     CallCounter dealloc;
     CallCounter realloc;
     CallCounter create;
+    CallCounter create_resize;
     CallCounter destroy;
 #if HAVE_MSEG_RECREATE
     CallCounter recreate;
@@ -165,17 +169,25 @@ typedef struct {
     CallCounter check_cache;
 } ErtsMsegCalls;
 
+typedef struct cache_t_ cache_t;
+
+struct cache_t_ {
+    Uint size;
+    void *seg;
+    cache_t *next;
+};
+
+
 typedef struct ErtsMsegAllctr_t_ ErtsMsegAllctr_t;
 
 struct mem_kind_t {
-    cache_desc_t cache_descs[MAX_CACHE_SIZE];
-    cache_desc_t *free_cache_descs;
-    cache_desc_t *cache;
-    cache_desc_t *cache_end;
-
-    Uint cache_size;
-    Uint min_cached_seg_size;
-    Uint max_cached_seg_size;
+
+    cache_t cache[MAX_CACHE_SIZE];
+    cache_t *cache_unpowered;
+    cache_t *cache_area[CACHE_AREAS];
+    cache_t *cache_free;
+
+    Sint cache_size;
     Uint cache_hits;
 
     struct {
@@ -320,8 +332,7 @@ static erts_mtx_t init_atoms_mutex; /* Also needed when !USE_THREADS */
 
 
 static ERTS_INLINE void
-schedule_cache_check(ErtsMsegAllctr_t *ma)
-{
+schedule_cache_check(ErtsMsegAllctr_t *ma) {
 
     if (!ma->is_cache_check_scheduled && ma->is_init_done) {
 	erts_set_aux_work_timeout(ma->ix,
@@ -331,12 +342,45 @@ schedule_cache_check(ErtsMsegAllctr_t *ma)
     }
 }
 
+/* remove ErtsMsegAllctr_t from arguments?
+ * only used for statistics
+ */
+static ERTS_INLINE void *
+mmap_align(ErtsMsegAllctr_t *ma, void *addr, size_t length, int prot, int flags, int fd, off_t offset) {
+
+    void *p, *q;
+    UWord d;
+
+    p = mmap(addr, length, prot, flags, fd, offset);
+
+    if (MAP_IS_ALIGNED(p) || p == MAP_FAILED)
+	return p;
+
+    if (ma)
+	INC_CC(ma, create_resize);
+
+    munmap(p, length);
+
+    if ((p = mmap(addr, length + MSEG_ALIGNED_SIZE, prot, flags, fd, offset)) == MAP_FAILED)
+	return MAP_FAILED;
+
+    q = (void *)ALIGNED_CEILING(p);
+    d = q - p;
+
+    if (d > 0)
+	munmap(p, d);
+
+    if (MSEG_ALIGNED_SIZE - d > 0)
+	munmap((void *) (q + length), MSEG_ALIGNED_SIZE - d);
+
+    return q;
+}
+
 static ERTS_INLINE void *
 mseg_create(ErtsMsegAllctr_t *ma, MemKind* mk, Uint size)
 {
     void *seg;
-
-    ASSERT(size % page_size == 0);
+    ASSERT(size % MSEG_ALIGNED_SIZE == 0);
 
 #if HALFWORD_HEAP
     if (mk == &ma->low_mem) {
@@ -345,18 +389,17 @@ mseg_create(ErtsMsegAllctr_t *ma, MemKind* mk, Uint size)
 	    erts_fprintf(stderr,"Pointer mask failure (0x%08lx)\n",(unsigned long) seg);
 	    return NULL;
 	}
-    }
-    else
+    } else
 #endif
     {
-#if defined(ERTS_MSEG_FAKE_SEGMENTS)
-	seg = erts_sys_alloc(ERTS_ALC_N_INVALID, NULL, size);
-#elif HAVE_MMAP
+#if HAVE_MMAP
 	{
-	    seg = (void *) mmap((void *) 0, (size_t) size,
+	    seg = (void *) mmap_align(ma, (void *) 0, (size_t) size,
 				MMAP_PROT, MMAP_FLAGS, MMAP_FD, 0);
 	    if (seg == (void *) MAP_FAILED)
 		seg = NULL;
+
+	    ASSERT(MAP_IS_ALIGNED(seg) || !seg);
 	}
 #else
 # error "Missing mseg_create() implementation"
@@ -369,38 +412,24 @@ mseg_create(ErtsMsegAllctr_t *ma, MemKind* mk, Uint size)
 }
 
 static ERTS_INLINE void
-mseg_destroy(ErtsMsegAllctr_t *ma, MemKind* mk, void *seg, Uint size)
-{
-#ifdef DEBUG
-    int res;
-#endif
+mseg_destroy(ErtsMsegAllctr_t *ma, MemKind* mk, void *seg, Uint size) {
+    ERTS_DECLARE_DUMMY(int res);
 
 #if HALFWORD_HEAP
     if (mk == &ma->low_mem) {
-#ifdef DEBUG
-	res = 
-#endif
-	    pmunmap((void *) seg, size);
+	res = pmunmap((void *) seg, size);
     }
     else
 #endif
     {
-#ifdef ERTS_MSEG_FAKE_SEGMENTS
-	erts_sys_free(ERTS_ALC_N_INVALID, NULL, seg);
-#ifdef DEBUG
-	res = 0;
-#endif
-#elif HAVE_MMAP
-#ifdef DEBUG
-	res = 
-#endif
-	    munmap((void *) seg, size);
+#ifdef HAVE_MMAP
+	res = munmap((void *) seg, size);
 #else
 # error "Missing mseg_destroy() implementation"
 #endif
     }
 
-    ASSERT(size % page_size == 0);
+    ASSERT(size % MSEG_ALIGNED_SIZE == 0);
     ASSERT(res == 0);
 
     INC_CC(ma, destroy);
@@ -408,14 +437,36 @@ mseg_destroy(ErtsMsegAllctr_t *ma, MemKind* mk, void *seg, Uint size)
 }
 
 #if HAVE_MSEG_RECREATE
+#if defined(__NetBsd__)
+#define MREMAP_FLAGS  (0)
+#else
+#define MREMAP_FLAGS  (MREMAP_MAYMOVE)
+#endif
+
+
+/* mseg_recreate
+ * May return *unaligned* segments as in address not aligned to MSEG_ALIGNMENT
+ * it is still page aligned
+ *
+ * This is fine for single block carriers as long as we don't cache misaligned
+ * segments (since multiblock carriers may use them)
+ *
+ * For multiblock carriers we *need* MSEG_ALIGNMENT but mbc's will never be
+ * reallocated.
+ *
+ * This should probably be fixed the following way:
+ * 1) Use an option to segment allocation - NEED_ALIGNMENT
+ * 2) Add mremap_align which takes care of aligning a new a mremaped area
+ * 3) Fix the cache to handle of aligned and unaligned segments
+ */
 
 static ERTS_INLINE void *
 mseg_recreate(ErtsMsegAllctr_t *ma, MemKind* mk, void *old_seg, Uint old_size, Uint new_size)
 {
     void *new_seg;
 
-    ASSERT(old_size % page_size == 0);
-    ASSERT(new_size % page_size == 0);
+    ASSERT(old_size % MSEG_ALIGNED_SIZE == 0);
+    ASSERT(new_size % MSEG_ALIGNED_SIZE == 0);
 
 #if HALFWORD_HEAP
     if (mk == &ma->low_mem) {
@@ -426,22 +477,12 @@ mseg_recreate(ErtsMsegAllctr_t *ma, MemKind* mk, void *old_seg, Uint old_size, U
     else
 #endif
     {
-#if defined(ERTS_MSEG_FAKE_SEGMENTS)
-	new_seg = erts_sys_realloc(ERTS_ALC_N_INVALID, NULL, old_seg, new_size);
-#elif HAVE_MREMAP
-
-    #if defined(__NetBSD__)
-	new_seg = (void *) mremap((void *) old_seg,
-				  (size_t) old_size,
-				  NULL,
-				  (size_t) new_size,
-				  0);
-    #else
-	new_seg = (void *) mremap((void *) old_seg,
-				  (size_t) old_size,
-				  (size_t) new_size,
-				  MREMAP_MAYMOVE);
-    #endif
+#if HAVE_MREMAP
+#if defined(__NetBSD__)
+	new_seg = mremap(old_seg, (size_t)old_size, NULL, new_size, MREMAP_FLAGS);
+#else
+	new_seg = mremap(old_seg, (size_t)old_size, (size_t)new_size, MREMAP_FLAGS);
+#endif
 	if (new_seg == (void *) MAP_FAILED)
 	    new_seg = NULL;
 #else
@@ -475,151 +516,265 @@ do {									\
 #define ERTS_DBG_MK_CHK_THR_ACCESS(MK)
 #endif
 
-static ERTS_INLINE cache_desc_t * 
-alloc_cd(MemKind* mk)
-{    
-    cache_desc_t *cd = mk->free_cache_descs;
+/* NEW CACHE interface */
+
+static ERTS_INLINE cache_t *mseg_cache_alloc_descriptor(MemKind *mk) {
+    cache_t *c = mk->cache_free;
+
     ERTS_DBG_MK_CHK_THR_ACCESS(mk);
-    if (cd)
-	mk->free_cache_descs = cd->next;
-    return cd;
+    if (c)
+	mk->cache_free = c->next;
+
+    return c;
 }
 
-static ERTS_INLINE void
-free_cd(MemKind* mk, cache_desc_t *cd)
-{
+static ERTS_INLINE void mseg_cache_free_descriptor(MemKind *mk, cache_t *c) {
     ERTS_DBG_MK_CHK_THR_ACCESS(mk);
-    cd->next = mk->free_cache_descs;
-    mk->free_cache_descs = cd;
+    ASSERT(c);
+    
+    c->seg  = NULL;
+    c->size = 0;
+    c->next = mk->cache_free;
+    mk->cache_free = c;
 }
 
+static ERTS_INLINE int cache_bless_segment(MemKind *mk, void *seg, Uint size) {
 
-static ERTS_INLINE void
-link_cd(MemKind* mk, cache_desc_t *cd)
-{
+    cache_t *c;
     ERTS_DBG_MK_CHK_THR_ACCESS(mk);
-    if (mk->cache)
-	mk->cache->prev = cd;
-    cd->next = mk->cache;
-    cd->prev = NULL;
-    mk->cache = cd;
-
-    if (!mk->cache_end) {
-	ASSERT(!cd->next);
-	mk->cache_end = cd;
+
+    if (mk->cache_free && MAP_IS_ALIGNED(seg)) {
+	if (IS_2POW(size)) {
+	    int ix = SIZE_TO_CACHE_AREA_IDX(size);
+
+	    ASSERT(ix < CACHE_AREAS);
+	    ASSERT((1 << (ix + MSEG_ALIGN_BITS)) == size);
+
+	    /* unlink from free cache list */
+	    c = mseg_cache_alloc_descriptor(mk);
+
+	    /* link to cache area */
+	    c->seg  = seg;
+	    c->size = size;
+	    c->next = mk->cache_area[ix];
+
+	    mk->cache_area[ix] = c;
+	    mk->cache_size++;
+
+	    ASSERT(mk->cache_size <= mk->ma->max_cache_size);
+
+	    return 1;
+	} else {
+	    /* unlink from free cache list */
+	    c = mseg_cache_alloc_descriptor(mk);
+
+	    /* link to cache area */
+	    c->seg  = seg;
+	    c->size = size;
+	    c->next = mk->cache_unpowered;
+
+	    mk->cache_unpowered = c;
+	    mk->cache_size++;
+
+	    ASSERT(mk->cache_size <= mk->ma->max_cache_size);
+
+	    return 1;
+	}
     }
 
-    mk->cache_size++;
+    return 0;
 }
 
-#if CAN_PARTLY_DESTROY
-static ERTS_INLINE void
-end_link_cd(MemKind* mk, cache_desc_t *cd)
-{
+static ERTS_INLINE void *cache_get_segment(MemKind *mk, Uint *size_p) {
+
+    Uint size = *size_p;
+
     ERTS_DBG_MK_CHK_THR_ACCESS(mk);
-    if (mk->cache_end)
-	mk->cache_end->next = cd;
-    cd->next = NULL;
-    cd->prev = mk->cache_end;
-    mk->cache_end = cd;
-
-    if (!mk->cache) {
-	ASSERT(!cd->prev);
-	mk->cache = cd;
-    }
+    if (IS_2POW(size)) {
+
+	int i, ix = SIZE_TO_CACHE_AREA_IDX(size);
+	void *seg;
+	cache_t *c;
+	Uint csize;
+
+	for( i = ix; i < CACHE_AREAS; i++) {
+
+	    if ((c = mk->cache_area[i]) == NULL)
+		continue;
+
+	    ASSERT(IS_2POW(c->size));
+
+	    /* unlink from cache area */
+	    csize   = c->size;
+	    seg     = c->seg;
+	    mk->cache_area[i] = c->next;
+	    c->next = NULL;
+	    mk->cache_size--;
+	    mk->cache_hits++;
+
+	    /* link to free cache list */
+	    mseg_cache_free_descriptor(mk, c);
+
+	    ASSERT(!(mk->cache_size < 0));
+
+	    /* divvy up the cache - if needed */
+	    while( i > ix) {
+		csize = csize >> 1;
+		/* try to cache half of it */
+		if (!cache_bless_segment(mk, (char *)seg + csize, csize)) {
+		    /* wouldn't cache .. destroy it instead */
+		    mseg_destroy(mk->ma, mk, (char *)seg + csize, csize);
+		}
+		i--;
+	    }
+	    ASSERT(csize == size);
+	    return seg;
+	}
+    } 
+    else if (mk->cache_unpowered) {
+	void *seg;
+	cache_t *c, *pc;
+	Uint csize;
+	Uint bad_max_abs = mk->ma->abs_max_cache_bad_fit;
+	Uint bad_max_rel = mk->ma->rel_max_cache_bad_fit;
+
+	c = mk->cache_unpowered;
+	pc = c;
+
+	while (c) {
+	    csize = c->size;
+		if (csize >= size &&
+			((csize - size)*100 < bad_max_rel*size) &&
+			(csize - size) < bad_max_abs ) {
+
+		/* unlink from cache area */
+		seg = c->seg;
+
+		if (pc == c) {
+		    mk->cache_unpowered = c->next;
+		} else {
+		    pc->next = c->next;
+		}
+
+		c->next = NULL;
+		mk->cache_size--;
+		mk->cache_hits++;
+
+		/* link to free cache list */
+		mseg_cache_free_descriptor(mk, c);
+		*size_p = csize;
+
+		return seg;
+	    }
 
-    mk->cache_size++;
+	    pc = c;
+	    c  = c->next;
+	}
+    }
+    return NULL;
 }
-#endif
 
-static ERTS_INLINE void
-unlink_cd(MemKind* mk, cache_desc_t *cd)
-{
-    ERTS_DBG_MK_CHK_THR_ACCESS(mk);
-    if (cd->next)
-	cd->next->prev = cd->prev;
-    else
-	mk->cache_end = cd->prev;
-    
-    if (cd->prev)
-	cd->prev->next = cd->next;
-    else
-	mk->cache = cd->next;
-    ASSERT(mk->cache_size > 0);
+/* *_mseg_check_*_cache
+ * Slowly remove segments cached in the allocator by
+ * using callbacks from aux-work in the scheduler.
+ */
+
+static ERTS_INLINE Uint mseg_drop_one_memkind_cache_size(MemKind *mk, cache_t **head) {
+    cache_t *c = NULL;
+
+    c = *head;
+
+    ASSERT( c != NULL );
+
+    *head = c->next;
+
+    if (erts_mtrace_enabled)
+	erts_mtrace_crr_free(SEGTYPE, SEGTYPE, c->seg);
+
+    mseg_destroy(mk->ma, mk, c->seg, c->size);
+    mseg_cache_free_descriptor(mk, c);
+
+    mk->segments.current.watermark--;
     mk->cache_size--;
-}
 
-static ERTS_INLINE void
-check_cache_limits(MemKind* mk)
-{
-    cache_desc_t *cd;
-    ERTS_DBG_MK_CHK_THR_ACCESS(mk);
-    mk->max_cached_seg_size = 0;
-    mk->min_cached_seg_size = ~((Uint) 0);
-    for (cd = mk->cache; cd; cd = cd->next) {
-	if (cd->size < mk->min_cached_seg_size)
-	    mk->min_cached_seg_size = cd->size;
-	if (cd->size > mk->max_cached_seg_size)
-	    mk->max_cached_seg_size = cd->size;
-    }
+    ASSERT( mk->cache_size >= 0 );
+
+    return mk->cache_size;
 }
 
-static ERTS_INLINE void
-adjust_cache_size(MemKind* mk, int force_check_limits)
-{
-    cache_desc_t *cd;
-    int check_limits = force_check_limits;
-    Sint max_cached = ((Sint) mk->segments.current.watermark
-		       - (Sint) mk->segments.current.no);
-    ERTS_DBG_MK_CHK_THR_ACCESS(mk);
-    while (((Sint) mk->cache_size) > max_cached && ((Sint) mk->cache_size) > 0) {
-	ASSERT(mk->cache_end);
-	cd = mk->cache_end;
-	if (!check_limits &&
-	    !(mk->min_cached_seg_size < cd->size
-	      && cd->size < mk->max_cached_seg_size)) {
-	    check_limits = 1;
-	}
+static ERTS_INLINE Uint mseg_drop_memkind_cache_size(MemKind *mk, cache_t **head) {
+    cache_t *c = NULL, *next = NULL;
+
+    c = *head;
+    ASSERT( c != NULL );
+
+    while (c) {
+
+	next = c->next;
+
 	if (erts_mtrace_enabled)
-	    erts_mtrace_crr_free(SEGTYPE, SEGTYPE, cd->seg);
-	mseg_destroy(mk->ma, mk, cd->seg, cd->size);
-	unlink_cd(mk,cd);
-	free_cd(mk,cd);
-    }
+	    erts_mtrace_crr_free(SEGTYPE, SEGTYPE, c->seg);
 
-    if (check_limits)
-	check_cache_limits(mk);
-}
+	mseg_destroy(mk->ma, mk, c->seg, c->size);
+	mseg_cache_free_descriptor(mk, c);
 
-static Uint
-check_one_cache(MemKind* mk)
-{
-    if (mk->segments.current.watermark > mk->segments.current.no)
 	mk->segments.current.watermark--;
-    adjust_cache_size(mk, 0);
+	mk->cache_size--;
+
+	c = next;
+    }
+
+    *head = NULL;
+
+    ASSERT( mk->cache_size >= 0 );
 
-    if (mk->cache_size)
-	schedule_cache_check(mk->ma);
     return mk->cache_size;
 }
 
-static void do_cache_check(ErtsMsegAllctr_t *ma)
-{
-    int empty_cache = 1;
+/* mseg_check_memkind_cache
+ * - Check if we can empty some cached segments in this
+ *   MemKind.
+ */
+
+
+static Uint mseg_check_memkind_cache(MemKind *mk) {
+    int i;
+
+    ERTS_DBG_MK_CHK_THR_ACCESS(mk);
+
+    for (i = 0; i < CACHE_AREAS; i++) {
+	if (mk->cache_area[i] != NULL)
+	    return mseg_drop_one_memkind_cache_size(mk, &(mk->cache_area[i]));
+    }
+
+    if (mk->cache_unpowered)
+	return mseg_drop_one_memkind_cache_size(mk, &(mk->cache_unpowered));
+
+    return 0;
+}
+
+/* mseg_cache_check
+ * - Check if we have some cache we can purge
+ *   in any of the memkinds.
+ */
+
+static void mseg_cache_check(ErtsMsegAllctr_t *ma) {
     MemKind* mk;
+    Uint empty_cache = 1;
 
     ERTS_MSEG_LOCK(ma);
 
-    for (mk=ma->mk_list; mk; mk=mk->next) {
-	if (check_one_cache(mk))
+    for (mk = ma->mk_list; mk; mk = mk->next) {
+	if (mseg_check_memkind_cache(mk))
 	    empty_cache = 0;
     }
 
+    /* If all MemKinds caches are empty,
+     * remove aux-work callback
+     */
     if (empty_cache) {
 	ma->is_cache_check_scheduled = 0;
-	erts_set_aux_work_timeout(ma->ix,
-				  ERTS_SSI_AUX_WORK_MSEG_CACHE_CHECK,
-				  0);
+	erts_set_aux_work_timeout(ma->ix, ERTS_SSI_AUX_WORK_MSEG_CACHE_CHECK, 0);
     }
 
     INC_CC(ma, check_cache);
@@ -627,27 +782,65 @@ static void do_cache_check(ErtsMsegAllctr_t *ma)
     ERTS_MSEG_UNLOCK(ma);
 }
 
-void erts_mseg_cache_check(void)
-{
-    do_cache_check(ERTS_MSEG_ALLCTR_SS());
+/* erts_mseg_cache_check
+ * - This is a callback that is scheduled as aux-work from
+ *   schedulers and is called at some interval if we have a cache
+ *   on this mseg-allocator and memkind.
+ * - Purpose: Empty cache slowly so we don't collect mapped areas
+ *   and bloat memory.
+ */
+
+void erts_mseg_cache_check(void) {
+    mseg_cache_check(ERTS_MSEG_ALLCTR_SS());
 }
 
-static void
-mseg_clear_cache(MemKind* mk)
-{
-    mk->segments.current.watermark = 0;
 
-    adjust_cache_size(mk, 1);
+/* *_mseg_clear_*_cache
+ * Remove cached segments from the allocator completely
+ */
+
+static void mseg_clear_memkind_cache(MemKind *mk) {
+    int i;
+
+    /* drop pow2 caches */
+    for (i = 0; i < CACHE_AREAS; i++) {
+	if (mk->cache_area[i] == NULL)
+	    continue;
+
+	mseg_drop_memkind_cache_size(mk, &(mk->cache_area[i]));
+	ASSERT(mk->cache_area[i] == NULL);
+    }
+    /* drop varied caches */
+    if(mk->cache_unpowered)
+	mseg_drop_memkind_cache_size(mk, &(mk->cache_unpowered));
+
+    ASSERT(mk->cache_unpowered == NULL);
+    ASSERT(mk->cache_size == 0);
+}
+
+static void mseg_clear_cache(ErtsMsegAllctr_t *ma) {
+    MemKind* mk;
+
+    ERTS_MSEG_LOCK(ma);
+    ERTS_DBG_MA_CHK_THR_ACCESS(ma);
+
 
-    ASSERT(!mk->cache);
-    ASSERT(!mk->cache_end);
-    ASSERT(!mk->cache_size);
+    for (mk = ma->mk_list; mk; mk = mk->next) {
+	mseg_clear_memkind_cache(mk);
+    }
 
-    mk->segments.current.watermark = mk->segments.current.no;
+    INC_CC(ma, clear_cache);
 
-    INC_CC(mk->ma, clear_cache);
+    ERTS_MSEG_UNLOCK(ma);
 }
 
+void erts_mseg_clear_cache(void) {
+    mseg_clear_cache(ERTS_MSEG_ALLCTR_SS());
+    mseg_clear_cache(ERTS_MSEG_ALLCTR_IX(0));
+}
+
+
+
 static ERTS_INLINE MemKind* memkind(ErtsMsegAllctr_t *ma,
 				    const ErtsMsegOpt_t *opt)
 {
@@ -660,116 +853,40 @@ static ERTS_INLINE MemKind* memkind(ErtsMsegAllctr_t *ma,
 
 static void *
 mseg_alloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, Uint *size_p,
-	   const ErtsMsegOpt_t *opt)
+	   Uint flags, const ErtsMsegOpt_t *opt)
 {
-    Uint max, min, diff_size, size;
-    cache_desc_t *cd, *cand_cd;
+    Uint size;
     void *seg;
     MemKind* mk = memkind(ma, opt);
 
     INC_CC(ma, alloc);
 
