Compact Regions

This brings in initial support for compact regions, as described in the
ICFP 2015 paper "Efficient Communication and Collection with Compact
Normal Forms" (Edward Z. Yang et.al.) and implemented by Giovanni
Campagna.

Some things may change before the 8.2 release, but I (Simon M.) wanted
to get the main patch committed so that we can iterate.

What documentation there is is in the Data.Compact module in the new
compact package.  We'll need to extend and polish the documentation
before the release.

Test Plan:
validate
(new test cases included)

Reviewers: ezyang, simonmar, hvr, bgamari, austin

Subscribers: vikraman, Yuras, RyanGlScott, qnikst, mboes, facundominguez, rrnewton, thomie, erikd

Differential Revision: https://phabricator.haskell.org/D1264

GHC Trac Issues: #11493

9 files changed, 1653 insertions, 12 deletions

diff --git a/rts/sm/BlockAlloc.c b/rts/sm/BlockAlloc.c
index 6c2e96414e..c729c1874f 100644
--- a/rts/sm/BlockAlloc.c
+++ b/rts/sm/BlockAlloc.c
@@ -795,6 +795,7 @@ countAllocdBlocks(bdescr *bd)
     W_ n;
     for (n=0; bd != NULL; bd=bd->link) {
         n += bd->blocks;
+
         // hack for megablock groups: see (*1) above
         if (bd->blocks > BLOCKS_PER_MBLOCK) {
             n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
diff --git a/rts/sm/CNF.c b/rts/sm/CNF.c
new file mode 100644
index 0000000000..3c681c2ee2
--- /dev/null
+++ b/rts/sm/CNF.c
@@ -0,0 +1,1352 @@
+/* -----------------------------------------------------------------------------
+ *
+ * (c) The GHC Team 1998-2014
+ *
+ * GC support for immutable non-GCed structures, also known as Compact
+ * Normal Forms (CNF for short). This provides the RTS support for
+ * the 'compact' package and the Data.Compact module.
+ *
+ * ---------------------------------------------------------------------------*/
+
+#define _GNU_SOURCE
+
+#include "PosixSource.h"
+#include <string.h>
+#include "Rts.h"
+#include "RtsUtils.h"
+
+#include "Capability.h"
+#include "GC.h"
+#include "Storage.h"
+#include "CNF.h"
+#include "Hash.h"
+#include "HeapAlloc.h"
+#include "BlockAlloc.h"
+
+#ifdef HAVE_UNISTD_H
+#include <unistd.h>
+#endif
+#ifdef HAVE_LIMITS_H
+#include <limits.h>
+#endif
+#include <dlfcn.h>
+#include <endian.h>
+
+/**
+ * Note [Compact Normal Forms]
+ *
+ * A Compact Normal Form, is at its essence a chain of memory blocks (multiple
+ * of block allocator blocks) containing other closures inside.
+ *
+ * Each block starts with a header, of type StgCompactNFDataBlock, that points
+ * to the first and to the next block in the chain. Right after the header
+ * in the first block we have a closure of type StgCompactNFData, which holds
+ * compact-wide metadata. This closure is the Compact# that Cmm and Haskell
+ * see, and it's mostly a regular Haskell closure.
+ *
+ * Blocks are appended to the chain automatically as needed, or manually with a
+ * compactResize() call, which also adjust the size of automatically appended
+ * blocks.
+ *
+ * Objects can be appended to the block currently marked to the nursery, or any
+ * of the later blocks if the nursery block is too full to fit the entire
+ * object. For each block in the chain (which can be multiple block allocator
+ * blocks), we use the bdescr of its beginning to store how full it is.
+ * After an object is appended, it is scavenged for any outgoing pointers,
+ * and all pointed to objects are appended, recursively, in a manner similar
+ * to copying GC (further discussion in the note [Appending to a Compact])
+ *
+ * We also flag each bdescr in each block allocator block of a compact
+ * (including those there were obtained as second or later from a single
+ * allocGroup(n) call) with the BF_COMPACT. This allows the GC to quickly
+ * realize that a given pointer is in a compact region, and trigger the
+ * CNF path.
+ *
+ * These two facts combined mean that in any compact block where some object
+ * begins bdescrs must be valid. For this simplicity this is achieved by
+ * restricting the maximum size of a compact block to 252 block allocator
+ * blocks (so that the total with the bdescr is one megablock).
+ *
+ * Compacts as a whole live in special list in each generation, where the
+ * list is held through the bd->link field of the bdescr of the StgCompactNFData
+ * closure (as for large objects). They live in a different list than large
+ * objects because the operation to free them is different (all blocks in
+ * a compact must be freed individually), and stats/sanity behavior are
+ * slightly different. This is also the reason that compact allocates memory
+ * using a special function instead of just calling allocate().
+ *
+ * Compacts are also suitable for network or disk serialization, and to
+ * that extent they support a pointer fixup operation, which adjusts pointers
+ * from a previous layout of the chain in memory to the new allocation.
+ * This works by constructing a temporary binary search table (in the C heap)
+ * of the old block addresses (which are known from the block header), and
+ * then searching for each pointer in the table, and adjusting it.
+ * It relies on ABI compatibility and static linking (or no ASLR) because it
+ * does not attempt to reconstruct info tables, and uses info tables to detect
+ * pointers. In practice this means only the exact same binary should be
+ * used.
+ */
+
+typedef enum {
+    ALLOCATE_APPEND,
+    ALLOCATE_NEW,
+    ALLOCATE_IMPORT_NEW,
+    ALLOCATE_IMPORT_APPEND,
+} AllocateOp;
+
+static StgCompactNFDataBlock *
+compactAllocateBlockInternal(Capability            *cap,
+                             StgWord                aligned_size,
+                             StgCompactNFDataBlock *first,
+                             AllocateOp             operation)
+{
+    StgCompactNFDataBlock *self;
+    bdescr *block, *head;
+    uint32_t n_blocks;
+    generation *g;
+
+    n_blocks = aligned_size / BLOCK_SIZE;
+
+    // Attempting to allocate an object larger than maxHeapSize
+    // should definitely be disallowed.  (bug #1791)
+    if ((RtsFlags.GcFlags.maxHeapSize > 0 &&
+         n_blocks >= RtsFlags.GcFlags.maxHeapSize) ||
+        n_blocks >= HS_INT32_MAX)   // avoid overflow when
+                                    // calling allocGroup() below
+    {
+        heapOverflow();
+        // heapOverflow() doesn't exit (see #2592), but we aren't
+        // in a position to do a clean shutdown here: we
+        // either have to allocate the memory or exit now.
+        // Allocating the memory would be bad, because the user
+        // has requested that we not exceed maxHeapSize, so we
+        // just exit.
+        stg_exit(EXIT_HEAPOVERFLOW);
+    }
+
+    // It is imperative that first is the first block in the compact
+    // (or NULL if the compact does not exist yet)
+    // because the evacuate code does not update the generation of
+    // blocks other than the first (so we would get the statistics
+    // wrong and crash in Sanity)
+    if (first != NULL) {
+        block = Bdescr((P_)first);
+        g = block->gen;
+    } else {
+        g = g0;
+    }
+
+    ACQUIRE_SM_LOCK;
+    block = allocGroup(n_blocks);
+    switch (operation) {
+    case ALLOCATE_NEW:
+        ASSERT (first == NULL);
+        ASSERT (g == g0);
+        dbl_link_onto(block, &g0->compact_objects);
