21. AMC pool class

21.1. Introduction

.intro: This document contains a guide (.guide) to the MPS AMC pool class, followed by the historical initial design (.initial-design).

.readership: Any MPS developer.

21.2. Guide

.guide: The AMC pool class is a general-purpose automatic (collecting) pool class. It is intended for most client objects. AMC is “Automatic, Mostly Copying”: it preserves objects by copying, except when an ambiguous reference ‘nails’ the object in place. It is generational. Chain: specify capacity and mortality of generations 0 to N − 1. Survivors from generation N − 1 get promoted into an arena-wide “top” generation (often anachronistically called the “dynamic” generation, which was the term on the Lisp Machine).

21.2.1. Segment states

.seg.state: AMC segments are in one of three states: “mobile”, “boarded”, or “stuck”.

.seg.state.mobile: Segments are normally mobile: all objects on the seg are un-nailed, and thus may be preserved by copying.

.seg.state.boarded: An ambiguous reference to any address within an segment makes that segment boarded: a nailboard is allocated to record ambiguous references (“nails”), but un-nailed objects on the segment are still preserved by copying.

.seg.state.stuck: Stuck segments only occur in emergency tracing: a discovery fix to an object in a mobile segment is recorded in the only non-allocating way available: by making the entire segment stuck.

21.2.2. Pads

(See job001809 and job001811, and mps/branch/2009-03-31/padding.)

.pad: A pad is logically a trivial client object. Pads are created by the MPS asking the client’s format code to create them, to fill up a space in a segment. Thereafter, the pad appears to the MPS as a normal client object (that is: the MPS cannot distinguish a pad from a client object).

.pad.reason: AMC creates pads for three reasons: buffer empty fragment (BEF), large segment padding (LSP), and non-mobile reclaim (NMR). (Large segment pads were new with job001811.)

.pad.reason.bef: Buffer empty fragment (BEF) pads are made by AMCBufferEmpty() whenever it detaches a non-empty buffer from an AMC segment. Buffer detachment is most often caused because the buffer is too small for the current buffer reserve request (which may be either a client requested or a forwarding allocation). Detachment may happen for other reasons, such as trace flip.

.pad.reason.lsp: Large segment padding (LSP) pads are made by AMCBufferFill() when the requested fill size is “large” (see The LSP payoff calculation below). AMCBufferFill() fills the buffer to exactly the size requested by the current buffer reserve operation; that is: it does not round up to the whole segment size. This prevents subsequent small objects being placed in the same segment as a single very large object. If the buffer fill size is less than the segment size, AMCBufferFill() fills any remainder with an large segment pad.

.pad.reason.nmr: Non-mobile reclaim (NMR) pads are made by amcReclaimNailed(), when performing reclaim on a non-mobile (that is, either boarded or stuck) segment:

The more common NMR scenario is reclaim of a boarded segment after a non-emergency trace. Ambiguous references into the segment are recorded as nails. Subsequent exact references to a nailed object do nothing further, but exact refs that do not match a nail cause preserve-by-copy and leave a forwarding object. Unreachable objects are not touched during the scan+fix part of the trace. On reclaim, only nailed objects need to be preserved; others (namely forwarding pointers and unreachable objects) are replaced by an NMR pad. (Note that a BEF or LSP pad appears to be an unreachable object, and is therefore overwritten by an NMR pad).

The less common NMR scenario is after emergency tracing. Boarded segments still occur; they may have nailed objects from ambiguous references, forwarding objects from pre-emergency exact fixes, nailed objects from mid-emergency exact fixes, and unpreserved objects; reclaim is as in the non-emergency case. Stuck segments may have forwarding objects from pre-emergency exact fixes, objects from mid-emergency fixes, and unreachable objects – but the latter two are not distinguishable because there is no nailboard. On reclaim, all objects except forwarding pointers are preserved; each forwarding object is replaced by an NMR pad.

If amcReclaimNailed() finds no objects to be preserved then it calls SegFree() (new with job001809).

21.2.3. Placement pads are okay

Placement pads are the BEF and LSP pads created in “to-space” when placing objects into segments. This wasted space is an expected space-cost of AMC’s naive (but time-efficient) approach to placement of objects into segments. This is normally not a severe problem. (The worst case is a client that always requests ArenaAlign() + 1 byte objects: this has a nearly 100% overhead).

21.2.4. Retained pads could be a problem

Retained pads are the NMR pads stuck in “from-space”: non-mobile segments that were condemned but have preserved-in-place objects cannot be freed by amcReclaimNailed(). The space around the preserved objects is filled with NMR pads.