-    size = PAGE_CEILING(*size_p);
+    /* Carrier align */
+    size = ALIGNED_CEILING(*size_p);
+
+    /* Cache optim (if applicable) */
+    if (MSEG_FLG_IS_2POW(flags) && !IS_2POW(size))
+	size = ceil_2pow(size);
 
 #if CAN_PARTLY_DESTROY
     if (size < ma->min_seg_size)	
 	ma->min_seg_size = size;
 #endif
+    
+    if (opt->cache && mk->cache_size > 0 && (seg = cache_get_segment(mk, &size)) != NULL) 
+	goto done;
 
-    if (!opt->cache) {
-    create_seg:
-	adjust_cache_size(mk,0);
-	seg = mseg_create(ma, mk, size);
-	if (!seg) {
-	    mseg_clear_cache(mk);
-	    seg = mseg_create(ma, mk, size);
-	    if (!seg)
-		size = 0;
-	}
-
-	*size_p = size;
-	if (seg) {
-	    if (erts_mtrace_enabled)
-		erts_mtrace_crr_alloc(seg, atype, ERTS_MTRACE_SEGMENT_ID, size);
-	    ERTS_MSEG_ALLOC_STAT(mk,size);
-	}
-	return seg;
-    }
-
-    if (size > mk->max_cached_seg_size)
-	goto create_seg;
-
-    if (size < mk->min_cached_seg_size) {
-
-	diff_size = mk->min_cached_seg_size - size;
-
-	if (diff_size > ma->abs_max_cache_bad_fit)
-	    goto create_seg;
-
-	if (100*PAGES(diff_size) > ma->rel_max_cache_bad_fit*PAGES(size))
-	    goto create_seg;
-
-    }
-
-    max = 0;
-    min = ~((Uint) 0);
-    cand_cd = NULL;
-
-    for (cd = mk->cache; cd; cd = cd->next) {
-	if (cd->size >= size) {
-	    if (!cand_cd) {
-		cand_cd = cd;
-		continue;
-	    }
-	    else if (cd->size < cand_cd->size) {
-		if (max < cand_cd->size)
-		    max = cand_cd->size;
-		if (min > cand_cd->size)
-		    min = cand_cd->size;
-		cand_cd = cd;
-		continue;
-	    }
-	}
-	if (max < cd->size)
-	    max = cd->size;
-	if (min > cd->size)
-	    min = cd->size;
-    }
-
-    mk->min_cached_seg_size = min;
-    mk->max_cached_seg_size = max;
-
-    if (!cand_cd)
-	goto create_seg;
-
-    diff_size = cand_cd->size - size;
-
-    if (diff_size > ma->abs_max_cache_bad_fit
-	|| 100*PAGES(diff_size) > ma->rel_max_cache_bad_fit*PAGES(size)) {
-	if (mk->max_cached_seg_size < cand_cd->size)
-	    mk->max_cached_seg_size = cand_cd->size;
-	if (mk->min_cached_seg_size > cand_cd->size)
-	    mk->min_cached_seg_size = cand_cd->size;
-	goto create_seg;
-    }
-
-    mk->cache_hits++;
-
-    size = cand_cd->size;
-    seg = cand_cd->seg;
-
-    unlink_cd(mk,cand_cd);
-    free_cd(mk,cand_cd);
+    if ((seg = mseg_create(ma, mk, size)) == NULL)
+	size = 0;
 
+done:
     *size_p = size;
+    if (seg) {
+	if (erts_mtrace_enabled)
+	    erts_mtrace_crr_alloc(seg, atype, ERTS_MTRACE_SEGMENT_ID, size);
 
-    if (erts_mtrace_enabled) {
-	erts_mtrace_crr_free(SEGTYPE, SEGTYPE, seg);
-	erts_mtrace_crr_alloc(seg, atype, SEGTYPE, size);
-    }
-
-    if (seg)
 	ERTS_MSEG_ALLOC_STAT(mk,size);
+    }
 
     return seg;
 }
@@ -780,73 +897,42 @@ mseg_dealloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg, Uint size,
 	     const ErtsMsegOpt_t *opt)
 {
     MemKind* mk = memkind(ma, opt);
-    cache_desc_t *cd;
+
 
     ERTS_MSEG_DEALLOC_STAT(mk,size);
 
-    if (!opt->cache || ma->max_cache_size == 0) {
-	if (erts_mtrace_enabled)
-	    erts_mtrace_crr_free(atype, SEGTYPE, seg);
-	mseg_destroy(ma, mk, seg, size);
+    if (opt->cache && cache_bless_segment(mk, seg, size)) {
+	schedule_cache_check(ma);
+	goto done;
     }
-    else {
-	int check_limits = 0;
-	
-	if (size < mk->min_cached_seg_size)
-	    mk->min_cached_seg_size = size;
-	if (size > mk->max_cached_seg_size)
-	    mk->max_cached_seg_size = size;
-
-	if (!mk->free_cache_descs) {
-	    cd = mk->cache_end;
-	    if (!(mk->min_cached_seg_size < cd->size
-		  && cd->size < mk->max_cached_seg_size)) {
-		check_limits = 1;
-	    }
-	    if (erts_mtrace_enabled)
-		erts_mtrace_crr_free(SEGTYPE, SEGTYPE, cd->seg);
-	    mseg_destroy(ma, mk, cd->seg, cd->size);
-	    unlink_cd(mk,cd);
-	    free_cd(mk,cd);
-	}
 
-	cd = alloc_cd(mk);
-	ASSERT(cd);
-	cd->seg = seg;
-	cd->size = size;
-	link_cd(mk,cd);
+    if (erts_mtrace_enabled)
+	erts_mtrace_crr_free(atype, SEGTYPE, seg);
 
-	if (erts_mtrace_enabled) {
-	    erts_mtrace_crr_free(atype, SEGTYPE, seg);
-	    erts_mtrace_crr_alloc(seg, SEGTYPE, SEGTYPE, size);
-	}
-
-	/* ASSERT(segments.current.watermark >= segments.current.no + cache_size); */
-
-	if (check_limits)
-	    check_cache_limits(mk);
+    mseg_destroy(ma, mk, seg, size);
 
-	schedule_cache_check(ma);
-
-    }
+done:
 
     INC_CC(ma, dealloc);
 }
 
 static void *
 mseg_realloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg,
-	     Uint old_size, Uint *new_size_p, const ErtsMsegOpt_t *opt)
+	     Uint old_size, Uint *new_size_p, Uint flags, const ErtsMsegOpt_t *opt)
 {
     MemKind* mk;
     void *new_seg;
     Uint new_size;
 
+    /* Just allocate a new segment if we didn't have one before */
     if (!seg || !old_size) {
-	new_seg = mseg_alloc(ma, atype, new_size_p, opt);
+	new_seg = mseg_alloc(ma, atype, new_size_p, flags, opt);
 	DEC_CC(ma, alloc);
 	return new_seg;
     }
 
+
+    /* Dealloc old segment if new segment is of size 0 */
     if (!(*new_size_p)) {
 	mseg_dealloc(ma, atype, seg, old_size, opt);
 	DEC_CC(ma, dealloc);
@@ -855,7 +941,13 @@ mseg_realloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg,
 
     mk = memkind(ma, opt);
     new_seg = seg;
-    new_size = PAGE_CEILING(*new_size_p);
+
+    /* Carrier align */
+    new_size = ALIGNED_CEILING(*new_size_p);
+
+    /* Cache optim (if applicable) */
+    if (MSEG_FLG_IS_2POW(flags) && !IS_2POW(new_size))
+	new_size = ceil_2pow(new_size);
 
     if (new_size == old_size)
 	;
@@ -866,53 +958,27 @@ mseg_realloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg,
 	if (new_size < ma->min_seg_size)	
 	    ma->min_seg_size = new_size;
 #endif
-
+	/* +M<S>rsbcst <ratio> */
 	if (shrink_sz < opt->abs_shrink_th
-	    && 100*PAGES(shrink_sz) < opt->rel_shrink_th*PAGES(old_size)) {
+	    && 100*shrink_sz < opt->rel_shrink_th*old_size) {
 	    new_size = old_size;
 	}
 	else {
 
 #if CAN_PARTLY_DESTROY
 
-	    if (shrink_sz > ma->min_seg_size
-		&& mk->free_cache_descs
-		&& opt->cache) {
-		cache_desc_t *cd;
-
-		cd = alloc_cd(mk);
-		ASSERT(cd);
-		cd->seg = ((char *) seg) + new_size;
-		cd->size = shrink_sz;
-		end_link_cd(mk,cd);
-
-		if (erts_mtrace_enabled) {
-		    erts_mtrace_crr_realloc(new_seg,
-					    atype,
-					    SEGTYPE,
-					    seg,
-					    new_size);
-		    erts_mtrace_crr_alloc(cd->seg, SEGTYPE, SEGTYPE, cd->size);
-		}
-		schedule_cache_check(ma);
-	    }
-	    else {
-		if (erts_mtrace_enabled)
-		    erts_mtrace_crr_realloc(new_seg,
-					    atype,
-					    SEGTYPE,
-					    seg,
-					    new_size);
-		mseg_destroy(ma, mk, ((char *) seg) + new_size, shrink_sz);
-	    }
+	    if (erts_mtrace_enabled)
+		erts_mtrace_crr_realloc(new_seg, atype, SEGTYPE, seg, new_size);
 
-#elif HAVE_MSEG_RECREATE
+	    mseg_destroy(ma, mk, ((char *) seg) + new_size, shrink_sz);
 
+#elif HAVE_MSEG_RECREATE
 	    goto do_recreate;
-
 #else
+	    new_seg = mseg_alloc(ma, atype, &new_size, flags, opt);
+
+	    ASSERT(MAP_IS_ALIGNED(new_seg) || !new_seg);
 
-	    new_seg = mseg_alloc(ma, atype, &new_size, opt);
 	    if (!new_seg)
 		new_size = old_size;
 	    else {
@@ -921,16 +987,15 @@ mseg_realloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg,
 			   MIN(new_size, old_size));
 		mseg_dealloc(ma, atype, seg, old_size, opt);
 	    }
-
 #endif
-
 	}
     }
     else {
 
 	if (!opt->preserv) {
 	    mseg_dealloc(ma, atype, seg, old_size, opt);
-	    new_seg = mseg_alloc(ma, atype, &new_size, opt);
+	    new_seg = mseg_alloc(ma, atype, &new_size, flags, opt);
+	    ASSERT(MAP_IS_ALIGNED(new_seg) || !new_seg);
 	}
 	else {
 #if HAVE_MSEG_RECREATE
@@ -938,18 +1003,23 @@ mseg_realloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg,
 	do_recreate:
 #endif
 	    new_seg = mseg_recreate(ma, mk, (void *) seg, old_size, new_size);
+	    /* ASSERT(MAP_IS_ALIGNED(new_seg) || !new_seg);
+	     * will not always be aligned and it ok for now
+	     */
+
 	    if (erts_mtrace_enabled)
 		erts_mtrace_crr_realloc(new_seg, atype, SEGTYPE, seg, new_size);
 	    if (!new_seg)
 		new_size = old_size;
 #else
-	    new_seg = mseg_alloc(ma, atype, &new_size, opt);
+	    new_seg = mseg_alloc(ma, atype, &new_size, flags, opt);
+
+	    ASSERT(MAP_IS_ALIGNED(new_seg) || !new_seg);
+
 	    if (!new_seg)
 		new_size = old_size;
 	    else {
-		sys_memcpy(((char *) new_seg),
-			   ((char *) seg),
-			   MIN(new_size, old_size));
+		sys_memcpy(((char *) new_seg), ((char *) seg), MIN(new_size, old_size));
 		mseg_dealloc(ma, atype, seg, old_size, opt);
 	    }
 #endif
@@ -958,6 +1028,7 @@ mseg_realloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg,
 
     INC_CC(ma, realloc);
 
+    ASSERT(!MSEG_FLG_IS_2POW(flags) || IS_2POW(new_size));
     *new_size_p = new_size;
 
     ERTS_MSEG_REALLOC_STAT(mk, old_size, new_size);
@@ -990,6 +1061,7 @@ static struct {
     Eterm mseg_dealloc;
     Eterm mseg_realloc;
     Eterm mseg_create;
+    Eterm mseg_create_resize;
     Eterm mseg_destroy;
 #if HAVE_MSEG_RECREATE
     Eterm mseg_recreate;
@@ -1046,6 +1118,7 @@ init_atoms(ErtsMsegAllctr_t *ma)
 	AM_INIT(mseg_dealloc);
 	AM_INIT(mseg_realloc);
 	AM_INIT(mseg_create);
+	AM_INIT(mseg_create_resize);
 	AM_INIT(mseg_destroy);
 #if HAVE_MSEG_RECREATE
 	AM_INIT(mseg_recreate);
@@ -1065,14 +1138,12 @@ init_atoms(ErtsMsegAllctr_t *ma)
     erts_mtx_unlock(&init_atoms_mutex);
 }
 
-
 #define bld_uint	erts_bld_uint
 #define bld_cons	erts_bld_cons
 #define bld_tuple	erts_bld_tuple
 #define bld_string	erts_bld_string
 #define bld_2tup_list	erts_bld_2tup_list
 
-
 /*
  * bld_unstable_uint() (instead of bld_uint()) is used when values may
  * change between size check and actual build. This because a value
@@ -1116,6 +1187,7 @@ add_4tup(Uint **hpp, Uint *szp, Eterm *lp,
     *lp = bld_cons(hpp, szp, bld_tuple(hpp, szp, 4, el1, el2, el3, el4), *lp);
 }
 
+
 static Eterm
 info_options(ErtsMsegAllctr_t *ma,
 	     char *prefix,
@@ -1176,6 +1248,7 @@ info_calls(ErtsMsegAllctr_t *ma, int *print_to_p, void *print_to_arg, Uint **hpp
 	PRINT_CC(to, arg, dealloc);
 	PRINT_CC(to, arg, realloc);
 	PRINT_CC(to, arg, create);
+	PRINT_CC(to, arg, create_resize);
 	PRINT_CC(to, arg, destroy);
 #if HAVE_MSEG_RECREATE
 	PRINT_CC(to, arg, recreate);
@@ -1215,6 +1288,10 @@ info_calls(ErtsMsegAllctr_t *ma, int *print_to_p, void *print_to_arg, Uint **hpp
 		 bld_unstable_uint(hpp, szp, ma->calls.create.giga_no),
 		 bld_unstable_uint(hpp, szp, ma->calls.create.no));
 
+	add_3tup(hpp, szp, &res,
+		 am.mseg_create_resize,
+		 bld_unstable_uint(hpp, szp, ma->calls.create_resize.giga_no),
+		 bld_unstable_uint(hpp, szp, ma->calls.create_resize.no));
 
 	add_3tup(hpp, szp, &res,
 		 am.mseg_realloc,
@@ -1401,21 +1478,21 @@ erts_mseg_info(int ix,
 }
 
 void *
-erts_mseg_alloc_opt(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt)
+erts_mseg_alloc_opt(ErtsAlcType_t atype, Uint *size_p, Uint flags, const ErtsMsegOpt_t *opt)
 {
     ErtsMsegAllctr_t *ma = ERTS_MSEG_ALLCTR_OPT(opt);
     void *seg;
     ERTS_MSEG_LOCK(ma);
     ERTS_DBG_MA_CHK_THR_ACCESS(ma);
-    seg = mseg_alloc(ma, atype, size_p, opt);
+    seg = mseg_alloc(ma, atype, size_p, flags, opt);
     ERTS_MSEG_UNLOCK(ma);
     return seg;
 }
 
 void *
-erts_mseg_alloc(ErtsAlcType_t atype, Uint *size_p)
+erts_mseg_alloc(ErtsAlcType_t atype, Uint *size_p, Uint flags)
 {
-    return erts_mseg_alloc_opt(atype, size_p, &erts_mseg_default_opt);
+    return erts_mseg_alloc_opt(atype, size_p, flags, &erts_mseg_default_opt);
 }
 
 void
@@ -1438,44 +1515,24 @@ erts_mseg_dealloc(ErtsAlcType_t atype, void *seg, Uint size)
 void *
 erts_mseg_realloc_opt(ErtsAlcType_t atype, void *seg,
 		      Uint old_size, Uint *new_size_p,
+		      Uint flags,
 		      const ErtsMsegOpt_t *opt)
 {
     ErtsMsegAllctr_t *ma = ERTS_MSEG_ALLCTR_OPT(opt);
     void *new_seg;
     ERTS_MSEG_LOCK(ma);
     ERTS_DBG_MA_CHK_THR_ACCESS(ma);
-    new_seg = mseg_realloc(ma, atype, seg, old_size, new_size_p, opt);
+    new_seg = mseg_realloc(ma, atype, seg, old_size, new_size_p, flags, opt);
     ERTS_MSEG_UNLOCK(ma);
     return new_seg;
 }
 
 void *
 erts_mseg_realloc(ErtsAlcType_t atype, void *seg,
-		  Uint old_size, Uint *new_size_p)
+		  Uint old_size, Uint *new_size_p, Uint flags)
 {
     return erts_mseg_realloc_opt(atype, seg, old_size, new_size_p,
-				 &erts_mseg_default_opt);
-}
-
-void
-erts_mseg_clear_cache(void)
-{
-    ErtsMsegAllctr_t *ma = ERTS_MSEG_ALLCTR_SS();
-    MemKind* mk;
-
-start:
-
-    ERTS_MSEG_LOCK(ma);
-    ERTS_DBG_MA_CHK_THR_ACCESS(ma);
-    for (mk=ma->mk_list; mk; mk=mk->next) {
-	mseg_clear_cache(mk);
-    }
-    ERTS_MSEG_UNLOCK(ma);
-
-    if (ma->ix != 0) {
-	ma = ERTS_MSEG_ALLCTR_IX(0);
-	goto start;
-    }
+				 flags, &erts_mseg_default_opt);
 }
 
 Uint
@@ -1496,28 +1553,32 @@ erts_mseg_no(const ErtsMsegOpt_t *opt)
 Uint
 erts_mseg_unit_size(void)
 {
-    return page_size;
+    return MSEG_ALIGNED_SIZE;
 }
 
 static void mem_kind_init(ErtsMsegAllctr_t *ma, MemKind* mk, const char* name)
 {
-    unsigned i;
+    int i;
 
-    mk->cache = NULL;
-    mk->cache_end = NULL;
-    mk->max_cached_seg_size = 0;
-    mk->min_cached_seg_size = ~((Uint) 0);
-    mk->cache_size = 0;
-    mk->cache_hits = 0;
+    for (i = 0; i < CACHE_AREAS; i++) {
+	mk->cache_area[i] = NULL;
+    }
+
+    mk->cache_free = NULL;
 
-    if (ma->max_cache_size > 0) {
-	for (i = 0; i < ma->max_cache_size - 1; i++)
-	    mk->cache_descs[i].next = &mk->cache_descs[i + 1];
-	mk->cache_descs[ma->max_cache_size - 1].next = NULL;
-	mk->free_cache_descs = &mk->cache_descs[0];
+    ASSERT(ma->max_cache_size <= MAX_CACHE_SIZE);
+
+    for (i = 0; i < ma->max_cache_size; i++) {
+	mk->cache[i].seg  = NULL;
+	mk->cache[i].size = 0;
+	mk->cache[i].next = mk->cache_free;
+	mk->cache_free = &(mk->cache[i]);
     }
-    else
-	mk->free_cache_descs = NULL;
+
+    mk->cache_unpowered = NULL;
+
+    mk->cache_size = 0;
+    mk->cache_hits = 0;
 
     mk->segments.current.watermark = 0;
     mk->segments.current.no = 0;
@@ -1570,15 +1631,10 @@ erts_mseg_init(ErtsMsegInit_t *init)
     initialize_pmmap();
 #endif
 
-    page_size = GET_PAGE_SIZE;
+    if (!IS_2POW(GET_PAGE_SIZE))
+	erl_exit(ERTS_ABORT_EXIT, "erts_mseg: Unexpected page_size %beu\n", GET_PAGE_SIZE);
 
-    page_shift = 1;
-    while ((page_size >> page_shift) != 1) {
-	if ((page_size & (1 << (page_shift - 1))) != 0)
-	    erl_exit(ERTS_ABORT_EXIT,
-		     "erts_mseg: Unexpected page_size %beu\n", page_size);
-	page_shift++;
-    }
+    ASSERT((MSEG_ALIGNED_SIZE % GET_PAGE_SIZE) == 0);
 
     for (i = 0; i < no_mseg_allocators; i++) {
 	ErtsMsegAllctr_t *ma = ERTS_MSEG_ALLCTR_IX(i);
@@ -1663,7 +1719,7 @@ erts_mseg_test(unsigned long op,
     case 0x400: /* Have erts_mseg */
 	return (unsigned long) 1;
     case 0x401:
-	return (unsigned long) erts_mseg_alloc(ERTS_ALC_A_INVALID, (Uint *) a1);
+	return (unsigned long) erts_mseg_alloc(ERTS_ALC_A_INVALID, (Uint *) a1, (Uint) 0);
     case 0x402:
 	erts_mseg_dealloc(ERTS_ALC_A_INVALID, (void *) a1, (Uint) a2);
 	return (unsigned long) 0;
@@ -1671,7 +1727,8 @@ erts_mseg_test(unsigned long op,
 	return (unsigned long) erts_mseg_realloc(ERTS_ALC_A_INVALID,
 						 (void *) a1,
 						 (Uint) a2,
-						 (Uint *) a3);
+						 (Uint *) a3,
+						 (Uint) 0);
     case 0x404:
 	erts_mseg_clear_cache();
 	return (unsigned long) 0;
@@ -1707,7 +1764,40 @@ erts_mseg_test(unsigned long op,
  * mapping tricks.
  */
 
-/*#define HARDDEBUG 1*/
+/* #define HARDDEBUG 1 */
+
+#ifdef HARDDEBUG
+static void dump_freelist(void)
+{
+    FreeBlock *p = first;
+
+    while (p) {
+	fprintf(stderr, "p = %p\r\np->num = %ld\r\np->next = %p\r\n\r\n",
+	       (void *) p, (unsigned long) p->num, (void *) p->next);
+	p = p->next;
+    }
+}
+
+#define HARDDEBUG_HW_INCOMPLETE_ALIGNMENT(PTR, SZ)            \
+    fprintf(stderr,"Mapping of address %p with size %ld "     \
+	    "does not map complete pages (%s:%d)\r\n",        \
+	    (void *) (PTR), (unsigned long) (SZ),__FILE__, __LINE__)
+
+#define HARDDEBUG_HW_UNALIGNED_ALIGNMENT(PTR, SZ)             \
+    fprintf(stderr,"Mapping of address %p with size %ld "     \
+	    "is not page aligned (%s:%d)\r\n",                \
+	    (void *) (PTR), (unsigned long) (SZ),__FILE__, __LINE__)
+
+#define HARDDEBUG_MAP_FAILED(PTR, SZ) \
+    fprintf(stderr, "Could not actually map memory "          \
+	    "at address %p with size %ld (%s:%d) ..\r\n",     \
+	    (void *) (PTR), (unsigned long) (SZ),__FILE__, __LINE__)
+#else
+#define HARDDEBUG_HW_INCOMPLETE_ALIGNMENT(PTR, SZ) do{}while(0)
+#define HARDDEBUG_HW_UNALIGNED_ALIGNMENT(PTR, SZ)  do{}while(0)
+#define HARDDEBUG_MAP_FAILED(PTR, SZ)              do{}while(0)
+#endif
+
 
 #ifdef __APPLE__
 #define MAP_ANONYMOUS MAP_ANON
@@ -1726,49 +1816,20 @@ typedef struct _free_block {
     struct _free_block *next;
 } FreeBlock;
 
-/* Assigned once and for all */
-static size_t pagsz;
-
 /* Protect with lock */
 static FreeBlock *first;
 
-static size_t round_up_to_pagesize(size_t size)
-{
-    size_t x  = size / pagsz;
-
-    if ((size % pagsz)) {
-	++x;
-    }
-
-    return pagsz * x;
-}
-
-static size_t round_down_to_pagesize(size_t size)
-{
-    size_t x  = size / pagsz;
-
-    return pagsz * x;
-}
-
 static void *do_map(void *ptr, size_t sz)
 {
     void *res;
 