+        g->n_compact_blocks += block->blocks;
+        g->n_new_large_words += aligned_size / sizeof(StgWord);
+        break;
+
+    case ALLOCATE_IMPORT_NEW:
+        dbl_link_onto(block, &g0->compact_blocks_in_import);
+        /* fallthrough */
+    case ALLOCATE_IMPORT_APPEND:
+        ASSERT (first == NULL);
+        ASSERT (g == g0);
+        g->n_compact_blocks_in_import += block->blocks;
+        g->n_new_large_words += aligned_size / sizeof(StgWord);
+        break;
+
+    case ALLOCATE_APPEND:
+        g->n_compact_blocks += block->blocks;
+        if (g == g0)
+            g->n_new_large_words += aligned_size / sizeof(StgWord);
+        break;
+
+    default:
+#ifdef DEBUG
+        ASSERT(!"code should not be reached");
+#else
+        __builtin_unreachable();
+#endif
+    }
+    RELEASE_SM_LOCK;
+
+    cap->total_allocated += aligned_size / sizeof(StgWord);
+
+    self = (StgCompactNFDataBlock*) block->start;
+    self->self = self;
+    self->next = NULL;
+
+    head = block;
+    initBdescr(head, g, g);
+    head->flags = BF_COMPACT;
+    for (block = head + 1, n_blocks --; n_blocks > 0; block++, n_blocks--) {
+        block->link = head;
+        block->blocks = 0;
+        block->flags = BF_COMPACT;
+    }
+
+    return self;
+}
+
+static inline StgCompactNFDataBlock *
+compactGetFirstBlock(StgCompactNFData *str)
+{
+    return (StgCompactNFDataBlock*) ((W_)str - sizeof(StgCompactNFDataBlock));
+}
+
+static inline StgCompactNFData *
+firstBlockGetCompact(StgCompactNFDataBlock *block)
+{
+    return (StgCompactNFData*) ((W_)block + sizeof(StgCompactNFDataBlock));
+}
+
+static void
+freeBlockChain(StgCompactNFDataBlock *block)
+{
+    StgCompactNFDataBlock *next;
+    bdescr *bd;
+
+    for ( ; block; block = next) {
+        next = block->next;
+        bd = Bdescr((StgPtr)block);
+        ASSERT((bd->flags & BF_EVACUATED) == 0);
+        freeGroup(bd);
+    }
+}
+
+void
+compactFree(StgCompactNFData *str)
+{
+    StgCompactNFDataBlock *block;
+
+    block = compactGetFirstBlock(str);
+    freeBlockChain(block);
+}
+
+void
+compactMarkKnown(StgCompactNFData *str)
+{
+    bdescr *bd;
+    StgCompactNFDataBlock *block;
+
+    block = compactGetFirstBlock(str);
+    for ( ; block; block = block->next) {
+        bd = Bdescr((StgPtr)block);
+        bd->flags |= BF_KNOWN;
+    }
+}
+
+StgWord
+countCompactBlocks(bdescr *outer)
+{
+    StgCompactNFDataBlock *block;
+    W_ count;
+
+    count = 0;
+    while (outer) {
+        bdescr *inner;
+
+        block = (StgCompactNFDataBlock*)(outer->start);
+        do {
+            inner = Bdescr((P_)block);
+            ASSERT (inner->flags & BF_COMPACT);
+
+            count += inner->blocks;
+            block = block->next;
+        } while(block);
+
+        outer = outer->link;
+    }
+
+    return count;
+}
+
+StgCompactNFData *
+compactNew (Capability *cap, StgWord size)
+{
+    StgWord aligned_size;
+    StgCompactNFDataBlock *block;
+    StgCompactNFData *self;
+    bdescr *bd;
+
+    aligned_size = BLOCK_ROUND_UP(size + sizeof(StgCompactNFDataBlock)
+                                  + sizeof(StgCompactNFDataBlock));
+    if (aligned_size >= BLOCK_SIZE * BLOCKS_PER_MBLOCK)
+        aligned_size = BLOCK_SIZE * BLOCKS_PER_MBLOCK;
+
+    block = compactAllocateBlockInternal(cap, aligned_size, NULL,
+                                         ALLOCATE_NEW);
+
+    self = firstBlockGetCompact(block);
+    SET_INFO((StgClosure*)self, &stg_COMPACT_NFDATA_info);
+    self->totalDataW = aligned_size / sizeof(StgWord);
+    self->autoBlockW = aligned_size / sizeof(StgWord);
+    self->nursery = block;
+    self->last = block;
+
+    block->owner = self;
+
+    bd = Bdescr((P_)block);
+    bd->free = (StgPtr)((W_)self + sizeof(StgCompactNFData));
+    ASSERT (bd->free == (StgPtr)self + sizeofW(StgCompactNFData));
+
+    self->totalW = bd->blocks * BLOCK_SIZE_W;
+
+    return self;
+}
+
+static StgCompactNFDataBlock *
+compactAppendBlock (Capability       *cap,
+                    StgCompactNFData *str,
+                    StgWord           aligned_size)
+{
+    StgCompactNFDataBlock *block;
+    bdescr *bd;
+
+    block = compactAllocateBlockInternal(cap, aligned_size,
+                                         compactGetFirstBlock(str),
+                                         ALLOCATE_APPEND);
+    block->owner = str;
+    block->next = NULL;
+
+    ASSERT (str->last->next == NULL);
+    str->last->next = block;
+    str->last = block;
+    if (str->nursery == NULL)
+        str->nursery = block;
+    str->totalDataW += aligned_size / sizeof(StgWord);
+
+    bd = Bdescr((P_)block);
+    bd->free = (StgPtr)((W_)block + sizeof(StgCompactNFDataBlock));
+    ASSERT (bd->free == (StgPtr)block + sizeofW(StgCompactNFDataBlock));
+
+    str->totalW += bd->blocks * BLOCK_SIZE_W;
+
+    return block;
+}
+
+void
+compactResize (Capability *cap, StgCompactNFData *str, StgWord new_size)
+{
+    StgWord aligned_size;
+
+    aligned_size = BLOCK_ROUND_UP(new_size + sizeof(StgCompactNFDataBlock));
+    if (aligned_size >= BLOCK_SIZE * BLOCKS_PER_MBLOCK)
+        aligned_size = BLOCK_SIZE * BLOCKS_PER_MBLOCK;
+
+    str->autoBlockW = aligned_size / sizeof(StgWord);
+
+    compactAppendBlock(cap, str, aligned_size);
+}
+
+/* Note [Appending to a Compact]
+
+   This is a simple reimplementation of the copying GC.
+   One could be tempted to reuse the actual GC code here, but he
+   would quickly find out that it would bring all the generational
+   GC complexity for no need at all.
+
+   Plus, we don't need to scavenge/evacuate all kinds of weird
+   objects here, just constructors and primitives. Thunks are
+   expected to be evaluated before appending by the API layer
+   (in Haskell, above the primop which is implemented here).
+   Also, we have a different policy for large objects: instead
+   of relinking to the new large object list, we fully copy
+   them inside the compact and scavenge them normally.
+
+   Note that if we allowed thunks and lazy evaluation the compact
+   would be a mutable object, which would create all sorts of
+   GC problems (besides, evaluating a thunk could exaust the
+   compact space or yield an invalid object, and we would have
+   no way to signal that to the user)
+
+   Just like the real evacuate/scavenge pairs, we need to handle
+   object loops. We would want to use the same strategy of rewriting objects
+   with forwarding pointer, but in a real GC, at the end the
+   blocks from the old space are dropped (dropping all forwarding
+   pointers at the same time), which we can't do here as we don't
+   know all pointers to the objects being evacuated. Also, in parallel
+   we don't know which other threads are evaluating the thunks
+   that we just corrupted at the same time.
+
+   So instead we use a hash table of "visited" objects, and add
+   the pointer as we copy it. To reduce the overhead, we also offer
+   a version of the API that does not preserve sharing (TODO).
+
+   You might be tempted to replace the objects with StdInd to
+   the object in the compact, but you would be wrong: the haskell
+   code assumes that objects in the heap only become more evaluated
+   (thunks to blackholes to inds to actual objects), and in
+   particular it assumes that if a pointer is tagged the object