In the worst case, retained pads could waste an enormous amount of space! A small (one-byte) object could retain a multi-page segment for as long as the ambiguous reference persists; that is: indefinitely. Imagine a 256-page (1 MiB) segment containing a very large object followed by a handful of small objects. An ambiguous reference to one of the small objects will unfortunately cause the entire 256-page segment to be retained, mostly as an NMR pad; this is a massive overhead of wasted space.

AMC mitigates this worst-case behaviour, by treating large segments specially.

21.2.5. Small, medium, and large segments

AMC categorises segments as small (one page), medium (several pages), or large (AMCLargeSegPAGES or more):

pages = SegSize(seg) / ArenaAlign(arena);
if(pages == 1) {
  /* small */
} else if(pages < AMCLargeSegPAGES) {
  /* medium */
} else {
  /* large */
}</code></pre></blockquote>

AMCLargeSegPAGES is currently 8 – see The LSP payoff calculation below.

AMC might treat “Large” segments specially, in two ways:

• .large.single-reserve: A large segment is only used for a single (large) buffer reserve request; the remainder of the segment (if any) is immediately padded with an LSP pad.

• .large.lsp-no-retain: Nails to such an LSP pad do not cause AMCReclaimNailed() to retain the segment.

.large.single-reserve is implemented. See job001811.

.large.lsp-no-retain is not currently implemented.

The point of .large.lsp-no-retain would be to avoid retention of the (large) segment when there is a spurious ambiguous reference to the LSP pad at the end of the segment. Such an ambiguous reference might happen naturally and repeatably if the preceding large object is an array, the array is accessed by an ambiguous element pointer (for example, on the stack), and the element pointer ends up pointing just off the end of the large object (as is normal for sequential element access in C) and remains with that value for a while. (Such an ambiguous reference could also occur by chance, for example, by coincidence with an int or float, or when the stack grows to include old unerased values).

Implementing .large.lsp-no-retain is a little tricky. A pad is indistinguishable from a client object, so AMC has no direct way to detect, and safely ignore, the final LSP object in the seg. If AMC could guarantee that the single buffer reserve (.large.single-reserve) is only used for a single object, then AMCReclaimNailed() could honour a nail at the start of a large seg and ignore all others; this would be extremely simple to implement. But AMC cannot guarantee this, because in the MPS Allocation Point Protocol the client is permitted to make a large buffer reserve and then fill it with many small objects. In such a case, AMC must honour all nails (if the buffer reserve request was an exact multiple of ArenaAlign()), or all nails except to the last object (if there was a remainder filled with an LSP pad). Because an LSP pad cannot be distinguished from a client object, and the requested allocation size is not recorded, AMC cannot distinguish these two conditions at reclaim time. Therefore AMC must record whether or not the last object in the seg is a pad, in order to ignore nails to it. This could be done by adding a flag to AMCSegStruct. (This can be done without increasing the structure size, by making the Bool new field smaller than its current 32 bits.)

21.2.6. The LSP payoff calculation

The LSP fix for job001811 treats large segments differently. Without it, after allocating a very large object (in a new very large multi-page segment), MPS would happily place subsequent small objects in any remaining space at the end of the segment. This would risk pathological fragmentation: if these small objects were systematically preserved by ambiguous refs, enormous NMR pads would be retained along with them.

The payoff calculation is a bit like deciding whether or not to purchase insurance. For single-page and medium-sized segments, we go ahead and use the remaining space for subsequent small objects. This is equivalent to choosing not to purchase insurance. If the small objects were to be preserved by ambiguous refs, the retained NMR pads would be big, but not massive. We expect such ambiguous refs to be uncommon, so we choose to live with this slight risk of bad fragmentation. The benefit is that the remaining space is used.

For large segments, we decide that the risk of using the remainder is just too great, and the benefit too small, so we throw it away as an LSP pad. This is equivalent to purchasing insurance: we choose to pay a known small cost every time, to avoid risking an occasional disaster.

To decide what size of segment counts as “large”, we must decide how much uninsured risk we can tolerate, versus how much insurance cost we can tolerate. The likelihood of ambiguous references retaining objects is entirely dependent on client behaviour. However, as a sufficient “one size fits all” policy, I (RHSK 2009-09-14) have judged that segments smaller than eight pages long do not need to be treated as large: the insurance cost to “play safe” would be considerable (wasting up to one page of remainder per seven pages of allocation), and the fragmentation overhead risk is not that great (at most eight times worse than the unavoidable minimum). So AMCLargeSegPAGES is defined as 8 in config.h. As long as the assumption that most segments are not ambiguously referenced remains correct, I expect this policy will be satisfactory.

To verify that this threshold is acceptable for a given client, poolamc.c calculates metrics; see Feedback about retained pages below. If this one-size-fits-all approach is not satisfactory, AMCLargeSegPAGES could be made a client-tunable parameter.