-    if (round_up_to_pagesize(sz) != sz) {
-#ifdef HARDDEBUG
-	fprintf(stderr,"Mapping of address %p with size %ld "
-		"does not map complete pages\r\n",
-		(void *) ptr, (unsigned long) sz);
-#endif
+    if (ALIGNED_CEILING(sz) != sz) {
+	HARDDEBUG_HW_INCOMPLETE_ALIGNMENT(ptr, sz);
 	return NULL;
     }
 
-    if (((unsigned long) ptr) % pagsz) {
-#ifdef HARDDEBUG
-	fprintf(stderr,"Mapping of address %p with size %ld "
-		"is not page aligned\r\n",
-		(void *) ptr, (unsigned long) sz);
-#endif
+    if (((unsigned long) ptr) % MSEG_ALIGNED_SIZE) {
+	HARDDEBUG_HW_UNALIGNED_ALIGNMENT(ptr, sz);
 	return NULL;
     }
 
@@ -1782,10 +1843,7 @@ static void *do_map(void *ptr, size_t sz)
 #endif
 
     if (res == MAP_FAILED) {
-#ifdef HARDDEBUG
-	fprintf(stderr,"Mapping of address %p with size %ld failed!\r\n",
-		(void *) ptr, (unsigned long) sz);
-#endif
+	HARDDEBUG_MAP_FAILED(ptr, sz);
 	return NULL;
     }
 
@@ -1796,35 +1854,22 @@ static int do_unmap(void *ptr, size_t sz)
 {
     void *res;
 
-    if (round_up_to_pagesize(sz) != sz) {
-#ifdef HARDDEBUG
-	fprintf(stderr,"Mapping of address %p with size %ld "
-		"does not map complete pages\r\n",
-		(void *) ptr, (unsigned long) sz);
-#endif
+    if (ALIGNED_CEILING(sz) != sz) {
+	HARDDEBUG_HW_INCOMPLETE_ALIGNMENT(ptr, sz);
 	return 1;
     }
 
-    if (((unsigned long) ptr) % pagsz) {
-#ifdef HARDDEBUG
-	fprintf(stderr,"Mapping of address %p with size %ld "
-		"is not page aligned\r\n",
-		(void *) ptr, (unsigned long) sz);
-#endif
+    if (((unsigned long) ptr) % MSEG_ALIGNED_SIZE) {
+	HARDDEBUG_HW_UNALIGNED_ALIGNMENT(ptr, sz);
 	return 1;
     }
 
-
     res = mmap(ptr, sz,
-	       PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE
-	       | MAP_FIXED,
+	       PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE | MAP_FIXED,
 	       -1 , 0);
 
     if (res == MAP_FAILED) {
-#ifdef HARDDEBUG
-	fprintf(stderr,"Mapping of address %p with size %ld failed!\r\n",
-		(void *) ptr, (unsigned long) sz);
-#endif
+	HARDDEBUG_MAP_FAILED(ptr, sz);
 	return 1;
     }
 
@@ -1862,8 +1907,6 @@ static int initialize_pmmap(void)
     size_t rsz;
     FreeBlock *initial;
 
-
-    pagsz = getpagesize();
     SET_RANGE_MIN();
     if (sizeof(void *) != 8) {
 	erl_exit(1,"Halfword emulator cannot be run in 32bit mode");
@@ -1872,15 +1915,15 @@ static int initialize_pmmap(void)
     p = (char *) RANGE_MIN;
     q = (char *) RANGE_MAX;
 
-    rsz = round_down_to_pagesize(q - p);
+    rsz = ALIGNED_FLOOR(q - p);
 
-    rptr = mmap((void *) p, rsz,
+    rptr = mmap_align(NULL, (void *) p, rsz,
 		PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS |
 		MAP_NORESERVE | EXTRA_MAP_FLAGS,
 		-1 , 0);
 #ifdef HARDDEBUG
     printf("p=%p, rsz = %ld, pages = %ld, got range = %p -> %p\r\n",
-	   p, (unsigned long) rsz, (unsigned long) (rsz / pagsz),
+	   p, (unsigned long) rsz, (unsigned long) (rsz / MSEG_ALIGNED_SIZE),
 	   (void *) rptr, (void*)(rptr + rsz));
 #endif
     if ((UWord)(rptr + rsz) > RANGE_MAX) {
@@ -1892,39 +1935,27 @@ static int initialize_pmmap(void)
 	munmap((void*)RANGE_MAX, rsz - rsz_trunc);
 	rsz = rsz_trunc;
     }
-    if (!do_map(rptr,pagsz)) {
+    if (!do_map(rptr, MSEG_ALIGNED_SIZE)) {
 	erl_exit(1,"Could not actually mmap first page for halfword emulator...\n");
     }
     initial = (FreeBlock *) rptr;
-    initial->num = (rsz / pagsz);
+    initial->num = (rsz / MSEG_ALIGNED_SIZE);
     initial->next = NULL;
     first = initial;
     INIT_LOCK();
     return 0;
 }
 
-#ifdef HARDDEBUG
-static void dump_freelist(void)
-{
-    FreeBlock *p = first;
-
-    while (p) {
-	printf("p = %p\r\np->num = %ld\r\np->next = %p\r\n\r\n",
-	       (void *) p, (unsigned long) p->num, (void *) p->next);
-	p = p->next;
-    }
-}
-#endif
-
-
 static void *pmmap(size_t size)
 {
-    size_t real_size = round_up_to_pagesize(size);
-    size_t num_pages = real_size / pagsz;
+    size_t real_size = ALIGNED_CEILING(size);
+    size_t num_pages = real_size / MSEG_ALIGNED_SIZE;
     FreeBlock **block;
     FreeBlock *tail;
     FreeBlock *res;
+
     TAKE_LOCK();
+
     for (block = &first;
 	 *block != NULL && (*block)->num < num_pages;
 	 block = &((*block)->next))
@@ -1935,29 +1966,25 @@ static void *pmmap(size_t size)
     }
     if ((*block)->num == num_pages) {
 	/* nice, perfect fit */
-	res = *block;
+	res    = *block;
 	*block = (*block)->next;
     } else {
 	tail = (FreeBlock *) (((char *) ((void *) (*block))) + real_size);
-	if (!do_map(tail,pagsz)) {
-#ifdef HARDDEBUG
-	    fprintf(stderr, "Could not actually allocate page at %p...\r\n",
-		    (void *) tail);
-#endif
+	if (!do_map(tail, MSEG_ALIGNED_SIZE)) {
+	    HARDDEBUG_MAP_FAILED(tail, MSEG_ALIGNED_SIZE);
 	    RELEASE_LOCK();
 	    return NULL;
 	}
-	tail->num = (*block)->num - num_pages;
+	tail->num  = (*block)->num - num_pages;
 	tail->next = (*block)->next;
 	res = *block;
 	*block = tail;
     }
+
     RELEASE_LOCK();
-    if (!do_map(res,real_size)) {
-#ifdef HARDDEBUG
-	fprintf(stderr, "Could not actually allocate %ld at %p...\r\n",
-		(unsigned long) real_size, (void *) res);
-#endif
+
+    if (!do_map(res, real_size)) {
+	HARDDEBUG_MAP_FAILED(res, real_size);
 	return NULL;
     }
 
@@ -1966,15 +1993,17 @@ static void *pmmap(size_t size)
 
 static int pmunmap(void *p, size_t size)
 {
-    size_t real_size = round_up_to_pagesize(size);
-    size_t num_pages = real_size / pagsz;
+    size_t real_size = ALIGNED_CEILING(size);
+    size_t num_pages = real_size / MSEG_ALIGNED_SIZE;
+
     FreeBlock *block;
     FreeBlock *last;
     FreeBlock *nb = (FreeBlock *) p;
 
     ASSERT(((unsigned long)p & CHECK_POINTER_MASK)==0);
-    if (real_size > pagsz) {
-	if (do_unmap(((char *) p) + pagsz,real_size - pagsz)) {
+
+    if (real_size > MSEG_ALIGNED_SIZE) {
+	if (do_unmap(((char *) p) + MSEG_ALIGNED_SIZE, real_size - MSEG_ALIGNED_SIZE)) {
 	    return 1;
 	}
     }
@@ -1993,7 +2022,7 @@ static int pmunmap(void *p, size_t size)
 	/* Merge new free block with following */
 	nb->num = block->num + num_pages;
 	nb->next = block->next;
-	if (do_unmap(block,pagsz)) {
+	if (do_unmap(block, MSEG_ALIGNED_SIZE)) {
 	    RELEASE_LOCK();
 	    return 1;
 	}
@@ -2003,11 +2032,11 @@ static int pmunmap(void *p, size_t size)
 	nb->next = block;
     }
     if (last != NULL) {
-	if (p == ((void *) (((char *) last) + (last->num * pagsz)))) {
+	if (p == ((void *) (((char *) last) + (last->num * MSEG_ALIGNED_SIZE)))) {
 	    /* Merge with previous */
 	    last->num += nb->num;
 	    last->next = nb->next;
-	    if (do_unmap(nb,pagsz)) {
+	    if (do_unmap(nb, MSEG_ALIGNED_SIZE)) {
 		RELEASE_LOCK();
 		return 1;
 	    }
@@ -2024,10 +2053,10 @@ static int pmunmap(void *p, size_t size)
 static void *pmremap(void *old_address, size_t old_size,
 	      size_t new_size)
 {
-    size_t new_real_size = round_up_to_pagesize(new_size);
-    size_t new_num_pages = new_real_size / pagsz;
-    size_t old_real_size = round_up_to_pagesize(old_size);
-    size_t old_num_pages = old_real_size / pagsz;
+    size_t new_real_size = ALIGNED_CEILING(new_size);
+    size_t new_num_pages = new_real_size / MSEG_ALIGNED_SIZE;
+    size_t old_real_size = ALIGNED_CEILING(old_size);
+    size_t old_num_pages = old_real_size / MSEG_ALIGNED_SIZE;
     if (new_num_pages == old_num_pages) {
 	return old_address;
     } else if (new_num_pages < old_num_pages) { /* Shrink */
@@ -2045,8 +2074,8 @@ static void *pmremap(void *old_address, size_t old_size,
 	    (*block) > ((FreeBlock *)(((char *) vnfb) + nfb_real_size))) {
 	    /* Normal link in */
 	    if (nfb_pages > 1) {
-		if (do_unmap((void *)(((char *) vnfb) + pagsz),
-			     (nfb_pages - 1)*pagsz)) {
+		if (do_unmap((void *)(((char *) vnfb) + MSEG_ALIGNED_SIZE),
+			     (nfb_pages - 1)*MSEG_ALIGNED_SIZE)) {
 		    return NULL;
 		}
 	    }
@@ -2058,8 +2087,8 @@ static void *pmremap(void *old_address, size_t old_size,
 	    nfb->num = nfb_pages + (*block)->num;
 	    /* unmap also the first page of the next freeblock */
 	    (*block) = nfb;
-	    if (do_unmap((void *)(((char *) vnfb) + pagsz),
-			 nfb_pages*pagsz)) {
+	    if (do_unmap((void *)(((char *) vnfb) + MSEG_ALIGNED_SIZE),
+			 nfb_pages*MSEG_ALIGNED_SIZE)) {
 		return NULL;
 	    }
 	}
@@ -2094,9 +2123,9 @@ static void *pmremap(void *old_address, size_t old_size,
 	    size_t remaining_pages = (*block)->num -
 		(new_num_pages - old_num_pages);
 	    if (!remaining_pages) {
-		void *p = (void *) (((char *) (*block)) + pagsz);
+		void *p = (void *) (((char *) (*block)) + MSEG_ALIGNED_SIZE);
 		void *n = (*block)->next;
-		size_t x = ((*block)->num - 1) * pagsz;
+		size_t x = ((*block)->num - 1) * MSEG_ALIGNED_SIZE;
 		if (x > 0) {
 		    if (do_map(p,x) == NULL) {
 			RELEASE_LOCK();
@@ -2108,7 +2137,7 @@ static void *pmremap(void *old_address, size_t old_size,
 		FreeBlock *nfb = (FreeBlock *) ((void *)
 						(((char *) old_address) +
 						 new_real_size));
-		void *p = (void *) (((char *) (*block)) + pagsz);
+		void *p = (void *) (((char *) (*block)) + MSEG_ALIGNED_SIZE);
 		if (do_map(p,new_real_size - old_real_size) == NULL) {
 		    RELEASE_LOCK();
 		    return NULL;
@@ -2122,5 +2151,4 @@ static void *pmremap(void *old_address, size_t old_size,
 	}
     }
 }
-
 #endif /* HALFWORD_HEAP */
diff --git a/erts/emulator/sys/common/erl_mseg.h b/erts/emulator/sys/common/erl_mseg.h
index 741080fb78..6f373f13f9 100644
--- a/erts/emulator/sys/common/erl_mseg.h
+++ b/erts/emulator/sys/common/erl_mseg.h
@@ -32,12 +32,35 @@
 
 #if HAVE_MMAP
 #  define HAVE_ERTS_MSEG 1
+#  define HAVE_SUPER_ALIGNED_MB_CARRIERS 1
 #else
 #  define HAVE_ERTS_MSEG 0
+#  define HAVE_SUPER_ALIGNED_MB_CARRIERS 0
+#endif
+
+#if HAVE_SUPER_ALIGNED_MB_CARRIERS
+#  define MSEG_ALIGN_BITS (18)
+   /* Affects hard limits for sbct and lmbcs documented in erts_alloc.xml */
+#else
+/* If we don't use super aligned multiblock carriers
+ * we will mmap with page size alignment (and thus use corresponding
+ * align bits).
+ *
+ * Current implementation needs this to be a constant and
+ * only uses this for user dev testing so setting page size
+ * to 4096 (12 bits) is fine.
+ */
+#  define MSEG_ALIGN_BITS       (12)
 #endif
 
 #if HAVE_ERTS_MSEG
 
+#define MSEG_ALIGNED_SIZE     (1 << MSEG_ALIGN_BITS)
+
+#define ERTS_MSEG_FLG_NONE    ((Uint)(0))
+#define ERTS_MSEG_FLG_2POW    ((Uint)(1 << 0))
+
+
 #define ERTS_MSEG_VSN_STR "0.9"
 
 typedef struct {
@@ -68,13 +91,13 @@ typedef struct {
 
 extern const ErtsMsegOpt_t erts_mseg_default_opt;
 
-void *erts_mseg_alloc(ErtsAlcType_t, Uint *);
-void *erts_mseg_alloc_opt(ErtsAlcType_t, Uint *, const ErtsMsegOpt_t *);
+void *erts_mseg_alloc(ErtsAlcType_t, Uint *, Uint);
+void *erts_mseg_alloc_opt(ErtsAlcType_t, Uint *, Uint, const ErtsMsegOpt_t *);
 void  erts_mseg_dealloc(ErtsAlcType_t, void *, Uint);
 void  erts_mseg_dealloc_opt(ErtsAlcType_t, void *, Uint, const ErtsMsegOpt_t *);
-void *erts_mseg_realloc(ErtsAlcType_t, void *, Uint, Uint *);
+void *erts_mseg_realloc(ErtsAlcType_t, void *, Uint, Uint *, Uint);
 void *erts_mseg_realloc_opt(ErtsAlcType_t, void *, Uint, Uint *,
-			    const ErtsMsegOpt_t *);
+			    Uint, const ErtsMsegOpt_t *);
 void  erts_mseg_clear_cache(void);
 void  erts_mseg_cache_check(void);
 Uint  erts_mseg_no( const ErtsMsegOpt_t *);
diff --git a/erts/emulator/test/alloc_SUITE_data/allocator_test.h b/erts/emulator/test/alloc_SUITE_data/allocator_test.h
index cd4a91d34a..c0396ddb61 100644
--- a/erts/emulator/test/alloc_SUITE_data/allocator_test.h
+++ b/erts/emulator/test/alloc_SUITE_data/allocator_test.h
@@ -60,9 +60,9 @@ typedef void* erts_cond;
 #define IS_MMAP_C(C)		((Ulong)	ALC_TEST1(0x00a, (C)))
 #define C_SZ(C)			((Ulong)	ALC_TEST1(0x00b, (C)))
 #define SBC2BLK(A, C)		((Block_t *)	ALC_TEST2(0x00c, (A), (C)))
-#define BLK2SBC(A, B)		((Carrier_t *)	ALC_TEST2(0x00d, (A), (B)))
-#define MBC2FBLK(A, C)		((Block_t *)	ALC_TEST2(0x00e, (A), (C)))
-#define FBLK2MBC(A, B)		((Carrier_t *)	ALC_TEST2(0x00f, (A), (B)))
+#define BLK_TO_SBC(A, B)	((Carrier_t *)	ALC_TEST2(0x00d, (A), (B)))
+#define MBC_TO_FIRST_BLK(A, C)	((Block_t *)	ALC_TEST2(0x00e, (A), (C)))
+#define FIRST_BLK_TO_MBC(A, B)	((Carrier_t *)	ALC_TEST2(0x00f, (A), (B)))
 #define FIRST_MBC(A)		((Carrier_t *)	ALC_TEST1(0x010, (A)))
 #define LAST_MBC(A)		((Carrier_t *)	ALC_TEST1(0x011, (A)))
 #define FIRST_SBC(A)		((Carrier_t *)	ALC_TEST1(0x012, (A)))
@@ -73,7 +73,7 @@ typedef void* erts_cond;
 #define MIN_BLK_SZ(A)		((Ulong)	ALC_TEST1(0x017, (A)))
 #define NXT_BLK(B)		((Block_t *)	ALC_TEST1(0x018, (B)))
 #define PREV_BLK(B)		((Block_t *)	ALC_TEST1(0x019, (B)))
-#define IS_FIRST_BLK(B)		((Ulong)	ALC_TEST1(0x01a, (B)))
+#define IS_MBC_FIRST_BLK(A,B)	((Ulong)	ALC_TEST2(0x01a, (A), (B)))
 #define UNIT_SZ			((Ulong)	ALC_TEST0(0x01b))
 
 /* From erl_goodfit_alloc.c */
diff --git a/erts/emulator/test/alloc_SUITE_data/basic.c b/erts/emulator/test/alloc_SUITE_data/basic.c
index 4a5e888161..0c27665712 100644
--- a/erts/emulator/test/alloc_SUITE_data/basic.c
+++ b/erts/emulator/test/alloc_SUITE_data/basic.c
@@ -44,7 +44,7 @@ testcase_run(TestCaseState_t *tcs)
 
     c = FIRST_MBC(a);
     ASSERT(tcs, !NEXT_C(c));
-    blk = MBC2FBLK(a, c);
+    blk = MBC_TO_FIRST_BLK(a, c);
     ASSERT(tcs, IS_LAST_BLK(blk));
     ASSERT(tcs, IS_FREE_BLK(blk));
 
diff --git a/erts/emulator/test/alloc_SUITE_data/bucket_mask.c b/erts/emulator/test/alloc_SUITE_data/bucket_mask.c
index b214f87e4a..34979cacf1 100644
--- a/erts/emulator/test/alloc_SUITE_data/bucket_mask.c
+++ b/erts/emulator/test/alloc_SUITE_data/bucket_mask.c
@@ -22,7 +22,7 @@
 #include "allocator_test.h"
 #include <stdio.h>
 
-#ifdef __WIN32__ && SIZEOF_VOID_P == 8
+#if defined(__WIN32__) && SIZEOF_VOID_P == 8
 /* Use larger threashold for win64 as block alignment
    is 16 bytes and not 8 */
 #define SBCT ((1024*1024))
@@ -48,10 +48,16 @@ testcase_cleanup(TestCaseState_t *tcs)
 void
 testcase_run(TestCaseState_t *tcs)
 {
-    void *tmp;
-    void **fence;
+    typedef struct linked_block {
+	struct linked_block* next;
+    }Linked;
+    Linked* link;
+    Linked* fence_list;
+    Linked* pad_list;
+    void* tmp;
     void **blk;
     Ulong sz;
+    Ulong residue;
     Ulong smbcs;
     int i;
     int bi;
@@ -73,7 +79,7 @@ testcase_run(TestCaseState_t *tcs)
     ASSERT(tcs, a);
 
     min_blk_sz = MIN_BLK_SZ(a);
-    smbcs = 2*(no_bkts*sizeof(void *) + min_blk_sz) + min_blk_sz;
+    smbcs = (no_bkts*sizeof(void *) + min_blk_sz) + min_blk_sz;
     for (i = 0; i < no_bkts; i++) {
 	sz = BKT_MIN_SZ(a, i);
 	if (sz >= sbct)
@@ -98,26 +104,42 @@ testcase_run(TestCaseState_t *tcs)
     tcs->extra = (void *) a;
     ASSERT(tcs, a);
 
+
     blk = (void **) ALLOC(a, no_bkts*sizeof(void *));
-    fence = (void **) ALLOC(a, no_bkts*sizeof(void *));
 
-    ASSERT(tcs, blk && fence);
+    ASSERT(tcs, blk);
+    fence_list = NULL;
 
     testcase_printf(tcs, "Allocating blocks and fences\n");
     for (i = 0; i < bi_tests; i++) {
 	sz = BKT_MIN_SZ(a, i);
 	blk[i] = ALLOC(a, sz - ablk_hdr_sz);
-	fence[i] = ALLOC(a, 1);
-	ASSERT(tcs, blk[i] && fence[i]);
+	link = (Linked*) ALLOC(a, sizeof(Linked));
+	ASSERT(tcs, blk[i] && link);
+	link->next = fence_list;
+	fence_list = link;
     }
 
-    tmp = (void *) UMEM2BLK(fence[bi_tests - 1]);
-    tmp = (void *) NXT_BLK((Block_t *) tmp);
-    ASSERT(tcs, IS_LAST_BLK(tmp));
-    sz = BLK_SZ((Block_t *) tmp);
-    testcase_printf(tcs, "Allocating leftover size = %lu\n", sz);
-    tmp = ALLOC(a, sz - ablk_hdr_sz);
-    ASSERT(tcs, tmp);
+    pad_list = 0;
+    do {
+	tmp = (void *) UMEM2BLK(link); /* last allocated */
+	tmp = (void *) NXT_BLK((Block_t *) tmp);
+	ASSERT(tcs, IS_LAST_BLK(tmp));
+	sz = BLK_SZ((Block_t *) tmp);
+	if (sz >= sbct) {
+	    residue = sz;
+	    sz = sbct - min_blk_sz;
+	    residue -= sz;
+	}
+	else {
+	    residue = 0;
+	}
+	testcase_printf(tcs, "Allocating leftover size = %lu, residue = %lu\n", sz, residue);
+	link = (Linked*) ALLOC(a, sz - ablk_hdr_sz);
+	ASSERT(tcs, link);
+	link->next = pad_list;
+	pad_list = link;
+    } while (residue);
 
     bi = FIND_BKT(a, 0);
     ASSERT(tcs, bi < 0);
@@ -135,16 +157,23 @@ testcase_run(TestCaseState_t *tcs)
 
     for (i = 0; i < bi_tests; i++) {
 	FREE(a, blk[i]);
-	FREE(a, fence[i]);
+    }
+    while (fence_list) {
+	link = fence_list;
+	fence_list = link->next;
+	FREE(a, link);
     }
 
     FREE(a, (void *) blk);
-    FREE(a, (void *) fence);
 
     bi = FIND_BKT(a, 0);
     ASSERT(tcs, bi == no_bkts - 1);
 