+   is directly referenced and the values can be read directly,
+   without entering the closure.
+
+   FIXME: any better idea than the hash table?
+*/
+
+static void
+unroll_memcpy(StgPtr to, StgPtr from, StgWord size)
+{
+    for (; size > 0; size--)
+        *(to++) = *(from++);
+}
+
+static rtsBool
+allocate_in_compact (StgCompactNFDataBlock *block, StgWord sizeW, StgPtr *at)
+{
+    bdescr *bd;
+    StgPtr top;
+    StgPtr free;
+
+    bd = Bdescr((StgPtr)block);
+    top = bd->start + BLOCK_SIZE_W * bd->blocks;
+    if (bd->free + sizeW > top)
+        return rtsFalse;
+
+    free = bd->free;
+    bd->free += sizeW;
+    *at = free;
+
+    return rtsTrue;
+}
+
+static rtsBool
+block_is_full (StgCompactNFDataBlock *block)
+{
+    bdescr *bd;
+    StgPtr top;
+    StgWord sizeW;
+
+    bd = Bdescr((StgPtr)block);
+    top = bd->start + BLOCK_SIZE_W * bd->blocks;
+
+    // We consider a block full if we could not fit
+    // an entire closure with 7 payload items
+    // (this leaves a slop of 64 bytes at most, but
+    // it avoids leaving a block almost empty to fit
+    // a large byte array, while at the same time
+    // it avoids trying to allocate a large closure
+    // in a chain of almost empty blocks)
+    sizeW = sizeofW(StgHeader) + 7;
+    return (bd->free + sizeW > top);
+}
+
+static inline StgWord max(StgWord a, StgWord b)
+{
+    if (a > b)
+        return a;
+    else
+        return b;
+}
+
+static rtsBool
+allocate_loop (Capability       *cap,
+               StgCompactNFData *str,
+               StgWord           sizeW,
+               StgPtr           *at)
+{
+    StgCompactNFDataBlock *block;
+    StgWord next_size;
+
+    // try the nursery first
+ retry:
+    if (str->nursery != NULL) {
+        if (allocate_in_compact(str->nursery, sizeW, at))
+            return rtsTrue;
+
+        if (block_is_full (str->nursery)) {
+            str->nursery = str->nursery->next;
+            goto retry;
+        }
+
+        // try subsequent blocks
+        block = str->nursery->next;
+        while (block != NULL) {
+            if (allocate_in_compact(block, sizeW, at))
+                return rtsTrue;
+
+            block = block->next;
+        }
+    }
+
+    next_size = max(str->autoBlockW * sizeof(StgWord),
+                    BLOCK_ROUND_UP(sizeW * sizeof(StgWord)));
+    if (next_size >= BLOCKS_PER_MBLOCK * BLOCK_SIZE)
+        next_size = BLOCKS_PER_MBLOCK * BLOCK_SIZE;
+    if (next_size < sizeW * sizeof(StgWord) + sizeof(StgCompactNFDataBlock))
+        return rtsFalse;
+
+    block = compactAppendBlock(cap, str, next_size);
+    ASSERT (str->nursery != NULL);
+    return allocate_in_compact(block, sizeW, at);
+}
+
+static void
+copy_tag (Capability        *cap,
+          StgCompactNFData  *str,
+          HashTable         *hash,
+          StgClosure       **p,
+          StgClosure        *from,
+          StgWord            tag)
+{
+    StgPtr to;
+    StgWord sizeW;
+
+    sizeW = closure_sizeW(from);
+
+    if (!allocate_loop(cap, str, sizeW, &to)) {
+        barf("Failed to copy object in compact, object too large\n");
+        return;
+    }
+
+    // unroll memcpy for small sizes because we can
+    // benefit of known alignment
+    // (32 extracted from my magic hat)
+    if (sizeW < 32)
+        unroll_memcpy(to, (StgPtr)from, sizeW);
+    else
+        memcpy(to, from, sizeW * sizeof(StgWord));
+
+    if (hash != NULL)
+        insertHashTable(hash, (StgWord)from, to);
+
+    *p = TAG_CLOSURE(tag, (StgClosure*)to);
+}
+
+STATIC_INLINE rtsBool
+object_in_compact (StgCompactNFData *str, StgClosure *p)
+{
+    bdescr *bd;
+
+    if (!HEAP_ALLOCED(p))
+        return rtsFalse;
+
+    bd = Bdescr((P_)p);
+    return (bd->flags & BF_COMPACT) != 0 &&
+        objectGetCompact(p) == str;
+}
+
+static void
+simple_evacuate (Capability        *cap,
+                 StgCompactNFData  *str,
+                 HashTable         *hash,
+                 StgClosure       **p)
+{
+    StgWord tag;
+    StgClosure *from;
+    void *already;
+
+    from = *p;
+    tag = GET_CLOSURE_TAG(from);
+    from = UNTAG_CLOSURE(from);
+
+    // If the object referenced is already in this compact
+    // (for example by reappending an object that was obtained
+    // by compactGetRoot) then do nothing
+    if (object_in_compact(str, from))
+        return;
+
+    switch (get_itbl(from)->type) {
+    case BLACKHOLE:
+        // If tag == 0, the indirectee is the TSO that claimed the tag
+        //
+        // Not useful and not NFData
+        from = ((StgInd*)from)->indirectee;
+        if (GET_CLOSURE_TAG(from) == 0) {
+            debugBelch("Claimed but not updated BLACKHOLE in Compact,"
+                       " not normal form");
+            return;
+        }
+
+        *p = from;
+        return simple_evacuate(cap, str, hash, p);
+
+    case IND:
+    case IND_STATIC:
+        // follow chains of indirections, don't evacuate them
+        from = ((StgInd*)from)->indirectee;
+        *p = from;
+        // Evac.c uses a goto, but let's rely on a smart compiler
+        // and get readable code instead
+        return simple_evacuate(cap, str, hash, p);
+
+    default:
+        // This object was evacuated already, return the existing
+        // pointer
+        if (hash != NULL &&
+            (already = lookupHashTable (hash, (StgWord)from))) {
+            *p = TAG_CLOSURE(tag, (StgClosure*)already);
+            return;
+        }
+
+        copy_tag(cap, str, hash, p, from, tag);
+    }
+}
+
+static void
+simple_scavenge_mut_arr_ptrs (Capability       *cap,
+                              StgCompactNFData *str,
+                              HashTable        *hash,
+                              StgMutArrPtrs    *a)
+{
+    StgPtr p, q;
+
+    p = (StgPtr)&a->payload[0];
+    q = (StgPtr)&a->payload[a->ptrs];
+    for (; p < q; p++) {
+        simple_evacuate(cap, str, hash, (StgClosure**)p);
+    }
+}
+
+static void
+simple_scavenge_block (Capability            *cap,
+                       StgCompactNFData      *str,
+                       StgCompactNFDataBlock *block,
+                       HashTable             *hash,
+                       StgPtr                 p)
+{
+    const StgInfoTable *info;
+    bdescr *bd = Bdescr((P_)block);
+
+    while (p < bd->free) {
+        ASSERT (LOOKS_LIKE_CLOSURE_PTR(p));
+        info = get_itbl((StgClosure*)p);
+
+        switch (info->type) {
+        case CONSTR_1_0:
+            simple_evacuate(cap, str, hash, &((StgClosure*)p)->payload[0]);
+        case CONSTR_0_1:
+            p += sizeofW(StgClosure) + 1;
+            break;
+
+        case CONSTR_2_0:
+            simple_evacuate(cap, str, hash, &((StgClosure*)p)->payload[1]);
+        case CONSTR_1_1:
+            simple_evacuate(cap, str, hash, &((StgClosure*)p)->payload[0]);
+        case CONSTR_0_2:
+            p += sizeofW(StgClosure) + 2;
+            break;
+
+        case CONSTR:
+        case PRIM:
+        case CONSTR_NOCAF_STATIC:
+        case CONSTR_STATIC:
+        {
+            StgPtr end;
+
+            end = (P_)((StgClosure *)p)->payload + info->layout.payload.ptrs;
+            for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
+                simple_evacuate(cap, str, hash, (StgClosure **)p);
+            }