21.2.7. Retained pages

The reasons why a segment and its pages might be retained are:

1. ambiguous reference to first-obj: unavoidable page retention (only the mutator can reduce this, if they so wish, by nulling out ambig referencess);

2. ambiguous reference to rest-obj: tuning MPS LSP policy could mitigate this, reducing the likelihood of rest-objs being co-located with large first-objs;

3. ambiguous reference to final pad: implementing .large.lsp-no-retain could mitigate this;

4. ambiguous reference to other (NMR) pad: hard to mitigate, as pads are indistinguishable from client objects;

5. emergency trace;

6. non-object-aligned ambiguous ref: fixed by job001809;

7. other reason (for example, buffered at flip): not expected to be a problem.

This list puts the reasons that are more “obvious” to the client programmer first, and the more obscure reasons last.

21.2.8. Feedback about retained pages

(New with job001811). AMC now accumulates counts of pages condemned and retained during a trace, in categories according to size and reason for retention, and emits this via the AMCTraceEnd telemetry event. See comments on the PageRetStruct in poolamc.c. These page-based metrics are not as precise as actually counting the size of objects, but they require much less intrusive code to implement, and should be sufficient to assess whether AMC’s page retention policies and behaviour are acceptable.

21.3. Initial design

.initial-design: This section contains the original design for the AMC Pool Class.

21.3.1. Introduction

.intro: This is the design of the AMC Pool Class. AMC stands for Automatic Mostly-Copying. This design is highly fragmentory and some may even be sufficiently old to be misleading.

.readership: The intended readership is any MPS developer.

21.3.2. Overview

.overview: This class is intended to be the main pool class used by Harlequin Dylan. It provides garbage collection of objects (hence “automatic”). It uses generational copying algorithms, but with some facility for handling small numbers of ambiguous references. Ambiguous references prevent the pool from copying objects (hence “mostly copying”). It provides incremental collection.

Note

A lot of this design is awesomely old. David Jones, 1998-02-04.

21.3.3. Definitions

.def.grain: Grain. An quantity of memory which is both aligned to the pool’s alignment and equal to the pool’s alignment in size. That is, the smallest amount of memory worth talking about.

21.3.4. Segments

.seg.class: AMC allocates segments of class AMCSegClass, which is a subclass of GCSegClass. Instances contain a segTypeP field, which is of type int*.

.seg.gen: AMC organizes the segments it manages into generations.

.seg.gen.map: Every segment is in exactly one generation.

.seg.gen.ind: The segment’s segTypeP field indicates which generation (that the segment is in) (an AMCGenStruct see blah below).

.seg.typep: The segTypeP field actually points to either the type field of a generation or to the type field of a nail board.

.seg.typep.distinguish: The type field (which can be accessed in either case) determines whether the segTypeP field is pointing to a generation or to a nail board.

.seg.gen.get: The map from segment to generation is implemented by AMCSegGen() which deals with all this.

21.3.5. Fixing and nailing

Note

This section contains placeholders for design rather than design really. David Jones, 1998-02-04.

.nailboard: AMC uses a nail board structure for recording ambiguous references to segments. A nail board is a bit table with one bit per grain in the segment.

.nailboard.create: Nail boards are allocated dynamically whenever a segment becomes newly ambiguously referenced.

.nailboard.destroy: They are deallocated during reclaim. Ambiguous fixes simply set the appropriate bit in this table. This table is used by subsequent scans and reclaims in order to work out what objects were marked.

.nailboard.emergency: During emergency tracing two things relating to nail boards happen that don’t normally:

1. .nailboard.emergency.nonew: Nail boards aren’t allocated when we have new ambiguous references to segments.

   .nailboard.emergency.nonew.justify: We could try and allocate a nail board, but we’re in emergency mode so short of memory so it’s unlikely to succeed, and there would be additional code for yet another error path which complicates things.

2. .nailboard.emergency.exact: nail boards are used to record exact references in order to avoid copying the objects.

   .nailboard.hyper-conservative: Not creating new nail boards (.nailboard.emergency.nonew above) means that when we have a new reference to a segment during emergency tracing then we nail the entire segment and preserve everything in place.

.fix.nail.states: Partition the segment states into four sets:

1. white segment and not nailed (and has no nail board);

2. white segment and nailed and has no nail board;

3. white segment and nailed and has nail board;

4. the rest.

.fix.nail.why: A segment is recorded as being nailed when either there is an ambiguous reference to it, or there is an exact reference to it and the object couldn’t be copied off the segment (because there wasn’t enough memory to allocate the copy). In either of these cases reclaim cannot simply destroy the segment (usually the segment will not be destroyed because it will have live objects on it, though see .nailboard.limitations.middle below). If the segment is nailed then we might be using a nail board to mark objects on the segment. However, we cannot guarantee that being nailed implies a nail board, because we might not be able to allocate the nail board. Hence all these states actually occur in practice.