-    FREE(a, tmp);
+    while (pad_list) {
+	link = pad_list;
+	pad_list = link->next;
+	FREE(a, link);
+    }
 
     bi = FIND_BKT(a, 0);
     ASSERT(tcs, bi < 0);
diff --git a/erts/emulator/test/alloc_SUITE_data/coalesce.c b/erts/emulator/test/alloc_SUITE_data/coalesce.c
index 6f35d3279b..981fa6d43e 100644
--- a/erts/emulator/test/alloc_SUITE_data/coalesce.c
+++ b/erts/emulator/test/alloc_SUITE_data/coalesce.c
@@ -54,7 +54,7 @@ setup_sequence(TestCaseState_t *tcs, Allctr_t *a, Ulong bsz, int no,
 		    no, bsz);
     c = FIRST_MBC(a);
     ASSERT(tcs, !NEXT_C(c));
-    blk = MBC2FBLK(a, c);
+    blk = MBC_TO_FIRST_BLK(a, c);
     ASSERT(tcs, IS_LAST_BLK(blk));
 
     for (i = 0; i < no; i++)
@@ -266,7 +266,7 @@ testcase_name(void)
 void
 testcase_run(TestCaseState_t *tcs)
 {
-    char *argv_org[] = {"-tmmbcs1024", "-tsbct2048", "-trmbcmt100", "-tas", NULL, NULL};
+    char *argv_org[] = {"-tsmbcs511","-tmmbcs511", "-tsbct512", "-trmbcmt100", "-tas", NULL, NULL};
     char *alg[] = {"af", "gf", "bf", "aobf", "aoff", NULL};
     int i;
 
@@ -276,7 +276,7 @@ testcase_run(TestCaseState_t *tcs)
 	char *argv[sizeof(argv_org)/sizeof(argv_org[0])];
 	memcpy((void *) argv, (void *) argv_org, sizeof(argv_org));
 
-	argv[4] = alg[i];
+	argv[5] = alg[i];
 	testcase_printf(tcs, " *** Starting \"%s\" allocator *** \n", alg[i]);
 	a = START_ALC("coalesce_", 0, argv);
 	ASSERT(tcs, a);
diff --git a/erts/emulator/test/alloc_SUITE_data/mseg_clear_cache.c b/erts/emulator/test/alloc_SUITE_data/mseg_clear_cache.c
index 0277616bd0..7d5608f890 100644
--- a/erts/emulator/test/alloc_SUITE_data/mseg_clear_cache.c
+++ b/erts/emulator/test/alloc_SUITE_data/mseg_clear_cache.c
@@ -52,10 +52,10 @@ testcase_run(TestCaseState_t *tcs)
     tcs->extra = &seg[0];
 
     for (i = 0; i < MAX_SEGS; i++) {
-	seg[i].size = 1000;
+	seg[i].size = 1 << 18;
 	seg[i].ptr = MSEG_ALLOC(&seg[i].size);
 	ASSERT(tcs, seg[i].ptr);
-	ASSERT(tcs, seg[i].size >= 1000);
+	ASSERT(tcs, seg[i].size >= (1 << 18));
     }
 
     n = MSEG_NO();
diff --git a/erts/emulator/test/beam_SUITE.erl b/erts/emulator/test/beam_SUITE.erl
index 02c6e19686..3197a4c137 100644
--- a/erts/emulator/test/beam_SUITE.erl
+++ b/erts/emulator/test/beam_SUITE.erl
@@ -54,7 +54,7 @@ end_per_group(_GroupName, Config) ->
 
 %% Verify that apply(M, F, A) is really tail recursive.
 apply_last(Config) when is_list(Config) ->
-    Pid=spawn(?MODULE, applied, [self(), 10000]),
+    Pid  = spawn(?MODULE, applied, [self(), 10000]),
     Size =
 	receive
 	    {Pid, finished} ->
@@ -94,32 +94,32 @@ apply_last_bif(Config) when is_list(Config) ->
 %% Test three high register numbers in a put_list instruction
 %% (to test whether packing works properly).
 packed_registers(Config) when is_list(Config) ->
-    ?line PrivDir = ?config(priv_dir, Config),
-    ?line Mod = packed_regs,
-    ?line Name = filename:join(PrivDir, atom_to_list(Mod) ++ ".erl"),
+    PrivDir = ?config(priv_dir, Config),
+    Mod = packed_regs,
+    Name = filename:join(PrivDir, atom_to_list(Mod) ++ ".erl"),
 
     %% Generate a module which generates a list of tuples.
     %% put_list(A) -> [{A, 600}, {A, 999}, ... {A, 0}].
-    ?line Code = gen_packed_regs(600, ["-module("++atom_to_list(Mod)++").\n",
+    Code = gen_packed_regs(600, ["-module("++atom_to_list(Mod)++").\n",
 				       "-export([put_list/1]).\n",
 				       "put_list(A) ->\n["]),
-    ?line ok = file:write_file(Name, Code),
+    ok = file:write_file(Name, Code),
 
     %% Compile the module.
-    ?line io:format("Compiling: ~s\n", [Name]),
-    ?line CompRc = compile:file(Name, [{outdir, PrivDir}, report]),
-    ?line io:format("Result: ~p\n",[CompRc]),
-    ?line {ok, Mod} = CompRc,
+    io:format("Compiling: ~s\n", [Name]),
+    CompRc = compile:file(Name, [{outdir, PrivDir}, report]),
+    io:format("Result: ~p\n",[CompRc]),
+    {ok, Mod} = CompRc,
 
     %% Load it.
-    ?line io:format("Loading...\n",[]),
-    ?line LoadRc = code:load_abs(filename:join(PrivDir, atom_to_list(Mod))),
-    ?line {module,_Module} = LoadRc,
+    io:format("Loading...\n",[]),
+    LoadRc = code:load_abs(filename:join(PrivDir, atom_to_list(Mod))),
+    {module,_Module} = LoadRc,
 
     %% Call it and verify result.
-    ?line Term = {a, b},
-    ?line L = Mod:put_list(Term),
-    ?line verify_packed_regs(L, Term, 600),
+    Term = {a, b},
+    L = Mod:put_list(Term),
+    verify_packed_regs(L, Term, 600),
     ok.
 
 gen_packed_regs(0, Acc) ->
@@ -131,11 +131,11 @@ verify_packed_regs([], _, -1) -> ok;
 verify_packed_regs([{Term, N}| T], Term, N) ->
     verify_packed_regs(T, Term, N-1);
 verify_packed_regs(L, Term, N) ->
-    ?line ok = io:format("Expected [{~p, ~p}|T]; got\n~p\n", [Term, N, L]),
-    ?line test_server:fail().
+    ok = io:format("Expected [{~p, ~p}|T]; got\n~p\n", [Term, N, L]),
+    test_server:fail().
 
 buildo_mucho(Config) when is_list(Config) ->
-    ?line buildo_mucho_1(),
+    buildo_mucho_1(),
     ok.
 
 buildo_mucho_1() ->
@@ -206,20 +206,27 @@ buildo_mucho_1() ->
      {<<>>,1},{<<>>,1},{<<>>,1},{<<>>,1},{<<>>,1},{<<>>,1},{<<>>,1},{<<>>,1}].
 
 heap_sizes(Config) when is_list(Config) ->
-    ?line Sizes = erlang:system_info(heap_sizes),
-    ?line io:format("~p heap sizes\n", [length(Sizes)]),
-    ?line io:format("~p\n", [Sizes]),
+    Sizes = erlang:system_info(heap_sizes),
+    io:format("~p heap sizes\n", [length(Sizes)]),
+    io:format("~p\n", [Sizes]),
 
     %% Verify that heap sizes increase monotonically.
-    ?line Largest = lists:foldl(fun(E, P) when is_integer(P), E > P -> E;
+    Largest = lists:foldl(fun(E, P) when is_integer(P), E > P -> E;
 				   (E, []) -> E
 				end, [], Sizes),
 
-    %% Verify that the largest heap size consists of 31 or 63 bits.
-    ?line
-	case Largest bsr (erlang:system_info(wordsize)*8-2) of
-	    R when R > 0 -> ok
-	end,
+    %% Verify that the largest heap size consists of
+    %%  - 31 bits of bytes on 32 bits arch
+    %%  - atleast 52 bits of bytes (48 is the maximum virtual address)
+    %%    and at the most 63 bits on 64 bit archs
+    %% heap sizes are in words
+    case erlang:system_info(wordsize) of
+	8 ->
+	    0    = (Largest*8) bsr 63,
+	    true = (Largest*8) > (1 bsl 52);
+	4 ->
+	    1 = (Largest*4) bsr 31
+    end,
     ok.
 
 %% Thanks to Igor Goryachev.
@@ -302,10 +309,10 @@ b() ->
     end.
 
 fconv(Config) when is_list(Config) ->
-    ?line do_fconv(atom),
-    ?line do_fconv(nil),
-    ?line do_fconv(tuple_literal),
-    ?line 3.0 = do_fconv(1.0, 2.0),
+    do_fconv(atom),
+    do_fconv(nil),
+    do_fconv(tuple_literal),
+    3.0 = do_fconv(1.0, 2.0),
     ok.
 
 do_fconv(Type) ->
@@ -325,9 +332,9 @@ do_fconv(tuple_literal, Float) when is_float(Float) ->
     Float + {a,b}.
 
 select_val(Config) when is_list(Config) ->
-    ?line zero = do_select_val(0),
-    ?line big = do_select_val(1 bsl 64),
-    ?line integer = do_select_val(42),
+    zero = do_select_val(0),
+    big = do_select_val(1 bsl 64),
+    integer = do_select_val(42),
     ok.
 
 do_select_val(X) ->
diff --git a/erts/emulator/test/process_SUITE.erl b/erts/emulator/test/process_SUITE.erl
index 898535693f..11dd88413f 100644
--- a/erts/emulator/test/process_SUITE.erl
+++ b/erts/emulator/test/process_SUITE.erl
@@ -118,15 +118,15 @@ fun_spawn(Fun) ->
 %% (unclear if this test case will actually prove anything on
 %% a modern computer with lots of memory).
 spawn_with_binaries(Config) when is_list(Config) ->
-    ?line L = lists:duplicate(2048, 42),
-    ?line TwoMeg = lists:duplicate(1024, L),
-    ?line Fun = fun() -> spawn(?MODULE, binary_owner, [list_to_binary(TwoMeg)]),
+    L = lists:duplicate(2048, 42),
+    TwoMeg = lists:duplicate(1024, L),
+    Fun = fun() -> spawn(?MODULE, binary_owner, [list_to_binary(TwoMeg)]),
 			 receive after 1 -> ok end end,
-    ?line Iter = case test_server:purify_is_running() of
+    Iter = case test_server:purify_is_running() of
 		     true -> 10;
 		     false -> 150
 		 end,
-    ?line test_server:do_times(Iter, Fun),
+    test_server:do_times(Iter, Fun),
     ok.
 
 binary_owner(Bin) when is_binary(Bin) ->
@@ -134,87 +134,87 @@ binary_owner(Bin) when is_binary(Bin) ->
 
 %% Tests exit/1 with a big message.
 t_exit_1(Config) when is_list(Config) ->
-    ?line start_spawner(),
-    ?line Dog = test_server:timetrap(test_server:seconds(20)),
-    ?line process_flag(trap_exit, true),
-    ?line test_server:do_times(10, fun t_exit_1/0),
-    ?line test_server:timetrap_cancel(Dog),
-    ?line stop_spawner(),
+    start_spawner(),
+    Dog = test_server:timetrap(test_server:seconds(20)),
+    process_flag(trap_exit, true),
+    test_server:do_times(10, fun t_exit_1/0),
+    test_server:timetrap_cancel(Dog),
+    stop_spawner(),
     ok.
 
 t_exit_1() ->
-    ?line Pid = fun_spawn(fun() -> exit(kb_128()) end),
-    ?line Garbage = kb_128(),
-    ?line receive
+    Pid = fun_spawn(fun() -> exit(kb_128()) end),
+    Garbage = kb_128(),
+    receive
 	      {'EXIT', Pid, Garbage} -> ok
 	  end.
 
 
 %% Tests exit/2 with a lot of data in the exit message.
 t_exit_2_other(Config) when is_list(Config) ->
-    ?line start_spawner(),
-    ?line Dog = test_server:timetrap(test_server:seconds(20)),
-    ?line process_flag(trap_exit, true),
-    ?line test_server:do_times(10, fun t_exit_2_other/0),
-    ?line test_server:timetrap_cancel(Dog),
-    ?line stop_spawner(),
+    start_spawner(),
+    Dog = test_server:timetrap(test_server:seconds(20)),
+    process_flag(trap_exit, true),
+    test_server:do_times(10, fun t_exit_2_other/0),
+    test_server:timetrap_cancel(Dog),
+    stop_spawner(),
     ok.
 
 t_exit_2_other() ->
-    ?line Pid = fun_spawn(fun() -> receive x -> ok end end),
-    ?line Garbage = kb_128(),
-    ?line exit(Pid, Garbage),
-    ?line receive
+    Pid = fun_spawn(fun() -> receive x -> ok end end),
+    Garbage = kb_128(),
+    exit(Pid, Garbage),
+    receive
 	      {'EXIT', Pid, Garbage} -> ok
 	  end.
 
 %% Tests that exit(Pid, normal) does not kill another process.;
 t_exit_2_other_normal(Config) when is_list(Config) ->
-    ?line Dog = test_server:timetrap(test_server:seconds(20)),
-    ?line process_flag(trap_exit, true),
-    ?line Pid = fun_spawn(fun() -> receive x -> ok end end),
-    ?line exit(Pid, normal),
-    ?line receive
+    Dog = test_server:timetrap(test_server:seconds(20)),
+    process_flag(trap_exit, true),
+    Pid = fun_spawn(fun() -> receive x -> ok end end),
+    exit(Pid, normal),
+    receive
 	      {'EXIT', Pid, Reason} ->
-		  ?line test_server:fail({process_died, Reason})
+		  test_server:fail({process_died, Reason})
 	  after 1000 ->
 		  ok
 	  end,
-    ?line case process_info(Pid) of
+    case process_info(Pid) of
 	      undefined ->
 		  test_server:fail(process_died_on_normal);
 	      List when is_list(List) ->
 		  ok
 	  end,
     exit(Pid, kill),
-    ?line test_server:timetrap_cancel(Dog),
+    test_server:timetrap_cancel(Dog),
     ok.
 
 %% Tests that we can trap an exit message sent with exit/2 from
 %% the same process.
 self_exit(Config) when is_list(Config) ->
-    ?line start_spawner(),
-    ?line Dog = test_server:timetrap(test_server:seconds(10)),
-    ?line process_flag(trap_exit, true),
-    ?line test_server:do_times(200, fun self_exit/0),
-    ?line test_server:timetrap_cancel(Dog),
-    ?line stop_spawner(),
+    start_spawner(),
+    Dog = test_server:timetrap(test_server:seconds(10)),
+    process_flag(trap_exit, true),
+    test_server:do_times(200, fun self_exit/0),
+    test_server:timetrap_cancel(Dog),
+    stop_spawner(),
     ok.
 
 self_exit() ->
-    ?line Garbage = eight_kb(),
-    ?line P = self(),
-    ?line true = exit(P, Garbage),
-    ?line receive
+    Garbage = eight_kb(),
+    P = self(),
+    true = exit(P, Garbage),
+    receive
 	      {'EXIT', P, Garbage} -> ok
 	  end.
 
 %% Tests exit(self(), normal) is equivalent to exit(normal) for a process
 %% that doesn't trap exits.
 normal_suicide_exit(Config) when is_list(Config) ->
-    ?line process_flag(trap_exit, true),
-    ?line Pid = fun_spawn(fun() -> exit(self(), normal) end),
-    ?line receive
+    process_flag(trap_exit, true),
+    Pid = fun_spawn(fun() -> exit(self(), normal) end),
+    receive
 	      {'EXIT', Pid, normal} -> ok;
 	      Other -> test_server:fail({bad_message, Other})
 	  end.
@@ -222,19 +222,19 @@ normal_suicide_exit(Config) when is_list(Config) ->
 %% Tests exit(self(), Term) is equivalent to exit(Term) for a process
 %% that doesn't trap exits.";
 abnormal_suicide_exit(Config) when is_list(Config) ->
-    ?line Garbage = eight_kb(),
-    ?line process_flag(trap_exit, true),
-    ?line Pid = fun_spawn(fun() -> exit(self(), Garbage) end),
-    ?line receive
+    Garbage = eight_kb(),
+    process_flag(trap_exit, true),
+    Pid = fun_spawn(fun() -> exit(self(), Garbage) end),
+    receive
 	      {'EXIT', Pid, Garbage} -> ok;
 	      Other -> test_server:fail({bad_message, Other})
 	  end.
 
 %% Tests that exit(self(), die) cannot be catched.
 t_exit_2_catch(Config) when is_list(Config) ->
-    ?line process_flag(trap_exit, true),
-    ?line Pid = fun_spawn(fun() -> catch exit(self(), die) end),
-    ?line receive
+    process_flag(trap_exit, true),
+    Pid = fun_spawn(fun() -> catch exit(self(), die) end),
+    receive
 	      {'EXIT', Pid, normal} ->
 		  test_server:fail(catch_worked);
 	      {'EXIT', Pid, die} ->
@@ -246,29 +246,29 @@ t_exit_2_catch(Config) when is_list(Config) ->
 %% Tests trapping of an 'EXIT' message generated by a bad argument to
 %% the abs/1 bif.  The 'EXIT' message will intentionally be very big.
 trap_exit_badarg(Config) when is_list(Config) ->
-    ?line start_spawner(),
-    ?line Dog = test_server:timetrap(test_server:seconds(10)),
-    ?line process_flag(trap_exit, true),
-    ?line test_server:do_times(10, fun trap_exit_badarg/0),
-    ?line test_server:timetrap_cancel(Dog),
-    ?line stop_spawner(),
+    start_spawner(),
+    Dog = test_server:timetrap(test_server:seconds(10)),
+    process_flag(trap_exit, true),
+    test_server:do_times(10, fun trap_exit_badarg/0),
+    test_server:timetrap_cancel(Dog),
+    stop_spawner(),
     ok.
 
 trap_exit_badarg() ->
-    ?line Pid = fun_spawn(fun() -> bad_guy(kb_128()) end),
-    ?line Garbage = kb_128(),
-    ?line receive
+    Pid = fun_spawn(fun() -> bad_guy(kb_128()) end),
+    Garbage = kb_128(),
+    receive
 	      {'EXIT',Pid,{badarg,[{erlang,abs,[Garbage],Loc1},
 				   {?MODULE,bad_guy,1,Loc2}|_]}}
 	      when is_list(Loc1), is_list(Loc2) ->
 		  ok;
 	      Other ->
-		  ?line ok = io:format("Bad EXIT message: ~P", [Other, 30]),
-		  ?line test_server:fail(bad_exit_message)
+		  ok = io:format("Bad EXIT message: ~P", [Other, 30]),
+		  test_server:fail(bad_exit_message)
 	  end.
 
 bad_guy(Arg) ->
-    ?line abs(Arg).
+    abs(Arg).
 
 
 kb_128() ->
@@ -281,11 +281,11 @@ kb_128() ->
 
 eight_kb() ->
     B64 = lists:seq(1, 64),
-    ?line B512 = {<<1>>,B64,<<2,3>>,B64,make_unaligned_sub_binary(<<4,5,6,7,8,9>>),
+    B512 = {<<1>>,B64,<<2,3>>,B64,make_unaligned_sub_binary(<<4,5,6,7,8,9>>),
 		  B64,make_sub_binary([1,2,3,4,5,6]),
 		  B64,make_sub_binary(lists:seq(1, ?heap_binary_size+1)),
 		  B64,B64,B64,B64,big_binary()},
-    ?line lists:duplicate(8, {B512,B512}).
+    lists:duplicate(8, {B512,B512}).
 
 big_binary() ->
     big_binary(10, [42]).
@@ -296,19 +296,19 @@ big_binary(N, Acc) ->
 
 %% Test receiving an EXIT message when spawning a BIF with bad arguments.
 trap_exit_badarg_in_bif(Config) when is_list(Config) ->
-    ?line Dog = test_server:timetrap(test_server:seconds(10)),
-    ?line process_flag(trap_exit, true),
-    ?line test_server:do_times(10, fun trap_exit_badarg_bif/0),
-    ?line test_server:timetrap_cancel(Dog),
+    Dog = test_server:timetrap(test_server:seconds(10)),
+    process_flag(trap_exit, true),
+    test_server:do_times(10, fun trap_exit_badarg_bif/0),
+    test_server:timetrap_cancel(Dog),
     ok.
     
 trap_exit_badarg_bif() ->
-    ?line Pid = spawn_link(erlang, node, [1]),
-    ?line receive
+    Pid = spawn_link(erlang, node, [1]),
+    receive
 	      {'EXIT', Pid, {badarg, _}} ->
 		  ok;
 	      Other ->
-		  ?line test_server:fail({unexpected, Other})
+		  test_server:fail({unexpected, Other})
 	  end.
 