+            p += info->layout.payload.nptrs;
+            break;
+        }
+
+        case ARR_WORDS:
+            p += arr_words_sizeW((StgArrBytes*)p);
+            break;
+
+        case MUT_ARR_PTRS_FROZEN:
+        case MUT_ARR_PTRS_FROZEN0:
+            simple_scavenge_mut_arr_ptrs(cap, str, hash, (StgMutArrPtrs*)p);
+            p += mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
+            break;
+
+        case SMALL_MUT_ARR_PTRS_FROZEN:
+        case SMALL_MUT_ARR_PTRS_FROZEN0:
+        {
+            uint32_t i;
+            StgSmallMutArrPtrs *arr = (StgSmallMutArrPtrs*)p;
+
+            for (i = 0; i < arr->ptrs; i++)
+                simple_evacuate(cap, str, hash, &arr->payload[i]);
+
+            p += sizeofW(StgSmallMutArrPtrs) + arr->ptrs;
+            break;
+        }
+
+        case IND:
+        case BLACKHOLE:
+        case IND_STATIC:
+            // They get shortcircuited by simple_evaluate()
+            barf("IND/BLACKHOLE in Compact");
+            break;
+
+        default:
+            barf("Invalid non-NFData closure in Compact\n");
+        }
+    }
+}
+
+static void
+scavenge_loop (Capability            *cap,
+               StgCompactNFData      *str,
+               StgCompactNFDataBlock *first_block,
+               HashTable             *hash,
+               StgPtr                 p)
+{
+    // Scavenge the first block
+    simple_scavenge_block(cap, str, first_block, hash, p);
+
+    // Note: simple_scavenge_block can change str->last, which
+    // changes this check, in addition to iterating through
+    while (first_block != str->last) {
+        // we can't allocate in blocks that were already scavenged
+        // so push the nursery forward
+        if (str->nursery == first_block)
+            str->nursery = str->nursery->next;
+
+        first_block = first_block->next;
+        simple_scavenge_block(cap, str, first_block, hash,
+                              (P_)first_block + sizeofW(StgCompactNFDataBlock));
+    }
+}
+
+#ifdef DEBUG
+static rtsBool
+objectIsWHNFData (StgClosure *what)
+{
+    switch (get_itbl(what)->type) {
+    case CONSTR:
+    case CONSTR_1_0:
+    case CONSTR_0_1:
+    case CONSTR_2_0:
+    case CONSTR_1_1:
+    case CONSTR_0_2:
+    case CONSTR_STATIC:
+    case CONSTR_NOCAF_STATIC:
+    case ARR_WORDS:
+    case MUT_ARR_PTRS_FROZEN:
+    case MUT_ARR_PTRS_FROZEN0:
+    case SMALL_MUT_ARR_PTRS_FROZEN:
+    case SMALL_MUT_ARR_PTRS_FROZEN0:
+        return rtsTrue;
+
+    case IND:
+    case BLACKHOLE:
+        return objectIsWHNFData(UNTAG_CLOSURE(((StgInd*)what)->indirectee));
+
+    default:
+        return rtsFalse;
+    }
+}
+
+static rtsBool
+verify_mut_arr_ptrs (StgCompactNFData *str,
+                     StgMutArrPtrs    *a)
+{
+    StgPtr p, q;
+
+    p = (StgPtr)&a->payload[0];
+    q = (StgPtr)&a->payload[a->ptrs];
+    for (; p < q; p++) {
+        if (!object_in_compact(str, UNTAG_CLOSURE(*(StgClosure**)p)))
+            return rtsFalse;
+    }
+
+    return rtsTrue;
+}
+
+static rtsBool
+verify_consistency_block (StgCompactNFData *str, StgCompactNFDataBlock *block)
+{
+    bdescr *bd;
+    StgPtr p;
+    const StgInfoTable *info;
+    StgClosure *q;
+
+    p = (P_)firstBlockGetCompact(block);
+    bd = Bdescr((P_)block);
+    while (p < bd->free) {
+        q = (StgClosure*)p;
+
+        if (!LOOKS_LIKE_CLOSURE_PTR(q))
+            return rtsFalse;
+
+        info = get_itbl(q);
+        switch (info->type) {
+        case CONSTR_1_0:
+            if (!object_in_compact(str, UNTAG_CLOSURE(q->payload[0])))
+                return rtsFalse;
+        case CONSTR_0_1:
+            p += sizeofW(StgClosure) + 1;
+            break;
+
+        case CONSTR_2_0:
+            if (!object_in_compact(str, UNTAG_CLOSURE(q->payload[1])))
+                return rtsFalse;
+        case CONSTR_1_1:
+            if (!object_in_compact(str, UNTAG_CLOSURE(q->payload[0])))
+                return rtsFalse;
+        case CONSTR_0_2:
+            p += sizeofW(StgClosure) + 2;
+            break;
+
+        case CONSTR:
+        case PRIM:
+        case CONSTR_STATIC:
+        case CONSTR_NOCAF_STATIC:
+        {
+            uint32_t i;
+
+            for (i = 0; i < info->layout.payload.ptrs; i++)
+                if (!object_in_compact(str, UNTAG_CLOSURE(q->payload[i])))
+                    return rtsFalse;
+
+            p += sizeofW(StgClosure) + info->layout.payload.ptrs +
+                info->layout.payload.nptrs;
+            break;
+        }
+
+        case ARR_WORDS:
+            p += arr_words_sizeW((StgArrBytes*)p);
+            break;
+
+        case MUT_ARR_PTRS_FROZEN:
+        case MUT_ARR_PTRS_FROZEN0:
+            if (!verify_mut_arr_ptrs(str, (StgMutArrPtrs*)p))
+                return rtsFalse;
+            p += mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
+            break;
+
+        case SMALL_MUT_ARR_PTRS_FROZEN:
+        case SMALL_MUT_ARR_PTRS_FROZEN0:
+        {
+            uint32_t i;
+            StgSmallMutArrPtrs *arr = (StgSmallMutArrPtrs*)p;
+
+            for (i = 0; i < arr->ptrs; i++)
+                if (!object_in_compact(str, UNTAG_CLOSURE(arr->payload[i])))
+                    return rtsFalse;
+
+            p += sizeofW(StgSmallMutArrPtrs) + arr->ptrs;
+            break;
+        }
+
+        case COMPACT_NFDATA:
+            p += sizeofW(StgCompactNFData);
+            break;
+
+        default:
+            return rtsFalse;
+        }
+    }
+
+    return rtsTrue;
+}
+
+static rtsBool
+verify_consistency_loop (StgCompactNFData *str)
+{
+    StgCompactNFDataBlock *block;
+
+    block = compactGetFirstBlock(str);
+    do {
+        if (!verify_consistency_block(str, block))
+            return rtsFalse;
+        block = block->next;
+    } while (block && block->owner);
+
+    return rtsTrue;
+}
+#endif
+
+
+StgPtr
+compactAppend (Capability       *cap,
+               StgCompactNFData *str,
+               StgClosure       *what,
+               StgWord           share)
+{
+    StgClosure *root;
+    StgClosure *tagged_root;
+    HashTable *hash;
+    StgCompactNFDataBlock *evaced_block;
+
+    ASSERT(objectIsWHNFData(UNTAG_CLOSURE(what)));
+
+    tagged_root = what;
+    simple_evacuate(cap, str, NULL, &tagged_root);
+
+    root = UNTAG_CLOSURE(tagged_root);
+    evaced_block = objectGetCompactBlock(root);
+
+    if (share) {
+        hash = allocHashTable ();
+        insertHashTable(hash, (StgWord)UNTAG_CLOSURE(what), root);
+    } else
+        hash = NULL;
+
+    scavenge_loop(cap, str, evaced_block, hash, (P_)root);
+
+    if (share)
+        freeHashTable(hash, NULL);
+
+    ASSERT(verify_consistency_loop(str));
+
+    return (StgPtr)tagged_root;
+}
+
+StgWord
+compactContains (StgCompactNFData *str, StgPtr what)
+{
+    bdescr *bd;
+
+    // This check is the reason why this needs to be
+    // implemented in C instead of (possibly faster) Cmm
+    if (!HEAP_ALLOCED (what))
+        return 0;
+
+    // Note that we don't care about tags, they are eaten
+    // away by the Bdescr operation anyway
+    bd = Bdescr((P_)what);
+    return (bd->flags & BF_COMPACT) != 0 &&
+        (str == NULL || objectGetCompact((StgClosure*)what) == str);
+}
+
+StgCompactNFDataBlock *
+compactAllocateBlock(Capability            *cap,
+                     StgWord                size,
+                     StgCompactNFDataBlock *previous)
+{
+    StgWord aligned_size;
+    StgCompactNFDataBlock *block;
+    bdescr *bd;
+
+    aligned_size = BLOCK_ROUND_UP(size);
+
+    // We do not link the new object into the generation ever
+    // - we cannot let the GC know about this object until we're done
+    // importing it and we have fixed up all info tables and stuff
+    //
+    // but we do update n_compact_blocks, otherwise memInventory()
+    // in Sanity will think we have a memory leak, because it compares
+    // the blocks he knows about with the blocks obtained by the
+    // block allocator
+    // (if by chance a memory leak does happen due to a bug somewhere
+    // else, memInventory will also report that all compact blocks
+    // associated with this compact are leaked - but they are not really,
+    // we have a pointer to them and we're not losing track of it, it's
+    // just we can't use the GC until we're done with the import)
+    //
+    // (That btw means that the high level import code must be careful
+    // not to lose the pointer, so don't use the primops directly
+    // unless you know what you're doing!)
+
+    // Other trickery: we pass NULL as first, which means our blocks
+    // are always in generation 0
+    // This is correct because the GC has never seen the blocks so
+    // it had no chance of promoting them
+
+    block = compactAllocateBlockInternal(cap, aligned_size, NULL,
+                                         previous != NULL ? ALLOCATE_IMPORT_APPEND : ALLOCATE_IMPORT_NEW);
+    if (previous != NULL)
+        previous->next = block;
+
+    bd = Bdescr((P_)block);
+    bd->free = (P_)((W_)bd->start + size);
+
+    return block;
+}
+
+STATIC_INLINE rtsBool
+any_needs_fixup(StgCompactNFDataBlock *block)
+{
+    // ->next pointers are always valid, even if some blocks were
+    // not allocated where we want them, because compactAllocateAt()
+    // will take care to adjust them
+
+    do {
+        if (block->self != block)
+            return rtsTrue;
+        block = block->next;
+    } while (block && block->owner);
+
+    return rtsFalse;
+}
+
+#ifdef DEBUG
+static void
+spew_failing_pointer(StgWord *fixup_table, uint32_t count, StgWord address)
+{
+    uint32_t i;
+    StgWord key, value;
+    StgCompactNFDataBlock *block;
+    bdescr *bd;
+    StgWord size;
+
+    debugBelch("Failed to adjust 0x%lx. Block dump follows...\n",
+               address);
+
+    for (i  = 0; i < count; i++) {
+        key = fixup_table [2 * i];
+        value = fixup_table [2 * i + 1];
+
+        block = (StgCompactNFDataBlock*)value;
+        bd = Bdescr((P_)block);
+        size = (W_)bd->free - (W_)bd->start;
+
+        debugBelch("%d: was 0x%lx-0x%lx, now 0x%lx-0x%lx\n", i,
+                   key, key+size, value, value+size);
+    }
+}
+#endif
+
+STATIC_INLINE StgCompactNFDataBlock *
+find_pointer(StgWord *fixup_table, uint32_t count, StgClosure *q)
+{
+    StgWord address = (W_)q;
+    uint32_t a, b, c;
+    StgWord key, value;
+    bdescr *bd;
+
+    a = 0;
+    b = count;
+    while (a < b-1) {
+        c = (a+b)/2;
+
+        key = fixup_table[c * 2];
+        value = fixup_table[c * 2 + 1];
+
+        if (key > address)
+            b = c;
+        else
+            a = c;
+    }
+
+    // three cases here: 0, 1 or 2 blocks to check
+    for ( ; a < b; a++) {
+        key = fixup_table[a * 2];
+        value = fixup_table[a * 2 + 1];
+
+        if (key > address)
+            goto fail;
+
+        bd = Bdescr((P_)value);
+
+        if (key + bd->blocks * BLOCK_SIZE <= address)
+            goto fail;
+
+        return (StgCompactNFDataBlock*)value;
+    }
+
+ fail:
+    // We should never get here
+
+#ifdef DEBUG
+    spew_failing_pointer(fixup_table, count, address);
+#endif
+    return NULL;
+}
+
+static rtsBool
+fixup_one_pointer(StgWord *fixup_table, uint32_t count, StgClosure **p)
+{
+    StgWord tag;
+    StgClosure *q;
+    StgCompactNFDataBlock *block;
+
+    q = *p;
+    tag = GET_CLOSURE_TAG(q);
+    q = UNTAG_CLOSURE(q);
+
+    block = find_pointer(fixup_table, count, q);
+    if (block == NULL)
+        return rtsFalse;
+    if (block == block->self)
+        return rtsTrue;
+
+    q = (StgClosure*)((W_)q - (W_)block->self + (W_)block);
+    *p = TAG_CLOSURE(tag, q);
+
+    return rtsTrue;
+}
+
+static rtsBool
+fixup_mut_arr_ptrs (StgWord          *fixup_table,
+                    uint32_t               count,
+                    StgMutArrPtrs    *a)
+{
+    StgPtr p, q;
+
+    p = (StgPtr)&a->payload[0];
+    q = (StgPtr)&a->payload[a->ptrs];
+    for (; p < q; p++) {
+        if (!fixup_one_pointer(fixup_table, count, (StgClosure**)p))
+            return rtsFalse;
+    }
+
+    return rtsTrue;
+}
+
+static rtsBool
+fixup_block(StgCompactNFDataBlock *block, StgWord *fixup_table, uint32_t count)
+{
+    const StgInfoTable *info;
+    bdescr *bd;
+    StgPtr p;
+
+    bd = Bdescr((P_)block);
+    p = bd->start + sizeofW(StgCompactNFDataBlock);
+    while (p < bd->free) {
+        ASSERT (LOOKS_LIKE_CLOSURE_PTR(p));
+        info = get_itbl((StgClosure*)p);
+
+        switch (info->type) {
+        case CONSTR_1_0:
+            if (!fixup_one_pointer(fixup_table, count,
+                                   &((StgClosure*)p)->payload[0]))
+                return rtsFalse;
+        case CONSTR_0_1:
+            p += sizeofW(StgClosure) + 1;
+            break;
+
+        case CONSTR_2_0:
+            if (!fixup_one_pointer(fixup_table, count,
+                                   &((StgClosure*)p)->payload[1]))
+                return rtsFalse;
+        case CONSTR_1_1:
+            if (!fixup_one_pointer(fixup_table, count,
+                                   &((StgClosure*)p)->payload[0]))
+                return rtsFalse;
+        case CONSTR_0_2:
+            p += sizeofW(StgClosure) + 2;
+            break;
+
+        case CONSTR:
+        case PRIM:
+        case CONSTR_STATIC:
+        case CONSTR_NOCAF_STATIC:
+        {
+            StgPtr end;
+
+            end = (P_)((StgClosure *)p)->payload + info->layout.payload.ptrs;
+            for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
+                if (!fixup_one_pointer(fixup_table, count, (StgClosure **)p))
+                    return rtsFalse;
+            }
+            p += info->layout.payload.nptrs;
+            break;
+        }
+
+        case ARR_WORDS:
+            p += arr_words_sizeW((StgArrBytes*)p);
+            break;
+
+        case MUT_ARR_PTRS_FROZEN:
+        case MUT_ARR_PTRS_FROZEN0:
+            fixup_mut_arr_ptrs(fixup_table, count, (StgMutArrPtrs*)p);
+            p += mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
+            break;
+
+        case SMALL_MUT_ARR_PTRS_FROZEN:
+        case SMALL_MUT_ARR_PTRS_FROZEN0:
+        {
+            uint32_t i;
+            StgSmallMutArrPtrs *arr = (StgSmallMutArrPtrs*)p;
+
+            for (i = 0; i < arr->ptrs; i++) {
+                if (!fixup_one_pointer(fixup_table, count,
+                                       &arr->payload[i]))
+                    return rtsFalse;
+            }
+
+            p += sizeofW(StgSmallMutArrPtrs) + arr->ptrs;
+            break;
+        }
+
+        case COMPACT_NFDATA:
+            if (p == (bd->start + sizeofW(StgCompactNFDataBlock))) {