 %% The following sequences of events have crasched Beam.
@@ -321,27 +321,27 @@ trap_exit_badarg_bif() ->
 %% 3) The process will crash the next time it executes 'receive'.
 
 exit_and_timeout(Config) when is_list(Config) ->
-    ?line Dog = test_server:timetrap(test_server:seconds(20)),
+    Dog = test_server:timetrap(test_server:seconds(20)),
 
-    ?line process_flag(trap_exit, true),
-    ?line Parent = self(),
-    ?line Low = fun_spawn(fun() -> eat_low(Parent) end),
-    ?line High = fun_spawn(fun() -> eat_high(Low) end),
-    ?line eat_wait_for(Low, High),
+    process_flag(trap_exit, true),
+    Parent = self(),
+    Low = fun_spawn(fun() -> eat_low(Parent) end),
+    High = fun_spawn(fun() -> eat_high(Low) end),
+    eat_wait_for(Low, High),
 
-    ?line test_server:timetrap_cancel(Dog),
+    test_server:timetrap_cancel(Dog),
     ok.
 
 
 eat_wait_for(Low, High) ->
-    ?line receive
-	      {'EXIT', Low, {you, are, dead}} ->
-		  ok;
-	      {'EXIT', High, normal} ->
-		  eat_wait_for(Low, High);
-	      Other ->
-		  test_server:fail({bad_message, Other})
-	  end.
+    receive
+	{'EXIT', Low, {you, are, dead}} ->
+	    ok;
+	{'EXIT', High, normal} ->
+	    eat_wait_for(Low, High);
+	Other ->
+	    test_server:fail({bad_message, Other})
+    end.
 
 eat_low(_Parent) ->
     receive
@@ -374,27 +374,27 @@ loop(_, _) ->
 %% Tries to send two different exit messages to a process.
 %% (The second one should be ignored.)
 exit_twice(Config) when is_list(Config) ->
-    ?line Dog = test_server:timetrap(test_server:seconds(20)),
+    Dog = test_server:timetrap(test_server:seconds(20)),
 
-    ?line process_flag(trap_exit, true),
-    ?line Low = fun_spawn(fun etwice_low/0),
-    ?line High = fun_spawn(fun() -> etwice_high(Low) end),
-    ?line etwice_wait_for(Low, High),
+    process_flag(trap_exit, true),
+    Low = fun_spawn(fun etwice_low/0),
+    High = fun_spawn(fun() -> etwice_high(Low) end),
+    etwice_wait_for(Low, High),
 
-    ?line test_server:timetrap_cancel(Dog),
+    test_server:timetrap_cancel(Dog),
     ok.
 
 etwice_wait_for(Low, High) ->
-    ?line receive
-	      {'EXIT', Low, first} ->
-		  ok;
-	      {'EXIT', Low, Other} ->
-		  test_server:fail({wrong_exit_reason, Other});
-	      {'EXIT', High, normal} ->
-		  etwice_wait_for(Low, High);
-	      Other ->
-		  test_server:fail({bad_message, Other})
-	  end.
+    receive
+	{'EXIT', Low, first} ->
+	    ok;
+	{'EXIT', Low, Other} ->
+	    test_server:fail({wrong_exit_reason, Other});
+	{'EXIT', High, normal} ->
+	    etwice_wait_for(Low, High);
+	Other ->
+	    test_server:fail({bad_message, Other})
+    end.
 
 etwice_low() ->
     etwice_low().
@@ -406,15 +406,15 @@ etwice_high(Low) ->
 
 %% Tests the process_info/2 BIF.
 t_process_info(Config) when is_list(Config) ->
-    ?line [] = process_info(self(), registered_name),
-    ?line register(my_name, self()),
-    ?line {registered_name, my_name} = process_info(self(), registered_name),
-    ?line {status, running} = process_info(self(), status),
-    ?line {min_heap_size, 233} = process_info(self(), min_heap_size),
-    ?line {min_bin_vheap_size, 46368} = process_info(self(), min_bin_vheap_size),
-    ?line {current_function,{?MODULE,t_process_info,1}} =
+    [] = process_info(self(), registered_name),
+    register(my_name, self()),
+    {registered_name, my_name} = process_info(self(), registered_name),
+    {status, running} = process_info(self(), status),
+    {min_heap_size, 233} = process_info(self(), min_heap_size),
+    {min_bin_vheap_size,46422} = process_info(self(), min_bin_vheap_size),
+    {current_function,{?MODULE,t_process_info,1}} =
 	process_info(self(), current_function),
-    ?line {current_function,{?MODULE,t_process_info,1}} =
+    {current_function,{?MODULE,t_process_info,1}} =
 	apply(erlang, process_info, [self(),current_function]),
 
     %% current_location and current_stacktrace
@@ -425,9 +425,9 @@ t_process_info(Config) when is_list(Config) ->
     verify_loc(Line2, Res2),
     pi_stacktrace([{?MODULE,t_process_info,1,?LINE}]),
 
-    ?line Gleader = group_leader(),
-    ?line {group_leader, Gleader} = process_info(self(), group_leader),
-    ?line {'EXIT',{badarg,_Info}} = (catch process_info('not_a_pid')),
+    Gleader = group_leader(),
+    {group_leader, Gleader} = process_info(self(), group_leader),
+    {'EXIT',{badarg,_Info}} = (catch process_info('not_a_pid')),
     ok.
 
 pi_stacktrace(Expected0) ->
@@ -509,50 +509,50 @@ process_info_looper(Parent) ->
 %% Tests the process_info/1 BIF on another process with messages.
 process_info_other_msg(Config) when is_list(Config) ->
     Self = self(),
-    ?line Pid = spawn_link(fun() -> other_process(Self) end),
+    Pid = spawn_link(fun() -> other_process(Self) end),
     receive
 	{go_ahead,Pid} -> ok
     end,
 
-    ?line Own = {my,own,message},
+    Own = {my,own,message},
 
-    ?line {messages,[Own]} = process_info(Pid, messages),
+    {messages,[Own]} = process_info(Pid, messages),
     
-    ?line Garbage = kb_128(),
-    ?line MsgA = {a,Garbage},
-    ?line MsgB = {b,Garbage},
-    ?line MsgC = {c,Garbage},
-    ?line MsgD = {d,Garbage},
-    ?line MsgE = {e,Garbage},
-
-    ?line Pid ! MsgA,
-    ?line {messages,[Own,MsgA]} = process_info(Pid, messages),
-    ?line Pid ! MsgB,
-    ?line {messages,[Own,MsgA,MsgB]} = process_info(Pid, messages),
-    ?line Pid ! MsgC,
-    ?line {messages,[Own,MsgA,MsgB,MsgC]} = process_info(Pid, messages),
-    ?line Pid ! MsgD,
-    ?line {messages,[Own,MsgA,MsgB,MsgC,MsgD]} = process_info(Pid, messages),
-    ?line Pid ! MsgE,
-    ?line {messages,[Own,MsgA,MsgB,MsgC,MsgD,MsgE]=All} = process_info(Pid, messages),
-    ?line {memory,BytesOther} = process_info(Pid, memory),
-    ?line {memory,BytesSelf} = process_info(self(), memory),
+    Garbage = kb_128(),
+    MsgA = {a,Garbage},
+    MsgB = {b,Garbage},
+    MsgC = {c,Garbage},
+    MsgD = {d,Garbage},
+    MsgE = {e,Garbage},
+
+    Pid ! MsgA,
+    {messages,[Own,MsgA]} = process_info(Pid, messages),
+    Pid ! MsgB,
+    {messages,[Own,MsgA,MsgB]} = process_info(Pid, messages),
+    Pid ! MsgC,
+    {messages,[Own,MsgA,MsgB,MsgC]} = process_info(Pid, messages),
+    Pid ! MsgD,
+    {messages,[Own,MsgA,MsgB,MsgC,MsgD]} = process_info(Pid, messages),
+    Pid ! MsgE,
+    {messages,[Own,MsgA,MsgB,MsgC,MsgD,MsgE]=All} = process_info(Pid, messages),
+    {memory,BytesOther} = process_info(Pid, memory),
+    {memory,BytesSelf} = process_info(self(), memory),
 
     io:format("Memory ~p: ~p\n", [Pid,BytesOther]),
     io:format("Memory ~p (self): ~p\n", [self(),BytesSelf]),
 
     [Own,MsgA,MsgB,MsgC,MsgD,MsgE] = All,
 
-    ?line Pid ! {self(),empty},
-    ?line receive
+    Pid ! {self(),empty},
+    receive
 	      empty -> ok
 	  end,
-    ?line {messages,[]} = process_info(Pid, messages),
+    {messages,[]} = process_info(Pid, messages),
 
-    ?line {min_heap_size, 233} = process_info(Pid, min_heap_size),
-    ?line {min_bin_vheap_size, 46368} = process_info(Pid, min_bin_vheap_size),
+    {min_heap_size, 233} = process_info(Pid, min_heap_size),
+    {min_bin_vheap_size, 46422} = process_info(Pid, min_bin_vheap_size),
 
-    ?line Pid ! stop,
+    Pid ! stop,
     ok.
 
 process_info_other_dist_msg(Config) when is_list(Config) ->
@@ -560,52 +560,51 @@ process_info_other_dist_msg(Config) when is_list(Config) ->
     %% Check that process_info can handle messages that have not been
     %% decoded yet.
     %%
-    ?line {ok, Node} = start_node(Config),
-    ?line Self = self(),
-    ?line Pid = spawn_link(fun() -> other_process(Self) end),
-    ?line receive {go_ahead,Pid} -> ok end,
+    {ok, Node} = start_node(Config),
+    Self = self(),
+    Pid = spawn_link(fun() -> other_process(Self) end),
+    receive {go_ahead,Pid} -> ok end,
 
-    ?line Own = {my,own,message},
+    Own = {my,own,message},
 
-    ?line {messages,[Own]} = process_info(Pid, messages),
-    ?line Garbage = kb_128(),
-    ?line MsgA = {a,self(),Garbage},
-    ?line MsgB = {b,self(),Garbage},
-    ?line MsgC = {c,self(),Garbage},
-    ?line MsgD = {d,self(),Garbage},
-    ?line MsgE = {e,self(),Garbage},
+    {messages,[Own]} = process_info(Pid, messages),
+    Garbage = kb_128(),
+    MsgA = {a,self(),Garbage},
+    MsgB = {b,self(),Garbage},
+    MsgC = {c,self(),Garbage},
+    MsgD = {d,self(),Garbage},
+    MsgE = {e,self(),Garbage},
 
     %% We don't want the other process to decode messages itself
     %% therefore we suspend it.
-    ?line true =  erlang:suspend_process(Pid),
-    ?line spawn_link(Node, fun () ->
-				   Pid ! MsgA,
-				   Pid ! MsgB,
-				   Pid ! MsgC,
-				   Self ! check_abc
-			   end),
-    ?line receive check_abc -> ok end,
-    ?line [{status,suspended},
-	   {messages,[Own,MsgA,MsgB,MsgC]},
-	   {status,suspended}]= process_info(Pid, [status,messages,status]),
-    ?line spawn_link(Node, fun () ->
-				   Pid ! MsgD,
-				   Pid ! MsgE,
-				   Self ! check_de
-			   end),
-    ?line receive check_de -> ok end,
-    ?line {messages,[Own,MsgA,MsgB,MsgC,MsgD,MsgE]=All}
-	= process_info(Pid, messages),
-    ?line true = erlang:resume_process(Pid),
-    ?line Pid ! {self(), get_all_messages},
-    ?line receive
+    true =  erlang:suspend_process(Pid),
+    spawn_link(Node, fun () ->
+		Pid  ! MsgA,
+		Pid  ! MsgB,
+		Pid  ! MsgC,
+		Self ! check_abc
+	end),
+    receive check_abc -> ok end,
+    [{status,suspended},
+	{messages,[Own,MsgA,MsgB,MsgC]},
+	{status,suspended}]= process_info(Pid, [status,messages,status]),
+    spawn_link(Node, fun () ->
+		Pid  ! MsgD,
+		Pid  ! MsgE,
+		Self ! check_de
+	end),
+    receive check_de -> ok end,
+    {messages,[Own,MsgA,MsgB,MsgC,MsgD,MsgE]=All} = process_info(Pid, messages),
+    true = erlang:resume_process(Pid),
+    Pid ! {self(), get_all_messages},
+    receive
 	      {all_messages, AllMsgs} ->
-		  ?line All = AllMsgs
+		  All = AllMsgs
 	  end,
-    ?line {messages,[]} = process_info(Pid, messages),
-    ?line Pid ! stop,
-    ?line stop_node(Node),
-    ?line ok.
+    {messages,[]} = process_info(Pid, messages),
+    Pid ! stop,
+    stop_node(Node),
+    ok.
     
 
 other_process(Parent) ->
@@ -652,38 +651,36 @@ process_info_2_list(doc) ->
 process_info_2_list(suite) ->
     [];
 process_info_2_list(Config) when is_list(Config) ->
-    ?line Proc = spawn(fun () ->
-			       receive after infinity -> ok end end),
+    Proc = spawn(fun () -> receive after infinity -> ok end end),
     register(process_SUITE_process_info_2_list1, self()),
     register(process_SUITE_process_info_2_list2, Proc),
-    ?line erts_debug:set_internal_state(available_internal_state,true),
-    ?line AllArgs = erts_debug:get_internal_state(process_info_args),
-    ?line A1 = lists:sort(AllArgs) ++ [status] ++ lists:reverse(AllArgs),
+    erts_debug:set_internal_state(available_internal_state,true),
+    AllArgs = erts_debug:get_internal_state(process_info_args),
+    A1 = lists:sort(AllArgs) ++ [status] ++ lists:reverse(AllArgs),
 
     %% Verify that argument is accepted as single atom
-    ?line lists:foreach(fun (A) ->
-				?line {A, _} = process_info(Proc, A),
-				?line {A, _} = process_info(self(), A)
-			end,
-			A1),
+    lists:foreach(fun (A) ->
+		{A, _} = process_info(Proc, A),
+		{A, _} = process_info(self(), A)
+	end, A1),
 
     %% Verify that order is preserved
-    ?line ok = chk_pi_order(process_info(self(), A1), A1),
-    ?line ok = chk_pi_order(process_info(Proc, A1), A1),
+    ok = chk_pi_order(process_info(self(), A1), A1),
+    ok = chk_pi_order(process_info(Proc, A1), A1),
 
     %% Small arg list
-    ?line A2 = [status, stack_size, trap_exit, priority],
-    ?line [{status, _}, {stack_size, _}, {trap_exit, _}, {priority, _}]
+    A2 = [status, stack_size, trap_exit, priority],
+    [{status, _}, {stack_size, _}, {trap_exit, _}, {priority, _}]
 	= process_info(Proc, A2),
-    ?line [{status, _}, {stack_size, _}, {trap_exit, _}, {priority, _}]
+    [{status, _}, {stack_size, _}, {trap_exit, _}, {priority, _}]
 	= process_info(self(), A2),
 
     %% Huge arg list (note values are shared)
-    ?line A3 = lists:duplicate(5000,backtrace),
-    ?line V3 = process_info(Proc, A3),
-    ?line 5000 = length(V3),
-    ?line lists:foreach(fun ({backtrace, _}) -> ok end, V3),
-    ?line ok.
+    A3 = lists:duplicate(5000,backtrace),
+    V3 = process_info(Proc, A3),
+    5000 = length(V3),
+    lists:foreach(fun ({backtrace, _}) -> ok end, V3),
+    ok.
     
 process_info_lock_reschedule(doc) ->
     [];
@@ -692,43 +689,37 @@ process_info_lock_reschedule(suite) ->
 process_info_lock_reschedule(Config) when is_list(Config) ->
     %% We need a process that is running and an item that requires
     %% process_info to take the main process lock.
-    ?line Target1 = spawn_link(fun tok_loop/0),
-    ?line Name1 = process_info_lock_reschedule_running,
-    ?line register(Name1, Target1),
-    ?line Target2 = spawn_link(fun () -> receive after infinity -> ok end end),
-    ?line Name2 = process_info_lock_reschedule_waiting,
-    ?line register(Name2, Target2),
-    ?line PI = fun(_) ->
-		       ?line erlang:yield(),
-		       ?line [{registered_name, Name1}]
-			   = process_info(Target1, [registered_name]),
-		       ?line [{registered_name, Name2}]
-			   = process_info(Target2, [registered_name]),
-		       ?line erlang:yield(),
-		       ?line {registered_name, Name1}
-			   = process_info(Target1, registered_name),
-		       ?line {registered_name, Name2}
-			   = process_info(Target2, registered_name),
-		       ?line erlang:yield(),
-		       ?line [{registered_name, Name1}| _]
-			   = process_info(Target1),
-		       ?line [{registered_name, Name2}| _]
-			   = process_info(Target2)
-	       end,
-    ?line lists:foreach(PI, lists:seq(1,1000)),
+    Target1 = spawn_link(fun tok_loop/0),
+    Name1 = process_info_lock_reschedule_running,
+    register(Name1, Target1),
+    Target2 = spawn_link(fun () -> receive after infinity -> ok end end),
+    Name2 = process_info_lock_reschedule_waiting,
+    register(Name2, Target2),
+    PI = fun(_) ->
+	    erlang:yield(),
+	    [{registered_name, Name1}] = process_info(Target1, [registered_name]),
+	    [{registered_name, Name2}] = process_info(Target2, [registered_name]),
+	    erlang:yield(),
+	    {registered_name, Name1} = process_info(Target1, registered_name),
+	    {registered_name, Name2} = process_info(Target2, registered_name),
+	    erlang:yield(),
+	    [{registered_name, Name1}| _] = process_info(Target1),
+	    [{registered_name, Name2}| _] = process_info(Target2)
+    end,
+    lists:foreach(PI, lists:seq(1,1000)),
     %% Make sure Target1 still is willing to "tok loop"
-    ?line case process_info(Target1, status) of
-	      {status, OkStatus} when OkStatus == runnable;
-				      OkStatus == running;
-				      OkStatus == garbage_collecting ->
-		  ?line unlink(Target1),
-		  ?line unlink(Target2),
-		  ?line exit(Target1, bang),
-		  ?line exit(Target2, bang),
-		  ?line OkStatus;
-	      {status, BadStatus} ->
-		  ?line ?t:fail(BadStatus)
-	  end.
+    case process_info(Target1, status) of
+	{status, OkStatus} when OkStatus == runnable;
+				OkStatus == running;
+				OkStatus == garbage_collecting ->
+	    unlink(Target1),
+	    unlink(Target2),
+	    exit(Target1, bang),
+	    exit(Target2, bang),
+	    OkStatus;
+	{status, BadStatus} ->
+	    ?t:fail(BadStatus)
+    end.
 
 pi_loop(_Name, _Pid, 0) ->
     ok;
@@ -741,50 +732,50 @@ process_info_lock_reschedule2(doc) ->
 process_info_lock_reschedule2(suite) ->
     [];
 process_info_lock_reschedule2(Config) when is_list(Config) ->
-    ?line Parent = self(),
-    ?line Fun = fun () ->
-			receive {go, Name, Pid} -> ok end,
-			pi_loop(Name, Pid, 10000),
-			Parent ! {done, self()},
-			receive after infinity -> ok end
-		end,
-    ?line P1 = spawn_link(Fun),
-    ?line N1 = process_info_lock_reschedule2_1,
-    ?line true = register(N1, P1),
-    ?line P2 = spawn_link(Fun),
-    ?line N2 = process_info_lock_reschedule2_2,
-    ?line true = register(N2, P2),
-    ?line P3 = spawn_link(Fun),
-    ?line N3 = process_info_lock_reschedule2_3,
-    ?line true = register(N3, P3),
-    ?line P4 = spawn_link(Fun),
-    ?line N4 = process_info_lock_reschedule2_4,
-    ?line true = register(N4, P4),
-    ?line P5 = spawn_link(Fun),
-    ?line N5 = process_info_lock_reschedule2_5,
-    ?line true = register(N5, P5),
-    ?line P6 = spawn_link(Fun),
-    ?line N6 = process_info_lock_reschedule2_6,
-    ?line true = register(N6, P6),
-    ?line P1 ! {go, N2, P2},
-    ?line P2 ! {go, N1, P1},
-    ?line P3 ! {go, N1, P1},
-    ?line P4 ! {go, N1, P1},
-    ?line P5 ! {go, N6, P6},
-    ?line P6 ! {go, N5, P5},
-    ?line receive {done, P1} -> ok end,
-    ?line receive {done, P2} -> ok end,
-    ?line receive {done, P3} -> ok end,
-    ?line receive {done, P4} -> ok end,
-    ?line receive {done, P5} -> ok end,
-    ?line receive {done, P6} -> ok end,
-    ?line unlink(P1), exit(P1, bang),
-    ?line unlink(P2), exit(P2, bang),
-    ?line unlink(P3), exit(P3, bang),
-    ?line unlink(P4), exit(P4, bang),
-    ?line unlink(P5), exit(P5, bang),
-    ?line unlink(P6), exit(P6, bang),
-    ?line ok.
+    Parent = self(),
+    Fun = fun () ->
+	    receive {go, Name, Pid} -> ok end,
+	    pi_loop(Name, Pid, 10000),
+	    Parent ! {done, self()},
+	    receive after infinity -> ok end
+    end,
+    P1 = spawn_link(Fun),
+    N1 = process_info_lock_reschedule2_1,
+    true = register(N1, P1),
+    P2 = spawn_link(Fun),
+    N2 = process_info_lock_reschedule2_2,
+    true = register(N2, P2),
+    P3 = spawn_link(Fun),
+    N3 = process_info_lock_reschedule2_3,
+    true = register(N3, P3),
+    P4 = spawn_link(Fun),
+    N4 = process_info_lock_reschedule2_4,
+    true = register(N4, P4),
+    P5 = spawn_link(Fun),
+    N5 = process_info_lock_reschedule2_5,
+    true = register(N5, P5),
+    P6 = spawn_link(Fun),
+    N6 = process_info_lock_reschedule2_6,
+    true = register(N6, P6),
+    P1 ! {go, N2, P2},
+    P2 ! {go, N1, P1},
+    P3 ! {go, N1, P1},
+    P4 ! {go, N1, P1},
+    P5 ! {go, N6, P6},
+    P6 ! {go, N5, P5},
+    receive {done, P1} -> ok end,
+    receive {done, P2} -> ok end,
+    receive {done, P3} -> ok end,
+    receive {done, P4} -> ok end,
+    receive {done, P5} -> ok end,
+    receive {done, P6} -> ok end,
+    unlink(P1), exit(P1, bang),
+    unlink(P2), exit(P2, bang),
+    unlink(P3), exit(P3, bang),
+    unlink(P4), exit(P4, bang),
+    unlink(P5), exit(P5, bang),
+    unlink(P6), exit(P6, bang),
+    ok.
 
 many_args(0,_B,_C,_D,_E,_F,_G,_H,_I,_J) ->
     ok;
@@ -802,120 +793,115 @@ process_info_lock_reschedule3(suite) ->
 process_info_lock_reschedule3(Config) when is_list(Config) ->
     %% We need a process that is running and an item that requires
     %% process_info to take the main process lock.
-    ?line Target1 = spawn_link(fun tok_loop/0),
-    ?line Name1 = process_info_lock_reschedule_running,
-    ?line register(Name1, Target1),
-    ?line Target2 = spawn_link(fun () -> receive after infinity -> ok end end),
-    ?line Name2 = process_info_lock_reschedule_waiting,
-    ?line register(Name2, Target2),
-    ?line PI = fun(N) ->
-		       case N rem 10 of
-			   0 -> erlang:yield();
-			   _ -> ok
-		       end,
-		       ?line do_pi_msg_len({proc, Target1},
-					   {arg, message_queue_len})
-	       end,
-    ?line many_args(100000,1,2,3,4,5,6,7,8,9),
-    ?line lists:foreach(PI, lists:seq(1,1000000)),
+    Target1 = spawn_link(fun tok_loop/0),
+    Name1 = process_info_lock_reschedule_running,
+    register(Name1, Target1),
+    Target2 = spawn_link(fun () -> receive after infinity -> ok end end),
+    Name2 = process_info_lock_reschedule_waiting,
+    register(Name2, Target2),
+    PI = fun(N) ->
+	    case N rem 10 of
+		0 -> erlang:yield();
+		_ -> ok
+	    end,
+	    do_pi_msg_len({proc, Target1},
+		{arg, message_queue_len})
+    end,
+    many_args(100000,1,2,3,4,5,6,7,8,9),
+    lists:foreach(PI, lists:seq(1,1000000)),
     %% Make sure Target1 still is willing to "tok loop"
-    ?line case process_info(Target1, status) of
+    case process_info(Target1, status) of
 	      {status, OkStatus} when OkStatus == runnable;
 				      OkStatus == running;
 				      OkStatus == garbage_collecting ->
-		  ?line unlink(Target1),
-		  ?line unlink(Target2),
-		  ?line exit(Target1, bang),
-		  ?line exit(Target2, bang),
-		  ?line OkStatus;
+		  unlink(Target1),
+		  unlink(Target2),
+		  exit(Target1, bang),
+		  exit(Target2, bang),
+		  OkStatus;
 	      {status, BadStatus} ->
-		  ?line ?t:fail(BadStatus)
+		  ?t:fail(BadStatus)
 	  end.
 
 process_status_exiting(Config) when is_list(Config) ->
     %% Make sure that erts_debug:get_internal_state({process_status,P})
     %% returns exiting if it is in status P_EXITING.
-    ?line erts_debug:set_internal_state(available_internal_state,true),
-    ?line Prio = process_flag(priority, max),
-    ?line P = spawn_opt(fun () -> receive after infinity -> ok end end,
+    erts_debug:set_internal_state(available_internal_state,true),
+    Prio = process_flag(priority, max),
+    P = spawn_opt(fun () -> receive after infinity -> ok end end,
 			[{priority, normal}]),
-    ?line erlang:yield(),
+    erlang:yield(),
     %% The tok_loop processes are here to make it hard for the exiting
     %% process to be scheduled in for exit...
-    ?line TokLoops = lists:map(fun (_) ->
-				       spawn_opt(fun tok_loop/0,
-						 [link,{priority, high}])
-			       end,
-			       lists:seq(1, erlang:system_info(schedulers_online))),
-    ?line exit(P, boom),
-    ?line wait_until(
-	    fun () ->
-		    exiting =:= erts_debug:get_internal_state({process_status,P})
-	    end),
-    ?line lists:foreach(fun (Tok) -> unlink(Tok), exit(Tok,bang) end, TokLoops),
-    ?line process_flag(priority, Prio),
-    ?line ok.
+    TokLoops = lists:map(fun (_) ->
+		spawn_opt(fun tok_loop/0,
+		    [link,{priority, high}])
+	end, lists:seq(1, erlang:system_info(schedulers_online))),
+    exit(P, boom),
+    wait_until(fun() ->
+		exiting =:= erts_debug:get_internal_state({process_status,P})
+	end),
+    lists:foreach(fun (Tok) -> unlink(Tok), exit(Tok,bang) end, TokLoops),
+    process_flag(priority, Prio),
+    ok.
 