+                // Ignore the COMPACT_NFDATA header
+                // (it will be fixed up later)
+                p += sizeofW(StgCompactNFData);
+                break;
+            }
+
+            // fall through
+
+        default:
+            debugBelch("Invalid non-NFData closure (type %d) in Compact\n",
+                       info->type);
+            return rtsFalse;
+        }
+    }
+
+    return rtsTrue;
+}
+
+static int
+cmp_fixup_table_item (const void *e1, const void *e2)
+{
+    const StgWord *w1 = e1;
+    const StgWord *w2 = e2;
+
+    return *w1 - *w2;
+}
+
+static StgWord *
+build_fixup_table (StgCompactNFDataBlock *block, uint32_t *pcount)
+{
+    uint32_t count;
+    StgCompactNFDataBlock *tmp;
+    StgWord *table;
+
+    count = 0;
+    tmp = block;
+    do {
+        count++;
+        tmp = tmp->next;
+    } while(tmp && tmp->owner);
+
+    table = stgMallocBytes(sizeof(StgWord) * 2 * count, "build_fixup_table");
+
+    count = 0;
+    do {
+        table[count * 2] = (W_)block->self;
+        table[count * 2 + 1] = (W_)block;
+        count++;
+        block = block->next;
+    } while(block && block->owner);
+
+    qsort(table, count, sizeof(StgWord) * 2, cmp_fixup_table_item);
+
+    *pcount = count;
+    return table;
+}
+
+static rtsBool
+fixup_loop(StgCompactNFDataBlock *block, StgClosure **proot)
+{
+    StgWord *table;
+    rtsBool ok;
+    uint32_t count;
+
+    table = build_fixup_table (block, &count);
+
+    do {
+        if (!fixup_block(block, table, count)) {
+            ok = rtsFalse;
+            goto out;
+        }
+
+        block = block->next;
+    } while(block && block->owner);
+
+    ok = fixup_one_pointer(table, count, proot);
+
+ out:
+    stgFree(table);
+    return ok;
+}
+
+static void
+fixup_early(StgCompactNFData *str, StgCompactNFDataBlock *block)
+{
+    StgCompactNFDataBlock *last;
+
+    do {
+        last = block;
+        block = block->next;
+    } while(block);
+
+    str->last = last;
+}
+
+static void
+fixup_late(StgCompactNFData *str, StgCompactNFDataBlock *block)
+{
+    StgCompactNFDataBlock *nursery;
+    bdescr *bd;
+    StgWord totalW;
+    StgWord totalDataW;
+
+    nursery = block;
+    totalW = 0;
+    totalDataW = 0;
+    do {
+        block->self = block;
+
+        bd = Bdescr((P_)block);
+        totalW += bd->blocks * BLOCK_SIZE_W;
+
+        if (block->owner != NULL) {
+            if (bd->free != bd->start)
+                nursery = block;
+            block->owner = str;
+            totalDataW += bd->blocks * BLOCK_SIZE_W;
+        }
+
+        block = block->next;
+    } while(block);
+
+    str->nursery = nursery;
+    str->totalW = totalW;
+    str->totalDataW = totalDataW;
+}
+
+static StgClosure *
+maybe_fixup_internal_pointers (StgCompactNFDataBlock *block,
+                               StgClosure            *root)
+{
+    rtsBool ok;
+    StgClosure **proot;
+
+    // Check for fast path
+    if (!any_needs_fixup(block))
+        return root;
+
+    debugBelch("Compact imported at the wrong address, will fix up"
+               " internal pointers\n");
+
+    // I am PROOT!
+    proot = &root;
+
+    ok = fixup_loop(block, proot);
+    if (!ok)
+        *proot = NULL;
+
+    return *proot;
+}
+
+StgPtr
+compactFixupPointers(StgCompactNFData *str,
+                     StgClosure       *root)
+{
+    StgCompactNFDataBlock *block;
+    bdescr *bd;
+    StgWord total_blocks;
+
+    block = compactGetFirstBlock(str);
+
+    fixup_early(str, block);
+
+    root = maybe_fixup_internal_pointers(block, root);
+
+    // Do the late fixup even if we did not fixup all
+    // internal pointers, we need that for GC and Sanity
+    fixup_late(str, block);
+
+    // Now we're ready to let the GC, Sanity, the profiler
+    // etc. know about this object
+    bd = Bdescr((P_)block);
+
+    total_blocks = str->totalW / BLOCK_SIZE_W;
+
+    ACQUIRE_SM_LOCK;
+    ASSERT (bd->gen == g0);
+    ASSERT (g0->n_compact_blocks_in_import >= total_blocks);
+    g0->n_compact_blocks_in_import -= total_blocks;
+    g0->n_compact_blocks += total_blocks;
+    dbl_link_remove(bd, &g0->compact_blocks_in_import);
+    dbl_link_onto(bd, &g0->compact_objects);
+    RELEASE_SM_LOCK;
+
+#if DEBUG
+    if (root)
+        verify_consistency_loop(str);
+#endif
+
+    return (StgPtr)root;
+}
diff --git a/rts/sm/CNF.h b/rts/sm/CNF.h
new file mode 100644
index 0000000000..b34d9c96c1
--- /dev/null
+++ b/rts/sm/CNF.h
@@ -0,0 +1,71 @@
+/* -----------------------------------------------------------------------------
+ *
+ * (c) The GHC Team 1998-2014
+ *
+ * GC support for immutable non-GCed structures
+ *
+ * Documentation on the architecture of the Garbage Collector can be
+ * found in the online commentary:
+ *
+ *   http://ghc.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
+ *
+ * ---------------------------------------------------------------------------*/
+
+#ifndef SM_CNF_H
+#define SM_CNF_H
+
+#include "BeginPrivate.h"
+
+void              initCompact  (void);
+void              exitCompact  (void);
+
+StgCompactNFData *compactNew   (Capability      *cap,
+                                StgWord          size);
+StgPtr            compactAppend(Capability       *cap,
+                                StgCompactNFData *str,
+                                StgClosure       *what,
+                                StgWord           share);
+void              compactResize(Capability       *cap,
+                                StgCompactNFData *str,
+                                StgWord           new_size);
+void              compactFree  (StgCompactNFData *str);
+void              compactMarkKnown(StgCompactNFData *str);
+StgWord           compactContains(StgCompactNFData *str,
+                                  StgPtr            what);
+StgWord           countCompactBlocks(bdescr *outer);
+
+StgCompactNFDataBlock *compactAllocateBlock(Capability            *cap,
+                                            StgWord                size,
+                                            StgCompactNFDataBlock *previous);
+StgPtr                 compactFixupPointers(StgCompactNFData      *str,
+                                            StgClosure            *root);
+
+INLINE_HEADER StgCompactNFDataBlock *objectGetCompactBlock (StgClosure *closure);
+INLINE_HEADER StgCompactNFDataBlock *objectGetCompactBlock (StgClosure *closure)
+{
+    bdescr *object_block, *head_block;
+
+    object_block = Bdescr((StgPtr)closure);
+
+    ASSERT ((object_block->flags & BF_COMPACT) != 0);
+
+    if (object_block->blocks == 0)
+        head_block = object_block->link;
+    else
+        head_block = object_block;
+
+    ASSERT ((head_block->flags & BF_COMPACT) != 0);
+
+    return (StgCompactNFDataBlock*)(head_block->start);
+}
+
+INLINE_HEADER StgCompactNFData *objectGetCompact (StgClosure *closure);
+INLINE_HEADER StgCompactNFData *objectGetCompact (StgClosure *closure)
+{
+    StgCompactNFDataBlock *block = objectGetCompactBlock (closure);
+    return block->owner;
+}
+
+#include "EndPrivate.h"
+
+#endif // SM_COMPACT_H
diff --git a/rts/sm/Compact.c b/rts/sm/Compact.c
index ec178e91ef..3528fabb7b 100644
--- a/rts/sm/Compact.c
+++ b/rts/sm/Compact.c
@@ -470,6 +470,7 @@ update_fwd_large( bdescr *bd )
     switch (info->type) {
 
     case ARR_WORDS:
+    case COMPACT_NFDATA:
       // nothing to follow
       continue;
 
diff --git a/rts/sm/Evac.c b/rts/sm/Evac.c
index e53461de63..1f9c5cc8cd 100644
--- a/rts/sm/Evac.c
+++ b/rts/sm/Evac.c
@@ -25,6 +25,7 @@
 #include "Prelude.h"
 #include "Trace.h"
 #include "LdvProfile.h"
+#include "CNF.h"
 
 #if defined(PROF_SPIN) && defined(THREADED_RTS) && defined(PARALLEL_GC)
 StgWord64 whitehole_spin = 0;
@@ -245,7 +246,7 @@ copy(StgClosure **p, const StgInfoTable *info,
 
    This just consists of removing the object from the (doubly-linked)
    gen->large_objects list, and linking it on to the (singly-linked)
-   gen->new_large_objects list, from where it will be scavenged later.
+   gct->todo_large_objects list, from where it will be scavenged later.
 
    Convention: bd->flags has BF_EVACUATED set for a large object
    that has been evacuated, or unset otherwise.
@@ -305,12 +306,13 @@ evacuate_large(StgPtr p)
   bd->flags |= BF_EVACUATED;
   initBdescr(bd, new_gen, new_gen->to);
 
-  // If this is a block of pinned objects, we don't have to scan
-  // these objects, because they aren't allowed to contain any
+  // If this is a block of pinned or compact objects, we don't have to scan
+  // these objects, because they aren't allowed to contain any outgoing
   // pointers.  For these blocks, we skip the scavenge stage and put
   // them straight on the scavenged_large_objects list.
   if (bd->flags & BF_PINNED) {
       ASSERT(get_itbl((StgClosure *)p)->type == ARR_WORDS);
+
       if (new_gen != gen) { ACQUIRE_SPIN_LOCK(&new_gen->sync); }
       dbl_link_onto(bd, &new_gen->scavenged_large_objects);
       new_gen->n_scavenged_large_blocks += bd->blocks;
@@ -356,6 +358,110 @@ evacuate_static_object (StgClosure **link_field, StgClosure *q)
 }
 
 /* ----------------------------------------------------------------------------
+   Evacuate an object inside a CompactNFData
+
+   Don't actually evacuate the object. Instead, evacuate the structure
+   (which is a large object, so it is just relinked onto the new list
+   of large objects of the generation).
+
+   It is assumed that objects in the struct live in the same generation
+   as the struct itself all the time.
+   ------------------------------------------------------------------------- */
+STATIC_INLINE void
+evacuate_compact (StgPtr p)
+{
+    StgCompactNFData *str;
+    bdescr *bd;
+    generation *gen, *new_gen;
+    uint32_t gen_no, new_gen_no;
+
+    str = objectGetCompact((StgClosure*)p);
+    ASSERT(get_itbl((StgClosure*)str)->type == COMPACT_NFDATA);
+
+    bd = Bdescr((StgPtr)str);
+    gen_no = bd->gen_no;
+
+    // already evacuated? (we're about to do the same check,
+    // but we avoid taking the spin-lock)
+    if (bd->flags & BF_EVACUATED) {
+        /* Don't forget to set the gct->failed_to_evac flag if we didn't get
+         * the desired destination (see comments in evacuate()).
+         */
+        if (gen_no < gct->evac_gen_no) {
+            gct->failed_to_evac = rtsTrue;
+            TICK_GC_FAILED_PROMOTION();
+        }
+        return;
+    }
+
+    gen = bd->gen;
+    gen_no = bd->gen_no;
+    ACQUIRE_SPIN_LOCK(&gen->sync);
+
+    // already evacuated?
+    if (bd->flags & BF_EVACUATED) {
+        /* Don't forget to set the gct->failed_to_evac flag if we didn't get
+         * the desired destination (see comments in evacuate()).
+         */
+        if (gen_no < gct->evac_gen_no) {
+            gct->failed_to_evac = rtsTrue;
+            TICK_GC_FAILED_PROMOTION();
+        }
+        RELEASE_SPIN_LOCK(&gen->sync);
+        return;
+    }
+
+    // remove from large_object list
+    if (bd->u.back) {
+        bd->u.back->link = bd->link;
+    } else { // first object in the list
+        gen->compact_objects = bd->link;
+    }
+    if (bd->link) {
+        bd->link->u.back = bd->u.back;
+    }
+
+    /* link it on to the evacuated compact object list of the destination gen
+     */
+    new_gen_no = bd->dest_no;
+
+    if (new_gen_no < gct->evac_gen_no) {
+        if (gct->eager_promotion) {
+            new_gen_no = gct->evac_gen_no;
+        } else {
+            gct->failed_to_evac = rtsTrue;
+        }
+    }
+
+    new_gen = &generations[new_gen_no];
+
+    // Note: for speed we only update the generation of the first block here
+    // This means that bdescr of subsequent blocks will think they are in
+    // the wrong generation
+    // (This should not be a problem because there is no code that checks
+    // for that - the only code touching the generation of the block is
+    // in the GC, and that should never see blocks other than the first)
+    bd->flags |= BF_EVACUATED;
+    initBdescr(bd, new_gen, new_gen->to);
+
+    if (new_gen != gen) { ACQUIRE_SPIN_LOCK(&new_gen->sync); }
+    dbl_link_onto(bd, &new_gen->live_compact_objects);
+    new_gen->n_live_compact_blocks += str->totalW / BLOCK_SIZE_W;
+    if (new_gen != gen) { RELEASE_SPIN_LOCK(&new_gen->sync); }
+
+    RELEASE_SPIN_LOCK(&gen->sync);
+
+    // Note: the object did not move in memory, because it lives
+    // in pinned (BF_COMPACT) allocation, so we do not need to rewrite it
+    // or muck with forwarding pointers
+    // Also there is no tag to worry about on the struct (tags are used
+    // for constructors and functions, but a struct is neither). There
+    // might be a tag on the object pointer, but again we don't change
+    // the pointer because we don't move the object so we don't need to
+    // rewrite the tag.
+}
+
+/* ----------------------------------------------------------------------------
    Evacuate
 
    This is called (eventually) for every live object in the system.
@@ -459,8 +565,7 @@ loop:
 
   bd = Bdescr((P_)q);
 
-  if ((bd->flags & (BF_LARGE | BF_MARKED | BF_EVACUATED)) != 0) {
-
+  if ((bd->flags & (BF_LARGE | BF_MARKED | BF_EVACUATED | BF_COMPACT)) != 0) {
       // pointer into to-space: just return it.  It might be a pointer
       // into a generation that we aren't collecting (> N), or it
       // might just be a pointer into to-space.  The latter doesn't
@@ -478,6 +583,15 @@ loop:
           return;
       }
 
+      // Check for compact before checking for large, this allows doing the
+      // right thing for objects that are half way in the middle of the first
+      // block of a compact (and would be treated as large objects even though
+      // they are not)
+      if (bd->flags & BF_COMPACT) {
+          evacuate_compact((P_)q);
+          return;
+      }
+
       /* evacuate large objects by re-linking them onto a different list.
        */
       if (bd->flags & BF_LARGE) {
@@ -735,6 +849,12 @@ loop:
       copy(p,info,q,sizeofW(StgTRecChunk),gen_no);
       return;
 
+  case COMPACT_NFDATA:
+      // CompactNFData objects are at least one block plus the header
+      // so they are larger than the large_object_threshold (80% of
+      // block size) and never copied by value
+      barf("evacuate: compact nfdata is not large");
+      return;
   default:
     barf("evacuate: strange closure type %d", (int)(INFO_PTR_TO_STRUCT(info)->type));
   }
diff --git a/rts/sm/GC.c b/rts/sm/GC.c
index 54798719a4..7796f30965 100644
--- a/rts/sm/GC.c
+++ b/rts/sm/GC.c
@@ -47,6 +47,7 @@
 #include "RaiseAsync.h"
 #include "Stable.h"
 #include "CheckUnload.h"
+#include "CNF.h"
 
 #include <string.h> // for memset()
 #include <unistd.h>
@@ -592,6 +593,23 @@ GarbageCollect (uint32_t collect_gen,
         gen->n_large_blocks = gen->n_scavenged_large_blocks;
         gen->n_large_words  = countOccupied(gen->large_objects);
         gen->n_new_large_words = 0;
+
+        /* COMPACT_NFDATA. The currently live compacts are chained
+         * to live_compact_objects, quite like large objects. And
+         * objects left on the compact_objects list are dead.
+         *
+         * We don't run a simple freeChain because want to give the
+         * CNF module some chance to free memory that freeChain would
+         * not see (namely blocks appended to a CNF through a compactResize).
+         *
+         * See Note [Compact Normal Forms] for details.
+         */
+        for (bd = gen->compact_objects; bd; bd = next) {
+            next = bd->link;
+            compactFree(((StgCompactNFDataBlock*)bd->start)->owner);
+        }
+        gen->compact_objects = gen->live_compact_objects;
+        gen->n_compact_blocks = gen->n_live_compact_blocks;
     }
     else // for generations > N
     {
@@ -605,15 +623,27 @@ GarbageCollect (uint32_t collect_gen,
             gen->n_large_words += bd->free - bd->start;
         }
 