 otp_4725(Config) when is_list(Config) ->
-    ?line Tester = self(),
-    ?line Ref1 = make_ref(),
-    ?line Pid1 = spawn_opt(fun () ->
-				   Tester ! {Ref1, process_info(self())},
-				   receive
-				       Ref1 -> bye
-				   end
-			   end,
-			   [link,
-			    {priority, max},
-			    {fullsweep_after, 600}]),
-    ?line receive
-	      {Ref1, ProcInfo1A} ->
-		  ?line ProcInfo1B = process_info(Pid1),
-		  ?line Pid1 ! Ref1,
-		  ?line check_proc_infos(ProcInfo1A, ProcInfo1B)
-	  end,
-    ?line Ref2 = make_ref(),
-    ?line Pid2 = spawn_opt(fun () ->
-				   Tester ! {Ref2, process_info(self())},
-				   receive
-				       Ref2 -> bye
-				   end
-			   end,
-			   []),
-    ?line receive
-	      {Ref2, ProcInfo2A} ->
-		  ?line ProcInfo2B = process_info(Pid2),
-		  ?line Pid2 ! Ref2,
-		  ?line check_proc_infos(ProcInfo2A, ProcInfo2B)
-	  end,
-    ?line ok.
+    Tester = self(),
+    Ref1 = make_ref(),
+    Pid1 = spawn_opt(fun () ->
+		Tester ! {Ref1, process_info(self())},
+		receive
+		    Ref1 -> bye
+		end
+	end, [link, {priority, max}, {fullsweep_after, 600}]),
+    receive
+	{Ref1, ProcInfo1A} ->
+	    ProcInfo1B = process_info(Pid1),
+	    Pid1 ! Ref1,
+	    check_proc_infos(ProcInfo1A, ProcInfo1B)
+    end,
+    Ref2 = make_ref(),
+    Pid2 = spawn_opt(fun () ->
+		Tester ! {Ref2, process_info(self())},
+		receive
+		    Ref2 -> bye
+		end
+	end,
+	[]),
+    receive
+	{Ref2, ProcInfo2A} ->
+	    ProcInfo2B = process_info(Pid2),
+	    Pid2 ! Ref2,
+	    check_proc_infos(ProcInfo2A, ProcInfo2B)
+    end,
+    ok.
 
 check_proc_infos(A, B) ->
-    ?line IC = lists:keysearch(initial_call, 1, A),
-    ?line IC = lists:keysearch(initial_call, 1, B),
+    IC = lists:keysearch(initial_call, 1, A),
+    IC = lists:keysearch(initial_call, 1, B),
 
-    ?line L = lists:keysearch(links, 1, A),
-    ?line L = lists:keysearch(links, 1, B),
+    L = lists:keysearch(links, 1, A),
+    L = lists:keysearch(links, 1, B),
 
-    ?line D = lists:keysearch(dictionary, 1, A),
-    ?line D = lists:keysearch(dictionary, 1, B),
+    D = lists:keysearch(dictionary, 1, A),
+    D = lists:keysearch(dictionary, 1, B),
 
-    ?line TE = lists:keysearch(trap_exit, 1, A),
-    ?line TE = lists:keysearch(trap_exit, 1, B),
+    TE = lists:keysearch(trap_exit, 1, A),
+    TE = lists:keysearch(trap_exit, 1, B),
 
-    ?line EH = lists:keysearch(error_handler, 1, A),
-    ?line EH = lists:keysearch(error_handler, 1, B),
+    EH = lists:keysearch(error_handler, 1, A),
+    EH = lists:keysearch(error_handler, 1, B),
 
-    ?line P = lists:keysearch(priority, 1, A),
-    ?line P = lists:keysearch(priority, 1, B),
+    P = lists:keysearch(priority, 1, A),
+    P = lists:keysearch(priority, 1, B),
 
-    ?line GL = lists:keysearch(group_leader, 1, A),
-    ?line GL = lists:keysearch(group_leader, 1, B),
+    GL = lists:keysearch(group_leader, 1, A),
+    GL = lists:keysearch(group_leader, 1, B),
 
-    ?line GC = lists:keysearch(garbage_collection, 1, A),
-    ?line GC = lists:keysearch(garbage_collection, 1, B),
+    GC = lists:keysearch(garbage_collection, 1, A),
+    GC = lists:keysearch(garbage_collection, 1, B),
 
-    ?line ok.
+    ok.
 
 
 %% Dummies.
@@ -928,18 +914,18 @@ stop_spawner() ->
 
 %% Tests erlang:bump_reductions/1.
 bump_reductions(Config) when is_list(Config) ->
-    ?line erlang:garbage_collect(),
-    ?line receive after 1 -> ok end,		% Clear reductions.
-    ?line {reductions,R1} = process_info(self(), reductions),
-    ?line true = erlang:bump_reductions(100),
-    ?line {reductions,R2} = process_info(self(), reductions),
-    ?line case R2-R1 of
+    erlang:garbage_collect(),
+    receive after 1 -> ok end,		% Clear reductions.
+    {reductions,R1} = process_info(self(), reductions),
+    true = erlang:bump_reductions(100),
+    {reductions,R2} = process_info(self(), reductions),
+    case R2-R1 of
 	      Diff when Diff < 100 ->
-		  ?line ok = io:format("R1 = ~w, R2 = ~w", [R1, R2]),
-		  ?line test_server:fail({small_diff, Diff});
+		  ok = io:format("R1 = ~w, R2 = ~w", [R1, R2]),
+		  test_server:fail({small_diff, Diff});
 	      Diff when Diff > 110 ->
-		  ?line ok = io:format("R1 = ~w, R2 = ~w", [R1, R2]),
-		  ?line test_server:fail({big_diff, Diff});
+		  ok = io:format("R1 = ~w, R2 = ~w", [R1, R2]),
+		  test_server:fail({big_diff, Diff});
 	      Diff ->
 		  io:format("~p\n", [Diff]),
 		  ok
@@ -949,11 +935,11 @@ bump_reductions(Config) when is_list(Config) ->
     bump_big(R2, 16#08000000).
 
 bump_big(Prev, Limit) ->
-    ?line true = erlang:bump_reductions(100000), %Limited to CONTEXT_REDUCTIONS.
-    ?line case process_info(self(), reductions) of
+    true = erlang:bump_reductions(100000), %Limited to CONTEXT_REDUCTIONS.
+    case process_info(self(), reductions) of
 	      {reductions,Big} when is_integer(Big), Big > Limit ->
-		  ?line erlang:garbage_collect(),
-		  ?line io:format("~p\n", [Big]);
+		  erlang:garbage_collect(),
+		  io:format("~p\n", [Big]);
 	      {reductions,R} when is_integer(R), R > Prev ->
 		  bump_big(R, Limit)
 	  end,
@@ -964,34 +950,34 @@ bump_big(Prev, Limit) ->
 low_prio(Config) when is_list(Config) ->
     case erlang:system_info(schedulers_online) of
 	1 ->
-	    ?line ok = low_prio_test(Config);
+	    ok = low_prio_test(Config);
 	_ -> 
-	    ?line erlang:system_flag(multi_scheduling, block),
-	    ?line ok = low_prio_test(Config),
-	    ?line erlang:system_flag(multi_scheduling, unblock),
-	    ?line {comment,
+	    erlang:system_flag(multi_scheduling, block),
+	    ok = low_prio_test(Config),
+	    erlang:system_flag(multi_scheduling, unblock),
+	    {comment,
 		   "Test not written for SMP runtime system. "
 		   "Multi scheduling blocked during test."}
     end.
 
 low_prio_test(Config) when is_list(Config) ->
-    ?line process_flag(trap_exit, true),
-    ?line S = spawn_link(?MODULE, prio_server, [0, 0]),
-    ?line PCs = spawn_prio_clients(S, erlang:system_info(schedulers_online)),
-    ?line timer:sleep(2000),
-    ?line lists:foreach(fun (P) -> exit(P, kill) end, PCs),
-    ?line S ! exit,
-    ?line receive {'EXIT', S, {A, B}} -> check_prio(A, B) end,
+    process_flag(trap_exit, true),
+    S = spawn_link(?MODULE, prio_server, [0, 0]),
+    PCs = spawn_prio_clients(S, erlang:system_info(schedulers_online)),
+    timer:sleep(2000),
+    lists:foreach(fun (P) -> exit(P, kill) end, PCs),
+    S ! exit,
+    receive {'EXIT', S, {A, B}} -> check_prio(A, B) end,
     ok.
 
 check_prio(A, B) ->
-    ?line Prop = A/B,
-    ?line ok = io:format("Low=~p, High=~p, Prop=~p\n", [A, B, Prop]),
+    Prop = A/B,
+    ok = io:format("Low=~p, High=~p, Prop=~p\n", [A, B, Prop]),
 
     %% It isn't 1/8, it's more like 0.3, but let's check that
     %% the low-prio processes get some little chance to run at all.
-    ?line true = (Prop < 1.0),
-    ?line true = (Prop > 1/32).
+    true = (Prop < 1.0),
+    true = (Prop > 1/32).
 
 prio_server(A, B) ->
     receive
@@ -1051,25 +1037,25 @@ yield(Config) when is_list(Config) ->
     end.
 
 yield_test() ->
-    ?line erlang:garbage_collect(),
-    ?line receive after 1 -> ok end,		% Clear reductions.
-    ?line SC = schedcnt(start),
-    ?line {reductions, R1} = process_info(self(), reductions),
-    ?line {ok, true} = call_yield(middle),
-    ?line true = call_yield(final),
-    ?line true = call_yield(),
-    ?line true = apply(erlang, yield, []),
-    ?line {reductions, R2} = process_info(self(), reductions),
-    ?line Schedcnt = schedcnt(stop, SC),
-    ?line case {R2-R1, Schedcnt} of
-	      {Diff, 4} when Diff < 30 ->
-		  ?line ok = io:format("R1 = ~w, R2 = ~w, Schedcnt = ~w", 
-				       [R1, R2, Schedcnt]);
-	      {Diff, _} ->
-		  ?line ok = io:format("R1 = ~w, R2 = ~w, Schedcnt = ~w", 
-				       [R1, R2, Schedcnt]),
-		  ?line test_server:fail({measurement_error, Diff, Schedcnt})
-	  end.
+    erlang:garbage_collect(),
+    receive after 1 -> ok end,		% Clear reductions.
+    SC = schedcnt(start),
+    {reductions, R1} = process_info(self(), reductions),
+    {ok, true} = call_yield(middle),
+    true = call_yield(final),
+    true = call_yield(),
+    true = apply(erlang, yield, []),
+    {reductions, R2} = process_info(self(), reductions),
+    Schedcnt = schedcnt(stop, SC),
+    case {R2-R1, Schedcnt} of
+	{Diff, 4} when Diff < 30 ->
+	    ok = io:format("R1 = ~w, R2 = ~w, Schedcnt = ~w", 
+		[R1, R2, Schedcnt]);
+	{Diff, _} ->
+	    ok = io:format("R1 = ~w, R2 = ~w, Schedcnt = ~w", 
+		[R1, R2, Schedcnt]),
+	    test_server:fail({measurement_error, Diff, Schedcnt})
+    end.
 
 call_yield() ->
     erlang:yield().
@@ -1108,61 +1094,61 @@ schedcnt(stop, {Ref, Pid}) when is_reference(Ref), is_pid(Pid) ->
 yield2(doc) -> [];
 yield2(suite) -> [];
 yield2(Config) when is_list(Config) ->
-    ?line Me = self(),
-    ?line Go = make_ref(),
-    ?line RedDiff = make_ref(),
-    ?line Done = make_ref(),
-    ?line P = spawn(fun () ->
-			    receive Go -> ok end,
-			    {reductions, R1} = process_info(self(), reductions),
-			    {ok, true} = call_yield(middle),
-			    true = call_yield(final),
-			    true = call_yield(),
-			    true = apply(erlang, yield, []),
-			    {reductions, R2} = process_info(self(), reductions),
-			    Me ! {RedDiff, R2 - R1},
-			    exit(Done)
-		    end),
-    ?line erlang:yield(),
-
-    ?line 1 = erlang:trace(P, true, [running, procs, {tracer, self()}]),
-
-    ?line P ! Go,
+    Me = self(),
+    Go = make_ref(),
+    RedDiff = make_ref(),
+    Done = make_ref(),
+    P = spawn(fun () ->
+		receive Go -> ok end,
+		{reductions, R1} = process_info(self(), reductions),
+		{ok, true} = call_yield(middle),
+		true = call_yield(final),
+		true = call_yield(),
+		true = apply(erlang, yield, []),
+		{reductions, R2} = process_info(self(), reductions),
+		Me ! {RedDiff, R2 - R1},
+		exit(Done)
+	end),
+    erlang:yield(),
+
+    1 = erlang:trace(P, true, [running, procs, {tracer, self()}]),
+
+    P ! Go,
 
     %% receive Go -> ok end,
-    ?line {trace, P, in, _} = next_tmsg(P),
+    {trace, P, in, _} = next_tmsg(P),
 
     %% {ok, true} = call_yield(middle),
-    ?line {trace, P, out, _} = next_tmsg(P),
-    ?line {trace, P, in, _} = next_tmsg(P),
+    {trace, P, out, _} = next_tmsg(P),
+    {trace, P, in, _} = next_tmsg(P),
 
     %% true = call_yield(final),
-    ?line {trace, P, out, _} = next_tmsg(P),
-    ?line {trace, P, in, _} = next_tmsg(P),
+    {trace, P, out, _} = next_tmsg(P),
+    {trace, P, in, _} = next_tmsg(P),
 
     %% true = call_yield(),
-    ?line {trace, P, out, _} = next_tmsg(P),
-    ?line {trace, P, in, _} = next_tmsg(P),
+    {trace, P, out, _} = next_tmsg(P),
+    {trace, P, in, _} = next_tmsg(P),
 
     %% true = apply(erlang, yield, []),
-    ?line {trace, P, out, _} = next_tmsg(P),
-    ?line {trace, P, in, _} = next_tmsg(P),
+    {trace, P, out, _} = next_tmsg(P),
+    {trace, P, in, _} = next_tmsg(P),
 
     %% exit(Done)
-    ?line {trace, P, exit, Done} = next_tmsg(P),
+    {trace, P, exit, Done} = next_tmsg(P),
 
 
-    ?line receive
+    receive
 	      {RedDiff, Reductions} when Reductions < 30, Reductions > 0 ->
 		  io:format("Reductions = ~p~n", [Reductions]),
-		  ?line ok;
+		  ok;
 	      {RedDiff, Reductions} ->
-		  ?line ?t:fail({unexpected_reduction_count, Reductions})
+		  ?t:fail({unexpected_reduction_count, Reductions})
 	  end,
 
-    ?line none = next_tmsg(P),
+    none = next_tmsg(P),
 
-    ?line ok.
+    ok.
 
 next_tmsg(Pid) ->
     receive
@@ -1178,19 +1164,19 @@ next_tmsg(Pid) ->
 bad_register(Config) when is_list(Config) ->
     Name = a_long_and_unused_name,
 
-    ?line {'EXIT',{badarg,_}} = (catch register({bad,name}, self())),
-    ?line fail_register(undefined, self()),
-    ?line fail_register([bad,name], self()),
+    {'EXIT',{badarg,_}} = (catch register({bad,name}, self())),
+    fail_register(undefined, self()),
+    fail_register([bad,name], self()),
 
-    ?line {Dead,Mref} = spawn_monitor(fun() -> true end),
+    {Dead,Mref} = spawn_monitor(fun() -> true end),
     receive
 	{'DOWN',Mref,process,Dead,_} -> ok
     end,
-    ?line fail_register(Name, Dead),
-    ?line fail_register(Name, make_ref()),
-    ?line fail_register(Name, []),
-    ?line fail_register(Name, {bad,process}),
-    ?line fail_register(Name, <<>>),
+    fail_register(Name, Dead),
+    fail_register(Name, make_ref()),
+    fail_register(Name, []),
+    fail_register(Name, {bad,process}),
+    fail_register(Name, <<>>),
     ok.
 
 fail_register(Name, Process) ->
@@ -1201,50 +1187,50 @@ fail_register(Name, Process) ->
 garbage_collect(doc) -> [];
 garbage_collect(suite) -> [];
 garbage_collect(Config) when is_list(Config) ->
-    ?line Prio = process_flag(priority, high),
-    ?line true = erlang:garbage_collect(),
-    ?line TokLoopers = lists:map(fun (_) ->
-					 spawn_opt(fun tok_loop/0,
-						   [{priority, low}, link])
-				 end,
-				 lists:seq(1, 10)),
-    ?line lists:foreach(fun (Pid) ->
-				?line Mon = erlang:monitor(process, Pid),
-				?line DownBefore = receive
-						       {'DOWN', Mon, _, _, _} ->
-							   ?line true
-						   after 0 ->
-							   ?line false
-						   end,
-				?line GC = erlang:garbage_collect(Pid),
-				?line DownAfter = receive
-						      {'DOWN', Mon, _, _, _} ->
-							  ?line true
-						  after 0 ->
-							  ?line false
-						  end,
-				?line true = erlang:demonitor(Mon),
-				?line case {DownBefore, DownAfter} of
-					  {true, _} -> ?line false = GC;
-					  {false, false} -> ?line true = GC;
-					  _ -> ?line GC
-				      end
-			end,
-			processes()),
-    ?line lists:foreach(fun (Pid) ->
-				unlink(Pid),
-				exit(Pid, bang)
-			end, TokLoopers),
-    ?line process_flag(priority, Prio),
-    ?line ok.
+    Prio = process_flag(priority, high),
+    true = erlang:garbage_collect(),
+    
+    TokLoopers = lists:map(fun (_) ->
+		spawn_opt(fun tok_loop/0, [{priority, low}, link])
+	end, lists:seq(1, 10)),
+
+    lists:foreach(fun (Pid) ->
+		Mon = erlang:monitor(process, Pid),
+		DownBefore = receive
+		    {'DOWN', Mon, _, _, _} ->
+			true
+		after 0 ->
+			false
+		end,
+		GC = erlang:garbage_collect(Pid),
+		DownAfter = receive
+		    {'DOWN', Mon, _, _, _} ->
+			true
+		after 0 ->
+			false
+		end,
+		true = erlang:demonitor(Mon),
+		case {DownBefore, DownAfter} of
+		    {true, _} -> false = GC;
+		    {false, false} -> true = GC;
+		    _ -> GC
+		end
+	end, processes()),
+
+    lists:foreach(fun (Pid) ->
+		unlink(Pid),
+		exit(Pid, bang)
+	end, TokLoopers),
+    process_flag(priority, Prio),
+    ok.
 
 process_info_messages(doc) ->
     ["This used to cause the nofrag emulator to dump core"];
 process_info_messages(suite) ->
     [];
 process_info_messages(Config) when is_list(Config) ->
-    ?line process_info_messages_test(),
-    ?line ok.
+    process_info_messages_test(),
+    ok.
 
 process_info_messages_loop(0) -> ok;
 process_info_messages_loop(N) -> process_info_messages_loop(N-1).
@@ -1259,43 +1245,42 @@ process_info_messages_send_my_msgs_to(Rcvr) ->
     end.
 
 process_info_messages_test() ->
-    ?line Go = make_ref(),
-    ?line Done = make_ref(),
-    ?line Rcvr = self(),
-    ?line Rcvr2 = spawn_link(fun () ->
-				     receive {Go, Rcvr} -> ok end,
-				     garbage_collect(),
-				     Rcvr ! {Done, self()}
-			     end),
-    ?line Sndrs = lists:map(
-		    fun (_) ->
-			    spawn_link(fun () ->
-					       Rcvr ! {Go, self()},
-					       receive {Go, Rcvr} -> ok end,
-					       BigData = lists:seq(1, 1000),
-					       Rcvr ! BigData,
-					       Rcvr ! BigData,
-					       Rcvr ! BigData,
-					       Rcvr ! {Done, self()}
-				       end)
-		    end,
-		    lists:seq(1, 10)),
-    ?line lists:foreach(fun (Sndr) -> receive {Go, Sndr} -> ok end end,
+    Go = make_ref(),
+    Done = make_ref(),
+    Rcvr = self(),
+    Rcvr2 = spawn_link(fun () ->
+		receive {Go, Rcvr} -> ok end,
+		garbage_collect(),
+		Rcvr ! {Done, self()}
+	end),
+    Sndrs = lists:map(
+	fun (_) ->
+		spawn_link(fun () ->
+			    Rcvr ! {Go, self()},
+			    receive {Go, Rcvr} -> ok end,
+			    BigData = lists:seq(1, 1000),
+			    Rcvr ! BigData,
+			    Rcvr ! BigData,
+			    Rcvr ! BigData,
+			    Rcvr ! {Done, self()}
+		    end)
+	end, lists:seq(1, 10)),
+    lists:foreach(fun (Sndr) -> receive {Go, Sndr} -> ok end end,
 			Sndrs),
-    ?line garbage_collect(),
-    ?line erlang:yield(),
-    ?line lists:foreach(fun (Sndr) -> Sndr ! {Go, self()} end, Sndrs),
-    ?line process_info_messages_loop(100000000),
-    ?line Msgs = process_info(self(), messages),
-    ?line lists:foreach(fun (Sndr) -> receive {Done, Sndr} -> ok end end,
+    garbage_collect(),
+    erlang:yield(),
+    lists:foreach(fun (Sndr) -> Sndr ! {Go, self()} end, Sndrs),
+    process_info_messages_loop(100000000),
+    Msgs = process_info(self(), messages),
+    lists:foreach(fun (Sndr) -> receive {Done, Sndr} -> ok end end,
 			Sndrs),
-    ?line garbage_collect(),
-    ?line Rcvr2 ! Msgs,
-    ?line process_info_messages_send_my_msgs_to(Rcvr2),
-    ?line Rcvr2 ! {Go, self()},
-    ?line garbage_collect(),
-    ?line receive {Done, Rcvr2} -> ok end,
-    ?line Msgs.
+    garbage_collect(),
+    Rcvr2 ! Msgs,
+    process_info_messages_send_my_msgs_to(Rcvr2),
+    Rcvr2 ! {Go, self()},
+    garbage_collect(),
+    receive {Done, Rcvr2} -> ok end,
+    Msgs.
 
 chk_badarg(Fun) ->
     try Fun(), exit(no_badarg) catch error:badarg -> ok end.
@@ -1305,76 +1290,72 @@ process_flag_badarg(doc) ->
 process_flag_badarg(suite) ->
     [];
 process_flag_badarg(Config) when is_list(Config) ->
-    ?line chk_badarg(fun () -> process_flag(gurka, banan) end),
-    ?line chk_badarg(fun () -> process_flag(trap_exit, gurka) end),
-    ?line chk_badarg(fun () -> process_flag(error_handler, 1) end),
-    ?line chk_badarg(fun () -> process_flag(min_heap_size, gurka) end),
-    ?line chk_badarg(fun () -> process_flag(min_bin_vheap_size, gurka) end),
-    ?line chk_badarg(fun () -> process_flag(min_bin_vheap_size, -1) end),
-    ?line chk_badarg(fun () -> process_flag(priority, 4711) end),
-    ?line chk_badarg(fun () -> process_flag(save_calls, hmmm) end),
-    ?line P= spawn_link(fun () -> receive die -> ok end end),
-    ?line chk_badarg(fun () -> process_flag(P, save_calls, hmmm) end),
-    ?line chk_badarg(fun () -> process_flag(gurka, save_calls, hmmm) end),
-    ?line P ! die,
-    ?line ok.
+    chk_badarg(fun () -> process_flag(gurka, banan) end),
+    chk_badarg(fun () -> process_flag(trap_exit, gurka) end),
+    chk_badarg(fun () -> process_flag(error_handler, 1) end),
+    chk_badarg(fun () -> process_flag(min_heap_size, gurka) end),
+    chk_badarg(fun () -> process_flag(min_bin_vheap_size, gurka) end),
+    chk_badarg(fun () -> process_flag(min_bin_vheap_size, -1) end),
+    chk_badarg(fun () -> process_flag(priority, 4711) end),
+    chk_badarg(fun () -> process_flag(save_calls, hmmm) end),
+    P= spawn_link(fun () -> receive die -> ok end end),
+    chk_badarg(fun () -> process_flag(P, save_calls, hmmm) end),
+    chk_badarg(fun () -> process_flag(gurka, save_calls, hmmm) end),
+    P ! die,
+    ok.
 
 -include_lib("stdlib/include/ms_transform.hrl").
 
 otp_6237(doc) -> [];
 otp_6237(suite) -> [];
 otp_6237(Config) when is_list(Config) ->
-    ?line Slctrs = lists:map(fun (_) ->
-				     spawn_link(fun () ->
-							otp_6237_select_loop()
-						end)
-			     end,
-			     lists:seq(1,5)),
-    ?line lists:foreach(fun (_) -> otp_6237_test() end, lists:seq(1, 100)),
-    ?line lists:foreach(fun (S) -> unlink(S),exit(S, kill) end, Slctrs), 
-    ?line ok.
+    Slctrs = lists:map(fun (_) ->
+		spawn_link(fun () ->
+			    otp_6237_select_loop()
+		    end)
+	end,
+	lists:seq(1,5)),
+    lists:foreach(fun (_) -> otp_6237_test() end, lists:seq(1, 100)),
+    lists:foreach(fun (S) -> unlink(S),exit(S, kill) end, Slctrs), 
+    ok.
 				
 otp_6237_test() ->
-    ?line Parent = self(),
-    ?line Inited = make_ref(),
-    ?line Die = make_ref(),
-    ?line Pid = spawn_link(fun () ->
-				   register(otp_6237,self()),
-				   otp_6237 = ets:new(otp_6237,
-						      [named_table,
-						       ordered_set]),
-				   ets:insert(otp_6237,
-					      [{I,I}
-					       || I <- lists:seq(1, 100)]),
-				   %% Inserting a lot of bif timers
-				   %% increase the possibility that
-				   %% the test will fail when the
-				   %% original cleanup order is used
-				   lists:foreach(
-				     fun (_) ->
-					     erlang:send_after(1000000,
-							       self(),
-							       {a,b,c})
-				     end,
-				     lists:seq(1,1000)),
-				   Parent ! Inited,
-				   receive Die -> bye end
-			   end),
-    ?line receive
-	      Inited -> ?line ok
-	  end,
-    ?line Pid ! Die,
+    Parent = self(),
+    Inited = make_ref(),
+    Die = make_ref(),
+    Pid = spawn_link(fun () ->
+		register(otp_6237,self()),
+		otp_6237 = ets:new(otp_6237,
+		    [named_table,
+			ordered_set]),
+		ets:insert(otp_6237,
+		    [{I,I}
+			|| I <- lists:seq(1, 100)]),
+		%% Inserting a lot of bif timers
+		%% increase the possibility that
+		%% the test will fail when the
+		%% original cleanup order is used
+		lists:foreach( fun (_) ->
+			    erlang:send_after(1000000, self(), {a,b,c})
+		    end, lists:seq(1,1000)),
+		Parent ! Inited,
+		receive Die -> bye end
+	end),
+    receive
+	Inited -> ok
+    end,
+    Pid ! Die,
     otp_6237_whereis_loop().
 
 otp_6237_whereis_loop() ->
-    ?line case whereis(otp_6237) of
+    case whereis(otp_6237) of
 	      undefined ->
-		  ?line otp_6237 = ets:new(otp_6237,
+		  otp_6237 = ets:new(otp_6237,
 					   [named_table,ordered_set]),
-		  ?line ets:delete(otp_6237),
-		  ?line ok;
+		  ets:delete(otp_6237),
+		  ok;
 	      _ ->
-		  ?line otp_6237_whereis_loop()
+		  otp_6237_whereis_loop()
 	  end.
 	     
 otp_6237_select_loop() ->
@@ -1382,7 +1363,6 @@ otp_6237_select_loop() ->
     otp_6237_select_loop().
 
 
-
 -define(NoTestProcs, 10000).
 -record(ptab_list_bif_info, {min_start_reds,
 			     tab_chunks,
@@ -1399,42 +1379,42 @@ processes_large_tab(doc) ->
 processes_large_tab(suite) ->
     [];
 processes_large_tab(Config) when is_list(Config) ->
-    ?line enable_internal_state(),
-    ?line MaxDbgLvl = 20,
-    ?line MinProcTabSize = 2*(1 bsl 15),
-    ?line ProcTabSize0 = 1000000,
-    ?line ProcTabSize1 = case {erlang:system_info(schedulers_online),
-			       erlang:system_info(logical_processors)} of
-			     {Schdlrs, Cpus} when is_integer(Cpus),
-			                          Schdlrs =< Cpus ->
-				 ProcTabSize0;
-			     _ ->
-				 ProcTabSize0 div 4
-			 end,
-    ?line ProcTabSize2 = case erlang:system_info(debug_compiled) of
-		     true -> ProcTabSize1 - 500000;
-		     false -> ProcTabSize1
-		 end,
+    enable_internal_state(),
+    MaxDbgLvl = 20,
+    MinProcTabSize = 2*(1 bsl 15),
+    ProcTabSize0 = 1000000,
+    ProcTabSize1 = case {erlang:system_info(schedulers_online),
+	    erlang:system_info(logical_processors)} of
+	{Schdlrs, Cpus} when is_integer(Cpus),
+	Schdlrs =< Cpus ->
+	    ProcTabSize0;
+	_ ->
+	    ProcTabSize0 div 4
+    end,
+    ProcTabSize2 = case erlang:system_info(debug_compiled) of
+	true -> ProcTabSize1 - 500000;
+	false -> ProcTabSize1
+    end,
     %% With high debug levels this test takes so long time that
     %% the connection times out; therefore, shrink the test on
     %% high debug levels.
-    ?line DbgLvl = case erts_debug:get_internal_state(processes_bif_info) of
+    DbgLvl = case erts_debug:get_internal_state(processes_bif_info) of
 		       #ptab_list_bif_info{debug_level = Lvl} when Lvl > MaxDbgLvl ->
 			   20;
 		       #ptab_list_bif_info{debug_level = Lvl} when Lvl < 0 ->
-			   ?line ?t:fail({debug_level, Lvl});
+			   ?t:fail({debug_level, Lvl});
 		       #ptab_list_bif_info{debug_level = Lvl} ->
 			   Lvl
 		   end,
-    ?line ProcTabSize3 = ProcTabSize2 - (1300000 * DbgLvl div MaxDbgLvl),
-    ?line ProcTabSize = case ProcTabSize3 < MinProcTabSize of
+    ProcTabSize3 = ProcTabSize2 - (1300000 * DbgLvl div MaxDbgLvl),
+    ProcTabSize = case ProcTabSize3 < MinProcTabSize of
 			    true -> MinProcTabSize;
 			    false -> ProcTabSize3
 			end,
-    ?line {ok, LargeNode} = start_node(Config,
+    {ok, LargeNode} = start_node(Config,
 				       "+P " ++ integer_to_list(ProcTabSize)),
-    ?line Res = rpc:call(LargeNode, ?MODULE, processes_bif_test, []),
-    ?line case rpc:call(LargeNode,
+    Res = rpc:call(LargeNode, ?MODULE, processes_bif_test, []),
+    case rpc:call(LargeNode,
 			erts_debug,
 			get_internal_state,
 			[processes_bif_info]) of
@@ -1442,40 +1422,37 @@ processes_large_tab(Config) when is_list(Config) ->
 							    Chunks > 1 -> ok;
 	      PBInfo -> ?t:fail(PBInfo)
 	  end,
-    ?line stop_node(LargeNode),
-    ?line chk_processes_bif_test_res(Res).
+    stop_node(LargeNode),
+    chk_processes_bif_test_res(Res).
 
 processes_default_tab(doc) ->
     [];
 processes_default_tab(suite) ->
     [];
 processes_default_tab(Config) when is_list(Config) ->
-    ?line {ok, DefaultNode} = start_node(Config, ""),
-    ?line Res = rpc:call(DefaultNode, ?MODULE, processes_bif_test, []),
-    ?line stop_node(DefaultNode),
-    ?line chk_processes_bif_test_res(Res).
+    {ok, DefaultNode} = start_node(Config, ""),
+    Res = rpc:call(DefaultNode, ?MODULE, processes_bif_test, []),
+    stop_node(DefaultNode),
+    chk_processes_bif_test_res(Res).
 
 processes_small_tab(doc) ->
     [];
 processes_small_tab(suite) ->
     [];
 processes_small_tab(Config) when is_list(Config) ->
-    ?line {ok, SmallNode} = start_node(Config, "+P 1024"),
-    ?line Res = rpc:call(SmallNode, ?MODULE, processes_bif_test, []),
-    ?line PBInfo = rpc:call(SmallNode,
-			    erts_debug,
-			    get_internal_state,
-			    [processes_bif_info]),
-    ?line stop_node(SmallNode),
-    ?line true = PBInfo#ptab_list_bif_info.tab_chunks < 10,
-    ?line chk_processes_bif_test_res(Res).
+    {ok, SmallNode} = start_node(Config, "+P 1024"),
+    Res    = rpc:call(SmallNode, ?MODULE, processes_bif_test, []),
+    PBInfo = rpc:call(SmallNode, erts_debug, get_internal_state, [processes_bif_info]),
+    stop_node(SmallNode),
+    true = PBInfo#ptab_list_bif_info.tab_chunks < 10,
+    chk_processes_bif_test_res(Res).
 
 processes_this_tab(doc) ->
     [];
 processes_this_tab(suite) ->
     [];
 processes_this_tab(Config) when is_list(Config) ->
-    ?line chk_processes_bif_test_res(processes_bif_test()).
+    chk_processes_bif_test_res(processes_bif_test()).
 
 chk_processes_bif_test_res(ok) -> ok;
 chk_processes_bif_test_res({comment, _} = Comment) -> Comment;
@@ -1575,26 +1552,26 @@ do_processes(WantReds) ->
     processes().
 
 processes_bif_test() ->
-    ?line Tester = self(),
-    ?line enable_internal_state(),
-    ?line PBInfo = erts_debug:get_internal_state(processes_bif_info),
-    ?line print_processes_bif_info(PBInfo),
-    ?line WantReds = PBInfo#ptab_list_bif_info.min_start_reds + 10,
-    ?line WillTrap = case PBInfo of
-			#ptab_list_bif_info{tab_chunks = Chunks} when Chunks < 10 ->
-			    false; %% Skip for small tables
-			#ptab_list_bif_info{tab_chunks = Chunks,
-					    tab_chunks_size = ChunksSize,
-					    tab_indices_per_red = IndiciesPerRed
-					   } ->
-			    Chunks*ChunksSize >= IndiciesPerRed*WantReds
-		    end,
-    ?line Processes = fun () ->
-			      erts_debug:set_internal_state(reds_left,WantReds),
-			      processes()
-		      end,
+    Tester = self(),
+    enable_internal_state(),
+    PBInfo = erts_debug:get_internal_state(processes_bif_info),
+    print_processes_bif_info(PBInfo),
+    WantReds = PBInfo#ptab_list_bif_info.min_start_reds + 10,
+    WillTrap = case PBInfo of
+	#ptab_list_bif_info{tab_chunks = Chunks} when Chunks < 10 ->
+	    false; %% Skip for small tables
+	#ptab_list_bif_info{tab_chunks = Chunks,
+	    tab_chunks_size = ChunksSize,
+	    tab_indices_per_red = IndiciesPerRed
+	} ->
+	    Chunks*ChunksSize >= IndiciesPerRed*WantReds
+    end,
+    Processes = fun () ->
+	    erts_debug:set_internal_state(reds_left,WantReds),
+	    processes()
+    end,
 
-    ?line ok = do_processes_bif_test(WantReds, WillTrap, Processes),
+    ok = do_processes_bif_test(WantReds, WillTrap, Processes),
 
     case WillTrap of
 	false ->
@@ -1602,8 +1579,8 @@ processes_bif_test() ->
 	true ->
 	    %% Do it again with a process suspended while
 	    %% in the processes/0 bif.
-	    ?line erlang:system_flag(multi_scheduling, block),
-	    ?line Suspendee = spawn_link(fun () ->
+	    erlang:system_flag(multi_scheduling, block),
+	    Suspendee = spawn_link(fun () ->
 						 Tester ! {suspend_me, self()},
 						 Tester ! {self(),
 							   done,
@@ -1613,175 +1590,156 @@ processes_bif_test() ->
 							 ok
 						 end
 					 end),
-	    ?line receive {suspend_me, Suspendee} -> ok end,
-	    ?line erlang:suspend_process(Suspendee),
-	    ?line erlang:system_flag(multi_scheduling, unblock),
+	    receive {suspend_me, Suspendee} -> ok end,
+	    erlang:suspend_process(Suspendee),
+	    erlang:system_flag(multi_scheduling, unblock),
 	    
-	    ?line [{status,suspended},
-		   {current_function,{erlang,ptab_list_continue,2}}]
-		= process_info(Suspendee, [status, current_function]),
+	    [{status,suspended},{current_function,{erlang,ptab_list_continue,2}}] =
+		process_info(Suspendee, [status, current_function]),
 
-	    ?line ok = do_processes_bif_test(WantReds, WillTrap, Processes),
+	    ok = do_processes_bif_test(WantReds, WillTrap, Processes),
 	    
-	    ?line erlang:resume_process(Suspendee),
-	    ?line receive {Suspendee, done, _} -> ok end,
-	    ?line unlink(Suspendee),
-	    ?line exit(Suspendee, bang)
+	    erlang:resume_process(Suspendee),
+	    receive {Suspendee, done, _} -> ok end,
+	    unlink(Suspendee),
+	    exit(Suspendee, bang)
     end,
     case get(processes_bif_testcase_comment) of
-	undefined -> ?line ok;
-	Comment -> ?line {comment, Comment}
+	undefined -> ok;
+	Comment -> {comment, Comment}
     end.
     
 do_processes_bif_test(WantReds, DieTest, Processes) ->
-    ?line Tester = self(),
-    ?line SpawnProcesses = fun (Prio) ->
-				   spawn_opt(?MODULE,
-					     do_processes,
-					     [WantReds],
-					     [link, {priority, Prio}])
-			   end,
-    ?line Cleaner = spawn_link(fun () ->
-				       process_flag(trap_exit, true),
-				       Tester ! {cleaner_alive, self()},
-				       processes_bif_cleaner()
-			       end),
-    ?line receive {cleaner_alive, Cleaner} -> ok end,
+    Tester = self(),
+    SpawnProcesses = fun (Prio) ->
+	    spawn_opt(?MODULE, do_processes, [WantReds], [link, {priority, Prio}])
+    end,
+    Cleaner = spawn_link(fun () ->
+		process_flag(trap_exit, true),
+		Tester ! {cleaner_alive, self()},
+		processes_bif_cleaner()
+	end),
+    receive {cleaner_alive, Cleaner} -> ok end,
     try
-	?line DoIt = make_ref(),
-	?line GetGoing = make_ref(),
-	?line {NoTestProcs, TestProcs} = spawn_initial_hangarounds(Cleaner),
-	?line ?t:format("Testing with ~p processes~n", [NoTestProcs]),
-	?line SpawnHangAround = fun () ->
-					spawn(?MODULE,
-					      hangaround,
-					      [Cleaner, new_hangaround])
-				end,
-	?line Killer = spawn_opt(fun () ->
-					 Splt = NoTestProcs div 10,
-					 {TP1, TP23} = lists:split(Splt,
-								   TestProcs),
-					 {TP2, TP3} = lists:split(Splt, TP23),
-					 erlang:system_flag(multi_scheduling,
-							    block),
-					 Tester ! DoIt,
-					 receive GetGoing -> ok end,
-					 erlang:system_flag(multi_scheduling,
-							    unblock),
-					 SpawnProcesses(high),
-					 lists:foreach(
-					   fun (P) ->
-						   SpawnHangAround(),
-						   exit(P, bang)
-					   end,
-					   TP1),
-					 SpawnProcesses(high),
-					 erlang:yield(),
-					 lists:foreach(
-					   fun (P) ->
-						   SpawnHangAround(),
-						   exit(P, bang)
-					   end,
-					   TP2),
-					 SpawnProcesses(high),
-					 lists:foreach(
-					   fun (P) ->
-						   SpawnHangAround(),
-						   exit(P, bang)
-					   end,
-					   TP3)
-				 end,
-				 [{priority, high}, link]),
-	?line receive DoIt -> ok end,
-	?line process_flag(priority, low),
-	?line SpawnProcesses(low),
-	?line erlang:yield(),
-	?line process_flag(priority, normal),
-	?line CorrectProcs0 = erts_debug:get_internal_state(processes),
-	?line Killer ! GetGoing,
-	?line erts_debug:set_internal_state(reds_left, WantReds),
-	?line Procs0 = processes(),
-	?line Procs = lists:sort(Procs0),
-	?line CorrectProcs = lists:sort(CorrectProcs0),
-	?line LengthCorrectProcs = length(CorrectProcs),
-	?line ?t:format("~p = length(CorrectProcs)~n", [LengthCorrectProcs]),
-	?line true = LengthCorrectProcs > NoTestProcs,
-	?line case CorrectProcs =:= Procs of
-		  true ->
-		      ?line ok;
-		  false ->
-		      ?line processes_unexpected_result(CorrectProcs, Procs)
-	      end,
-	?line unlink(Killer),
-	?line exit(Killer, bang)
+	DoIt = make_ref(),
+	GetGoing = make_ref(),
+	{NoTestProcs, TestProcs} = spawn_initial_hangarounds(Cleaner),
+	?t:format("Testing with ~p processes~n", [NoTestProcs]),
+	SpawnHangAround = fun () ->
+		spawn(?MODULE, hangaround, [Cleaner, new_hangaround])
+	end,
+	Killer = spawn_opt(fun () ->
+		    Splt = NoTestProcs div 10,
+		    {TP1, TP23} = lists:split(Splt, TestProcs),
+		    {TP2, TP3} = lists:split(Splt, TP23),
+		    erlang:system_flag(multi_scheduling, block),
+		    Tester ! DoIt,
+		    receive GetGoing -> ok end,
+		    erlang:system_flag(multi_scheduling, unblock),
+		    SpawnProcesses(high),
+		    lists:foreach( fun (P) ->
+				SpawnHangAround(),
+				exit(P, bang)
+			end, TP1),
+		    SpawnProcesses(high),
+		    erlang:yield(),
+		    lists:foreach( fun (P) ->
+				SpawnHangAround(),
+				exit(P, bang)
+			end, TP2),
+		    SpawnProcesses(high),
+		    lists:foreach(
+			fun (P) ->
+				SpawnHangAround(),
+				exit(P, bang)
+			end, TP3)
+	    end, [{priority, high}, link]),
+	receive DoIt -> ok end,
+	process_flag(priority, low),
+	SpawnProcesses(low),
+	erlang:yield(),
+	process_flag(priority, normal),
+	CorrectProcs0 = erts_debug:get_internal_state(processes),
+	Killer ! GetGoing,
+	erts_debug:set_internal_state(reds_left, WantReds),
+	Procs0 = processes(),
+	Procs = lists:sort(Procs0),
+	CorrectProcs = lists:sort(CorrectProcs0),
+	LengthCorrectProcs = length(CorrectProcs),
+	?t:format("~p = length(CorrectProcs)~n", [LengthCorrectProcs]),
+	true = LengthCorrectProcs > NoTestProcs,
+	case CorrectProcs =:= Procs of
+	    true ->
+		ok;
+	    false ->
+		processes_unexpected_result(CorrectProcs, Procs)
+	end,
+	unlink(Killer),
+	exit(Killer, bang)
     after
 	unlink(Cleaner),
         exit(Cleaner, kill),
         %% Wait for the system to recover to a normal state...
 	wait_until_system_recover()
     end,
-    ?line do_processes_bif_die_test(DieTest, Processes),
-    ?line ok.
+    do_processes_bif_die_test(DieTest, Processes),
+    ok.
 
 
 do_processes_bif_die_test(false, _Processes) ->
-    ?line ?t:format("Skipping test killing process executing processes/0~n",[]),
-    ?line ok;
+    ?t:format("Skipping test killing process executing processes/0~n",[]),
+    ok;
 do_processes_bif_die_test(true, Processes) ->
-    ?line do_processes_bif_die_test(5, Processes);
+    do_processes_bif_die_test(5, Processes);
 do_processes_bif_die_test(N, Processes) ->
-    ?line ?t:format("Doing test killing process executing processes/0~n",[]),
+    ?t:format("Doing test killing process executing processes/0~n",[]),
     try
-	?line Tester = self(),
-	?line Oooh_Nooooooo = make_ref(),
-	?line {_, DieWhileDoingMon} = erlang:spawn_monitor(
-					fun () ->
-						Victim = self(),
-						spawn_opt(
-						  fun () ->
-							  exit(Victim, got_him)
-						  end,
-						  [link,
-						   {priority, max}]),
-						Tester ! {Oooh_Nooooooo,
-							  hd(Processes())},
-						exit(ohhhh_nooooo)
-					end),
-	?line receive
-		  {'DOWN', DieWhileDoingMon, _, _, Reason} ->
-		      case Reason of
-			  got_him -> ok;
-			  _ -> throw({kill_in_trap, Reason})
-		      end
-	      end,
-	?line receive
-		  {Oooh_Nooooooo, _} ->
-		      ?line throw({kill_in_trap, 'Oooh_Nooooooo'})
-	      after 0 ->
-		      ?line ok
-	      end,
-	?line PrcsCllrsSeqLen = 2*erlang:system_info(schedulers_online),
-	?line PrcsCllrsSeq = lists:seq(1, PrcsCllrsSeqLen),
-	?line ProcsCallers = lists:map(
-			       fun (_) ->
-				       spawn_link(
-					 fun () ->
-						 Tester ! hd(Processes())
-					 end)
-			       end,
-			       PrcsCllrsSeq),
-	?line erlang:yield(),
+	Tester = self(),
+	Oooh_Nooooooo = make_ref(),
+	{_, DieWhileDoingMon} = erlang:spawn_monitor( fun () ->
+		    Victim = self(),
+		    spawn_opt(
+			fun () ->
+				exit(Victim, got_him)
+			end,
+			[link, {priority, max}]),
+		    Tester ! {Oooh_Nooooooo,
+			hd(Processes())},
+		    exit(ohhhh_nooooo)
+	    end),
+	receive
+	    {'DOWN', DieWhileDoingMon, _, _, Reason} ->
+		case Reason of
+		    got_him -> ok;
+		    _ -> throw({kill_in_trap, Reason})
+		end
+	end,
+	receive
+	    {Oooh_Nooooooo, _} ->
+		throw({kill_in_trap, 'Oooh_Nooooooo'})
+	after 0 ->
+		ok
+	end,
+	PrcsCllrsSeqLen = 2*erlang:system_info(schedulers_online),
+	PrcsCllrsSeq = lists:seq(1, PrcsCllrsSeqLen),
+	ProcsCallers = lists:map( fun (_) ->
+		    spawn_link(
+			fun () ->
+				Tester ! hd(Processes())
+			end)
+	    end, PrcsCllrsSeq),
+	erlang:yield(),
 	{ProcsCallers1, ProcsCallers2} = lists:split(PrcsCllrsSeqLen div 2,
 						     ProcsCallers),
-	?line process_flag(priority, high),
-	?line lists:foreach(
+	process_flag(priority, high),
+	lists:foreach(
 		fun (P) ->
 			unlink(P),
 			exit(P, bang)
 		end,
 		lists:reverse(ProcsCallers2) ++ ProcsCallers1),
-	?line process_flag(priority, normal),
-	?line ok
+	process_flag(priority, normal),
+	ok
     catch
 	throw:{kill_in_trap, R} when N > 0 ->
 	    ?t:format("Failed to kill in trap: ~p~n", [R]),
@@ -1844,23 +1802,23 @@ processes_last_call_trap(doc) ->
 processes_last_call_trap(suite) ->
     [];
 processes_last_call_trap(Config) when is_list(Config) ->
-    ?line enable_internal_state(),
-    ?line Processes = fun () -> processes() end,
-    ?line PBInfo = erts_debug:get_internal_state(processes_bif_info),
-    ?line print_processes_bif_info(PBInfo),
-    ?line WantReds = case PBInfo#ptab_list_bif_info.min_start_reds of
-			 R when R > 10 -> R - 1;
-			 _R -> 9
-		     end,
-    ?line lists:foreach(fun (_) ->
-				?line erts_debug:set_internal_state(reds_left,
-								    WantReds),
-				Processes(),
-				?line erts_debug:set_internal_state(reds_left,
-								    WantReds),
-				my_processes()
-			end,
-			lists:seq(1,100)).
+    enable_internal_state(),
+    Processes = fun () -> processes() end,
+    PBInfo = erts_debug:get_internal_state(processes_bif_info),
+    print_processes_bif_info(PBInfo),
+    WantReds = case PBInfo#ptab_list_bif_info.min_start_reds of
+	R when R > 10 -> R - 1;
+	_R -> 9
+    end,
+    lists:foreach(fun (_) ->
+		erts_debug:set_internal_state(reds_left,
+		    WantReds),
+		Processes(),
+		erts_debug:set_internal_state(reds_left,
+		    WantReds),
+		my_processes()
+	end,
+	lists:seq(1,100)).
     