+        // And same for compacts
+        for (bd = gen->live_compact_objects; bd; bd = next) {
+            next = bd->link;
+            dbl_link_onto(bd, &gen->compact_objects);
+        }
+
         // add the new blocks we promoted during this GC
         gen->n_large_blocks += gen->n_scavenged_large_blocks;
+        gen->n_compact_blocks += gen->n_live_compact_blocks;
     }
 
     ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
     ASSERT(countOccupied(gen->large_objects) == gen->n_large_words);
+    // We can run the same assertion on compact objects because there
+    // is memory "the GC doesn't see" (directly), but which is still
+    // accounted in gen->n_compact_blocks
 
     gen->scavenged_large_objects = NULL;
     gen->n_scavenged_large_blocks = 0;
+    gen->live_compact_objects = NULL;
+    gen->n_live_compact_blocks = 0;
 
     // Count "live" data
     live_words  += genLiveWords(gen);
@@ -1207,6 +1237,8 @@ prepare_collected_gen (generation *gen)
     // initialise the large object queues.
     ASSERT(gen->scavenged_large_objects == NULL);
     ASSERT(gen->n_scavenged_large_blocks == 0);
+    ASSERT(gen->live_compact_objects == NULL);
+    ASSERT(gen->n_live_compact_blocks == 0);
 
     // grab all the partial blocks stashed in the gc_thread workspaces and
     // move them to the old_blocks list of this gen.
@@ -1246,6 +1278,11 @@ prepare_collected_gen (generation *gen)
         bd->flags &= ~BF_EVACUATED;
     }
 
+    // mark the compact objects as from-space
+    for (bd = gen->compact_objects; bd; bd = bd->link) {
+        bd->flags &= ~BF_EVACUATED;
+    }
+
     // for a compacted generation, we need to allocate the bitmap
     if (gen->mark) {
         StgWord bitmap_size; // in bytes
@@ -1472,7 +1509,8 @@ resize_generations (void)
             words = oldest_gen->n_words;
         }
         live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
-            oldest_gen->n_large_blocks;
+            oldest_gen->n_large_blocks +
+            oldest_gen->n_compact_blocks;
 
         // default max size for all generations except zero
         size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
diff --git a/rts/sm/Sanity.c b/rts/sm/Sanity.c
index 62d53e046d..6f6b15c4e8 100644
--- a/rts/sm/Sanity.c
+++ b/rts/sm/Sanity.c
@@ -28,6 +28,7 @@
 #include "Printer.h"
 #include "Arena.h"
 #include "RetainerProfile.h"
+#include "CNF.h"
 
 /* -----------------------------------------------------------------------------
    Forward decls.
@@ -424,7 +425,7 @@ checkClosure( const StgClosure* p )
       }
 
     default:
-            barf("checkClosure (closure type %d)", info->type);
+        barf("checkClosure (closure type %d)", info->type);
     }
 }
 
@@ -485,6 +486,37 @@ checkLargeObjects(bdescr *bd)
 }
 
 static void
+checkCompactObjects(bdescr *bd)
+{
+    // Compact objects are similar to large objects,
+    // but they have a StgCompactNFDataBlock at the beginning,
+    // before the actual closure
+
+    for ( ; bd != NULL; bd = bd->link) {
+        StgCompactNFDataBlock *block, *last;
+        StgCompactNFData *str;
+        StgWord totalW;
+
+        ASSERT (bd->flags & BF_COMPACT);
+
+        block = (StgCompactNFDataBlock*)bd->start;
+        str = block->owner;
+        ASSERT ((W_)str == (W_)block + sizeof(StgCompactNFDataBlock));
+
+        totalW = 0;
+        for ( ; block ; block = block->next) {
+            last = block;
+            ASSERT (block->owner == str);
+
+            totalW += Bdescr((P_)block)->blocks * BLOCK_SIZE_W;
+        }
+
+        ASSERT (str->totalW == totalW);
+        ASSERT (str->last == last);
+    }
+}
+
+static void
 checkSTACK (StgStack *stack)
 {
     StgPtr sp = stack->sp;
@@ -715,6 +747,7 @@ static void checkGeneration (generation *gen,
     }
 
     checkLargeObjects(gen->large_objects);
+    checkCompactObjects(gen->compact_objects);
 }
 
 /* Full heap sanity check. */
@@ -744,6 +777,14 @@ void checkSanity (rtsBool after_gc, rtsBool major_gc)
     }
 }
 
+static void
+markCompactBlocks(bdescr *bd)
+{
+    for (; bd != NULL; bd = bd->link) {
+        compactMarkKnown(((StgCompactNFDataBlock*)bd->start)->owner);
+    }
+}
+
 // If memInventory() calculates that we have a memory leak, this
 // function will try to find the block(s) that are leaking by marking
 // all the ones that we know about, and search through memory to find
@@ -764,6 +805,7 @@ findMemoryLeak (void)
         }
         markBlocks(generations[g].blocks);
         markBlocks(generations[g].large_objects);
+        markCompactBlocks(generations[g].compact_objects);
     }
 
     for (i = 0; i < n_nurseries; i++) {
@@ -833,8 +875,11 @@ genBlocks (generation *gen)
 {
     ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
     ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
+    ASSERT(countCompactBlocks(gen->compact_objects) == gen->n_compact_blocks);
+    ASSERT(countCompactBlocks(gen->compact_blocks_in_import) == gen->n_compact_blocks_in_import);
     return gen->n_blocks + gen->n_old_blocks +
-            countAllocdBlocks(gen->large_objects);
+        countAllocdBlocks(gen->large_objects) +
+        gen->n_compact_blocks + gen->n_compact_blocks_in_import;
 }
 
 void
diff --git a/rts/sm/Scav.c b/rts/sm/Scav.c
index 18a30d3bdf..1549df5021 100644
--- a/rts/sm/Scav.c
+++ b/rts/sm/Scav.c
@@ -795,6 +795,13 @@ scavenge_block (bdescr *bd)
         break;
       }
 
+    case COMPACT_NFDATA:
+        // CompactNFData blocks live in compact lists, which we don't
+        // scavenge, because there nothing to scavenge in them
+        // so we should never ever see them
+        barf("scavenge: found unexpected Compact structure");
+        break;
+
     default:
         barf("scavenge: unimplemented/strange closure type %d @ %p",
              info->type, p);
@@ -1953,7 +1960,7 @@ scavenge_large (gen_workspace *ws)
 
         // take this object *off* the large objects list and put it on
         // the scavenged large objects list.  This is so that we can
-        // treat new_large_objects as a stack and push new objects on
+        // treat todo_large_objects as a stack and push new objects on
         // the front when evacuating.
         ws->todo_large_objects = bd->link;
 
diff --git a/rts/sm/Storage.c b/rts/sm/Storage.c
index 7c41f8c64b..3f8889658f 100644
--- a/rts/sm/Storage.c
+++ b/rts/sm/Storage.c
@@ -100,8 +100,12 @@ initGeneration (generation *gen, int g)
     gen->n_large_blocks = 0;
     gen->n_large_words = 0;
     gen->n_new_large_words = 0;
+    gen->compact_objects = NULL;
+    gen->n_compact_blocks = 0;
     gen->scavenged_large_objects = NULL;
     gen->n_scavenged_large_blocks = 0;
+    gen->live_compact_objects = NULL;
+    gen->n_live_compact_blocks = 0;
     gen->mark = 0;
     gen->compact = 0;
     gen->bitmap = NULL;
@@ -1208,12 +1212,13 @@ W_ countOccupied (bdescr *bd)
 
 W_ genLiveWords (generation *gen)
 {
-    return gen->n_words + gen->n_large_words;
+    return gen->n_words + gen->n_large_words +
+        gen->n_compact_blocks * BLOCK_SIZE_W;
 }
 
 W_ genLiveBlocks (generation *gen)
 {
-    return gen->n_blocks + gen->n_large_blocks;
+    return gen->n_blocks + gen->n_large_blocks + gen->n_compact_blocks;
 }
 
 W_ gcThreadLiveWords (uint32_t i, uint32_t g)
@@ -1266,7 +1271,8 @@ calcNeeded (rtsBool force_major, memcount *blocks_needed)
         gen = &generations[g];
 
         blocks = gen->n_blocks // or: gen->n_words / BLOCK_SIZE_W (?)
-               + gen->n_large_blocks;
+               + gen->n_large_blocks
+               + gen->n_compact_blocks;
 
         // we need at least this much space
         needed += blocks;