 my_processes() ->
     processes().
@@ -1870,108 +1828,106 @@ processes_apply_trap(doc) ->
 processes_apply_trap(suite) ->
     [];
 processes_apply_trap(Config) when is_list(Config) ->
-    ?line enable_internal_state(),
-    ?line PBInfo = erts_debug:get_internal_state(processes_bif_info),
-    ?line print_processes_bif_info(PBInfo),
-    ?line WantReds = case PBInfo#ptab_list_bif_info.min_start_reds of
-			 R when R > 10 -> R - 1;
-			 _R -> 9
-		     end,
-    ?line lists:foreach(fun (_) ->
-				?line erts_debug:set_internal_state(reds_left,
-								    WantReds),
-				?line apply(erlang, processes, [])
-			end,
-			lists:seq(1,100)).
+    enable_internal_state(),
+    PBInfo = erts_debug:get_internal_state(processes_bif_info),
+    print_processes_bif_info(PBInfo),
+    WantReds = case PBInfo#ptab_list_bif_info.min_start_reds of
+	R when R > 10 -> R - 1;
+	_R -> 9
+    end,
+    lists:foreach(fun (_) ->
+		erts_debug:set_internal_state(reds_left,
+		    WantReds),
+		apply(erlang, processes, [])
+	end, lists:seq(1,100)).
 
 processes_gc_trap(doc) ->
     [];
 processes_gc_trap(suite) ->
     [];
 processes_gc_trap(Config) when is_list(Config) ->
-    ?line Tester = self(),
-    ?line enable_internal_state(),
-    ?line PBInfo = erts_debug:get_internal_state(processes_bif_info),
-    ?line print_processes_bif_info(PBInfo),
-    ?line WantReds = PBInfo#ptab_list_bif_info.min_start_reds + 10,
-    ?line Processes = fun () ->
-			      erts_debug:set_internal_state(reds_left,WantReds),
-			      processes()
-		      end,
+    Tester = self(),
+    enable_internal_state(),
+    PBInfo = erts_debug:get_internal_state(processes_bif_info),
+    print_processes_bif_info(PBInfo),
+    WantReds = PBInfo#ptab_list_bif_info.min_start_reds + 10,
+    Processes = fun () ->
+	    erts_debug:set_internal_state(reds_left,WantReds),
+	    processes()
+    end,
 
-    ?line erlang:system_flag(multi_scheduling, block),
-    ?line Suspendee = spawn_link(fun () ->
+    erlang:system_flag(multi_scheduling, block),
+    Suspendee = spawn_link(fun () ->
 					 Tester ! {suspend_me, self()},
 					 Tester ! {self(),
 						   done,
 						   hd(Processes())},
 					 receive after infinity -> ok end
 				 end),
-    ?line receive {suspend_me, Suspendee} -> ok end,
-    ?line erlang:suspend_process(Suspendee),
-    ?line erlang:system_flag(multi_scheduling, unblock),
+    receive {suspend_me, Suspendee} -> ok end,
+    erlang:suspend_process(Suspendee),
+    erlang:system_flag(multi_scheduling, unblock),
 	    
-    ?line [{status,suspended}, {current_function,{erlang,ptab_list_continue,2}}]
+    [{status,suspended}, {current_function,{erlang,ptab_list_continue,2}}]
 	= process_info(Suspendee, [status, current_function]),
 
-    ?line erlang:garbage_collect(Suspendee),
-    ?line erlang:garbage_collect(Suspendee),
+    erlang:garbage_collect(Suspendee),
+    erlang:garbage_collect(Suspendee),
 	    
-    ?line erlang:resume_process(Suspendee),
-    ?line receive {Suspendee, done, _} -> ok end,
-    ?line erlang:garbage_collect(Suspendee),
-    ?line erlang:garbage_collect(Suspendee),
+    erlang:resume_process(Suspendee),
+    receive {Suspendee, done, _} -> ok end,
+    erlang:garbage_collect(Suspendee),
+    erlang:garbage_collect(Suspendee),
 
-    ?line unlink(Suspendee),
-    ?line exit(Suspendee, bang),
-    ?line ok.
+    unlink(Suspendee),
+    exit(Suspendee, bang),
+    ok.
 
 process_flag_heap_size(doc) ->
     [];
 process_flag_heap_size(suite) ->
     [];
 process_flag_heap_size(Config) when is_list(Config) ->
-    HSize  = 2584,   % must be gc fib number
-    VHSize = 317811, % must be gc fib number
-    ?line OldHmin = erlang:process_flag(min_heap_size, HSize),
-    ?line {min_heap_size, HSize} = erlang:process_info(self(), min_heap_size),
-    ?line OldVHmin = erlang:process_flag(min_bin_vheap_size, VHSize),
-    ?line {min_bin_vheap_size, VHSize} = erlang:process_info(self(), min_bin_vheap_size),
-    ?line HSize = erlang:process_flag(min_heap_size, OldHmin),
-    ?line VHSize = erlang:process_flag(min_bin_vheap_size, OldVHmin),
-    ?line ok.
+    HSize  = 2586,   % must be gc fib+ number
+    VHSize = 318187, % must be gc fib+ number
+    OldHmin = erlang:process_flag(min_heap_size, HSize),
+    {min_heap_size, HSize} = erlang:process_info(self(), min_heap_size),
+    OldVHmin = erlang:process_flag(min_bin_vheap_size, VHSize),
+    {min_bin_vheap_size, VHSize} = erlang:process_info(self(), min_bin_vheap_size),
+    HSize = erlang:process_flag(min_heap_size, OldHmin),
+    VHSize = erlang:process_flag(min_bin_vheap_size, OldVHmin),
+    ok.
 
 spawn_opt_heap_size(doc) ->
     [];
 spawn_opt_heap_size(suite) ->
     [];
 spawn_opt_heap_size(Config) when is_list(Config) ->
-    HSize  = 987,   % must be gc fib number
-    VHSize = 46368, % must be gc fib number
-    ?line Pid  = spawn_opt(fun () -> receive stop -> ok end end,
+    HSize  = 987,   % must be gc fib+ number
+    VHSize = 46422, % must be gc fib+ number
+    Pid  = spawn_opt(fun () -> receive stop -> ok end end,
 	[{min_heap_size, HSize},{ min_bin_vheap_size, VHSize}]),
-    ?line {min_heap_size, HSize} = process_info(Pid, min_heap_size),
-    ?line {min_bin_vheap_size, VHSize} = process_info(Pid, min_bin_vheap_size),
-    ?line Pid ! stop,
-    ?line ok.
+    {min_heap_size, HSize} = process_info(Pid, min_heap_size),
+    {min_bin_vheap_size, VHSize} = process_info(Pid, min_bin_vheap_size),
+    Pid ! stop,
+    ok.
 
 processes_term_proc_list(doc) ->
     [];
 processes_term_proc_list(suite) ->
     [];
 processes_term_proc_list(Config) when is_list(Config) ->
-    ?line Tester = self(),
-    ?line as_expected = processes_term_proc_list_test(false),
-    ?line {ok, Node} = start_node(Config, "+Mis true"),
-    ?line RT = spawn_link(Node,
-			  fun () ->
-				  receive after 1000 -> ok end,
-				  processes_term_proc_list_test(false),
-				  Tester ! {it_worked, self()}
-			  end),
-    ?line receive {it_worked, RT} -> ok end,
-    ?line stop_node(Node),
-    ?line ok.
+    Tester = self(),
+    as_expected = processes_term_proc_list_test(false),
+    {ok, Node} = start_node(Config, "+Mis true"),
+    RT = spawn_link(Node, fun () ->
+		receive after 1000 -> ok end,
+		processes_term_proc_list_test(false),
+		Tester ! {it_worked, self()}
+	end),
+    receive {it_worked, RT} -> ok end,
+    stop_node(Node),
+    ok.
     
 -define(CHK_TERM_PROC_LIST(MC, XB),
 	chk_term_proc_list(?LINE, MC, XB)).
@@ -1997,35 +1953,34 @@ chk_term_proc_list(Line, MustChk, ExpectBlks) ->
     ok.
 
 processes_term_proc_list_test(MustChk) ->
-    ?line Tester = self(),
-    ?line enable_internal_state(),
-    ?line PBInfo = erts_debug:get_internal_state(processes_bif_info),
-    ?line print_processes_bif_info(PBInfo),
-    ?line WantReds = PBInfo#ptab_list_bif_info.min_start_reds + 10,
-    ?line #ptab_list_bif_info{tab_chunks = Chunks,
-			      tab_chunks_size = ChunksSize,
-			      tab_indices_per_red = IndiciesPerRed
-			     } = PBInfo,
-    ?line true = Chunks > 1,
-    ?line true = Chunks*ChunksSize >= IndiciesPerRed*WantReds,
-    ?line Processes = fun () ->
-			      erts_debug:set_internal_state(reds_left,
-							    WantReds),
-			      processes()
-		      end,
-    ?line Exit = fun (P) ->
-			 unlink(P),
-			 exit(P, bang),
-			 wait_until(
-			   fun () ->
-				   not lists:member(
-					 P,
-					 erts_debug:get_internal_state(
-					   processes))
-			   end)
-		 end,
-    ?line SpawnSuspendProcessesProc
-	= fun () ->
+    Tester = self(),
+    enable_internal_state(),
+    PBInfo = erts_debug:get_internal_state(processes_bif_info),
+    print_processes_bif_info(PBInfo),
+    WantReds = PBInfo#ptab_list_bif_info.min_start_reds + 10,
+    #ptab_list_bif_info{tab_chunks = Chunks,
+	tab_chunks_size = ChunksSize,
+	tab_indices_per_red = IndiciesPerRed
+    } = PBInfo,
+    true = Chunks > 1,
+    true = Chunks*ChunksSize >= IndiciesPerRed*WantReds,
+    Processes = fun () ->
+	    erts_debug:set_internal_state(reds_left,
+		WantReds),
+	    processes()
+    end,
+    Exit = fun (P) ->
+	    unlink(P),
+	    exit(P, bang),
+	    wait_until(
+		fun () ->
+			not lists:member(
+			    P,
+			    erts_debug:get_internal_state(
+				processes))
+		end)
+    end,
+    SpawnSuspendProcessesProc = fun () ->
 		  erlang:system_flag(multi_scheduling, block),
 		  P = spawn_link(fun () ->
 					 Tester ! {suspend_me, self()},
@@ -2042,72 +1997,72 @@ processes_term_proc_list_test(MustChk) ->
 		      = process_info(P, [status, current_function]),
 		  P
 	  end,
-    ?line ResumeProcessesProc = fun (P) ->
+    ResumeProcessesProc = fun (P) ->
 					erlang:resume_process(P),
 					receive {P, done, _} -> ok end
 				end,
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 0),
-    ?line HangAround = fun () -> receive after infinity -> ok end end,
-    ?line HA1 = spawn_link(HangAround),
-    ?line HA2 = spawn_link(HangAround),
-    ?line HA3 = spawn_link(HangAround),
-    ?line S1 = SpawnSuspendProcessesProc(),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 1),
-    ?line Exit(HA1),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 2),
-    ?line S2 = SpawnSuspendProcessesProc(),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 3),
-    ?line S3 = SpawnSuspendProcessesProc(),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 4),
-    ?line Exit(HA2),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 5),
-    ?line S4 = SpawnSuspendProcessesProc(),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 6),
-    ?line Exit(HA3),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 7),
-    ?line ResumeProcessesProc(S1),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 5),
-    ?line ResumeProcessesProc(S3),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 4),
-    ?line ResumeProcessesProc(S4),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 3),
-    ?line ResumeProcessesProc(S2),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 0),
-    ?line Exit(S1),
-    ?line Exit(S2),
-    ?line Exit(S3),
-    ?line Exit(S4),
-
-
-    ?line HA4 = spawn_link(HangAround),
-    ?line HA5 = spawn_link(HangAround),
-    ?line HA6 = spawn_link(HangAround),
-    ?line S5 = SpawnSuspendProcessesProc(),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 1),
-    ?line Exit(HA4),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 2),
-    ?line S6 = SpawnSuspendProcessesProc(),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 3),
-    ?line Exit(HA5),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 4),
-    ?line S7 = SpawnSuspendProcessesProc(),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 5),
-    ?line Exit(HA6),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 6),
-    ?line S8 = SpawnSuspendProcessesProc(),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 7),
-
-    ?line erlang:system_flag(multi_scheduling, block),
-    ?line Exit(S8),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 7),
-    ?line Exit(S5),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 6),
-    ?line Exit(S7),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 6),
-    ?line Exit(S6),
-    ?line ?CHK_TERM_PROC_LIST(MustChk, 0),
-    ?line erlang:system_flag(multi_scheduling, unblock),
-    ?line as_expected.
+    ?CHK_TERM_PROC_LIST(MustChk, 0),
+    HangAround = fun () -> receive after infinity -> ok end end,
+    HA1 = spawn_link(HangAround),
+    HA2 = spawn_link(HangAround),
+    HA3 = spawn_link(HangAround),
+    S1 = SpawnSuspendProcessesProc(),
+    ?CHK_TERM_PROC_LIST(MustChk, 1),
+    Exit(HA1),
+    ?CHK_TERM_PROC_LIST(MustChk, 2),
+    S2 = SpawnSuspendProcessesProc(),
+    ?CHK_TERM_PROC_LIST(MustChk, 3),
+    S3 = SpawnSuspendProcessesProc(),
+    ?CHK_TERM_PROC_LIST(MustChk, 4),
+    Exit(HA2),
+    ?CHK_TERM_PROC_LIST(MustChk, 5),
+    S4 = SpawnSuspendProcessesProc(),
+    ?CHK_TERM_PROC_LIST(MustChk, 6),
+    Exit(HA3),
+    ?CHK_TERM_PROC_LIST(MustChk, 7),
+    ResumeProcessesProc(S1),
+    ?CHK_TERM_PROC_LIST(MustChk, 5),
+    ResumeProcessesProc(S3),
+    ?CHK_TERM_PROC_LIST(MustChk, 4),
+    ResumeProcessesProc(S4),
+    ?CHK_TERM_PROC_LIST(MustChk, 3),
+    ResumeProcessesProc(S2),
+    ?CHK_TERM_PROC_LIST(MustChk, 0),
+    Exit(S1),
+    Exit(S2),
+    Exit(S3),
+    Exit(S4),
+
+
+    HA4 = spawn_link(HangAround),
+    HA5 = spawn_link(HangAround),
+    HA6 = spawn_link(HangAround),
+    S5 = SpawnSuspendProcessesProc(),
+    ?CHK_TERM_PROC_LIST(MustChk, 1),
+    Exit(HA4),
+    ?CHK_TERM_PROC_LIST(MustChk, 2),
+    S6 = SpawnSuspendProcessesProc(),
+    ?CHK_TERM_PROC_LIST(MustChk, 3),
+    Exit(HA5),
+    ?CHK_TERM_PROC_LIST(MustChk, 4),
+    S7 = SpawnSuspendProcessesProc(),
+    ?CHK_TERM_PROC_LIST(MustChk, 5),
+    Exit(HA6),
+    ?CHK_TERM_PROC_LIST(MustChk, 6),
+    S8 = SpawnSuspendProcessesProc(),
+    ?CHK_TERM_PROC_LIST(MustChk, 7),
+
+    erlang:system_flag(multi_scheduling, block),
+    Exit(S8),
+    ?CHK_TERM_PROC_LIST(MustChk, 7),
+    Exit(S5),
+    ?CHK_TERM_PROC_LIST(MustChk, 6),
+    Exit(S7),
+    ?CHK_TERM_PROC_LIST(MustChk, 6),
+    Exit(S6),
+    ?CHK_TERM_PROC_LIST(MustChk, 0),
+    erlang:system_flag(multi_scheduling, unblock),
+    as_expected.
 
 
 otp_7738_waiting(doc) ->
@@ -2115,88 +2070,88 @@ otp_7738_waiting(doc) ->
 otp_7738_waiting(suite) ->
     [];
 otp_7738_waiting(Config) when is_list(Config) ->
-    ?line otp_7738_test(waiting).
+    otp_7738_test(waiting).
 
 otp_7738_suspended(doc) ->
     [];
 otp_7738_suspended(suite) ->
     [];
 otp_7738_suspended(Config) when is_list(Config) ->
-    ?line otp_7738_test(suspended).
+    otp_7738_test(suspended).
 
 otp_7738_resume(doc) ->
     [];
 otp_7738_resume(suite) ->
     [];
 otp_7738_resume(Config) when is_list(Config) ->
-    ?line otp_7738_test(resume).
+    otp_7738_test(resume).
 
 otp_7738_test(Type) ->
-    ?line T = self(),
-    ?line S = spawn_link(fun () ->
-			   receive
-			       {suspend, Suspendee} ->
-				   erlang:suspend_process(Suspendee),
-				   T ! {suspended, Suspendee},
-				   receive
-				   after 10 ->
-					   erlang:resume_process(Suspendee),
-					   Suspendee ! wake_up
-				   end;
-			       {send, To, Msg} ->
-				   receive after 10 -> ok end,
-				   To ! Msg
-			   end
-			 end),
-    ?line R = spawn_link(fun () ->
-				 X = lists:seq(1, 20000000),
-				 T ! {initialized, self()},
-				 ?line case Type of
-					   _ when Type == suspended;
-						  Type == waiting ->
-					       receive _ -> ok end;
-					   _ when Type == resume ->
-					       Receive = fun (F) ->
-								 receive
-								     _ ->
-									 ok
-								 after 0 ->
-									 F(F)
-								 end
-							 end,
-					       Receive(Receive)
-				       end,
-				 T ! {woke_up, self()},
-				 id(X)
-			 end),
-    ?line receive {initialized, R} -> ok end,
-    ?line receive after 10 -> ok end,
-    ?line case Type of
+    T = self(),
+    S = spawn_link(fun () ->
+		receive
+		    {suspend, Suspendee} ->
+			erlang:suspend_process(Suspendee),
+			T ! {suspended, Suspendee},
+			receive
+			after 10 ->
+				erlang:resume_process(Suspendee),
+				Suspendee ! wake_up
+			end;
+		    {send, To, Msg} ->
+			receive after 10 -> ok end,
+			To ! Msg
+		end
+	end),
+    R = spawn_link(fun () ->
+		X = lists:seq(1, 20000000),
+		T ! {initialized, self()},
+		case Type of
+		    _ when Type == suspended;
+		Type == waiting ->
+		    receive _ -> ok end;
+		_ when Type == resume ->
+		    Receive = fun (F) ->
+			    receive
+				_ ->
+				    ok
+			    after 0 ->
+				    F(F)
+			    end
+		    end,
+		    Receive(Receive)
+	    end,
+	    T ! {woke_up, self()},
+	    id(X)
+    end),
+    receive {initialized, R} -> ok end,
+    receive after 10 -> ok end,
+    case Type of
 	      suspended ->
-		  ?line erlang:suspend_process(R),
-		  ?line S ! {send, R, wake_up};
+		  erlang:suspend_process(R),
+		  S ! {send, R, wake_up};
 	      waiting ->
-		  ?line S ! {send, R, wake_up};
+		  S ! {send, R, wake_up};
 	      resume ->
-		  ?line S ! {suspend, R},
-		  ?line receive {suspended, R} -> ok end
+		  S ! {suspend, R},
+		  receive {suspended, R} -> ok end
 	  end,
-    ?line erlang:garbage_collect(R),
-    ?line case Type of
+    erlang:garbage_collect(R),
+    case Type of
 	      suspended ->
-		  ?line erlang:resume_process(R);
+		  erlang:resume_process(R);
 	      _ ->
-		  ?line ok
+		  ok
 	  end,
-    ?line receive
+    receive
 	      {woke_up, R} ->
-		  ?line ok
+		  ok
 	  after 2000 ->
-		  ?line I = process_info(R, [status, message_queue_len]),
-		  ?line ?t:format("~p~n", [I]),
-		  ?line ?t:fail(no_progress)
+		  I = process_info(R, [status, message_queue_len]),
+		  ?t:format("~p~n", [I]),
+		  ?t:fail(no_progress)
 	  end,
-    ?line ok.
+    ok.
 
 gor(Reds, Stop) ->
     receive
@@ -2210,28 +2165,28 @@ gor(Reds, Stop) ->
     end.
 
 garb_other_running(Config) when is_list(Config) ->
-    ?line Stop = make_ref(),
-    ?line {Pid, Mon} = spawn_monitor(fun () -> gor(0, Stop) end),
-    ?line Reds = lists:foldl(fun (_, OldReds) ->
-				     ?line erlang:garbage_collect(Pid),
-				     ?line receive after 1 -> ok end,
-				     ?line Pid ! {self(), reds},
-				     ?line receive
+    Stop = make_ref(),
+    {Pid, Mon} = spawn_monitor(fun () -> gor(0, Stop) end),
+    Reds = lists:foldl(fun (_, OldReds) ->
+				     erlang:garbage_collect(Pid),
+				     receive after 1 -> ok end,
+				     Pid ! {self(), reds},
+				     receive
 					       {reds, NewReds, Pid} ->
-						   ?line true = (NewReds > OldReds),
-						   ?line NewReds
+						   true = (NewReds > OldReds),
+						   NewReds
 					   end
 			     end,
 			     0,
 			     lists:seq(1, 10000)),
-    ?line receive after 1 -> ok end,
-    ?line Pid ! {self(), Stop},
-    ?line receive
+    receive after 1 -> ok end,
+    Pid ! {self(), Stop},
+    receive
 	      {stopped, Stop, StopReds, Pid} ->
-		  ?line true = (StopReds > Reds)
+		  true = (StopReds > Reds)
 	  end,
-    ?line receive {'DOWN', Mon, process, Pid, normal} -> ok end,
-    ?line ok.
+    receive {'DOWN', Mon, process, Pid, normal} -> ok end,
+    ok.
 
 %% Internal functions
 
@@ -2255,9 +2210,9 @@ start_node(Config) ->
     start_node(Config, "").
 
 start_node(Config, Args) when is_list(Config) ->
-    ?line Pa = filename:dirname(code:which(?MODULE)),
-    ?line {A, B, C} = now(),
-    ?line Name = list_to_atom(atom_to_list(?MODULE)
+    Pa = filename:dirname(code:which(?MODULE)),
+    {A, B, C} = now(),
+    Name = list_to_atom(atom_to_list(?MODULE)
 			      ++ "-"
 			      ++ atom_to_list(?config(testcase, Config))
 			      ++ "-"
@@ -2266,7 +2221,7 @@ start_node(Config, Args) when is_list(Config) ->
 			      ++ integer_to_list(B)
 			      ++ "-"
 			      ++ integer_to_list(C)),
-    ?line ?t:start_node(Name, slave, [{args, "-pa "++Pa++" "++Args}]).
+    ?t:start_node(Name, slave, [{args, "-pa "++Pa++" "++Args}]).
 
 stop_node(Node) ->
     ?t:stop_node(Node).