.fix.nail.distinguish: The nailed bits in the segment descriptor (SegStruct) are used to record whether a segment is nailed or not. The segTypeP field of the segment either points to (the “type” field of) an AMCGen or to an AMCNailBoard, the type field can be used to determine which of these is the case. (see .seg.typep above).

.nailboard.limitations.single: Just having a single nail board per segment prevents traces from improving on the findings of each other: a later trace could find that a nailed object is no longer nailed or even dead. Until the nail board is discarded, that is.

.nailboard.limitations.middle: An ambiguous reference into the middle of an object will cause the segment to survive, even if there are no surviving objects on it.

.nailboard.limitations.reclaim: AMCReclaimNailed() could cover each block of reclaimed objects between two nailed objects with a single padding object, speeding up further scans.

21.3.6. Emergency tracing

.emergency.fix: AMCFixEmergency() is at the core of AMC’s emergency tracing policy (unsurprisingly). AMCFixEmergency() chooses exactly one of three options:

1. use the existing nail board structure to record the fix;

2. preserve and nail the segment in its entirety;

3. snapout an exact (or high rank) pointer to a broken heart to the broken heart’s forwarding pointer.

If the rank of the reference is RankAMBIG then it either does (1) or (2) depending on wether there is an existing nail board or not. Otherwise (the rank is exact or higher) if there is a broken heart it is used to snapout the pointer. Otherwise it is as for an RankAMBIG reference: we either do (1) or (2).

.emergency.scan: This is basically as before, the only complication is that when scanning a nailed segment we may need to do multiple passes, as FixEmergency() may introduce new marks into the nail board.

21.3.7. Buffers

.buffer.class: AMC uses buffer of class AMCBufClass (a subclass of SegBufClass).

.buffer.gen: Each buffer allocates into exactly one generation.

.buffer.field.gen: AMCBuf buffer contain a gen field which points to the generation that the buffer allocates into.

.buffer.fill.gen: AMCBufferFill() uses the generation (obtained from the gen field) to initialise the segment’s segTypeP field which is how segments get allocated in that generation.

.buffer.condemn: We condemn buffered segments, but not the contents of the buffers themselves, because we can’t reclaim uncommitted buffers (see design.mps.buffer for details). If the segment has a forwarding buffer on it, we detach it.

Note

Why? Forwarding buffers are detached because they used to cause objects on the same segment to not get condemned, hence caused retention of garbage. Now that we condemn the non-buffered portion of buffered segments this is probably unnecessary. David Jones, 1998-06-01.

But it’s probably more efficient than keeping the buffer on the segment, because then the other stuff gets nailed – Pekka P. Pirinen, 1998-07-10.

If the segment has a mutator buffer on it, we nail the buffer. If the buffer cannot be nailed, we give up condemning, since nailing the whole segment would make it survive anyway. The scan methods skip over buffers and fix methods don’t do anything to things that have already been nailed, so the buffer is effectively black.

21.3.8. Types

.struct: AMCStruct is the pool class AMC instance structure.

.struct.pool: Like other pool class instances, it contains a PoolStruct containing the generic pool fields.

.struct.format: The format field points to a Format structure describing the object format of objects allocated in the pool. The field is intialized by AMCInit() from a parameter, and thereafter it is not changed until the pool is destroyed.

Note

Actually the format field is in the generic PoolStruct these days. David Jones, 1998-09-21.

Note

There are lots more fields here.

21.3.9. Generations

.gen: Generations partition the segments that a pool manages (see .seg.gen.map above).

.gen.collect: Generations are more or less the units of condemnation in AMC. And also the granularity for forwarding (when copying objects during a collection): all the objects which are copied out of a generation use the same forwarding buffer for allocating the new copies, and a forwarding buffer results in allocation in exactly one generation.

.gen.rep: Generations are represented using an AMCGenStruct structure.

.gen.create: All the generation are create when the pool is created (during AMCInitComm()).

.gen.manage.ring: An AMC’s generations are kept on a ring attached to the AMCStruct (the genRing field).

.gen.manage.array: They are also kept in an array which is allocated when the pool is created and attached to the AMCStruct (the gens field holds the number of generations, the gen field points to an array of AMCGen).

Note

it seems to me that we could probably get rid of the ring. David Jones, 1998-09-22.

.gen.number: There are AMCTopGen + 2 generations in total. “normal” generations numbered from 0 to AMCTopGen inclusive and an extra “ramp” generation (see .gen.ramp below).

.gen.forward: Each generation has an associated forwarding buffer (stored in the forward field of AMCGen). This is the buffer that is used to forward objects out of this generation. When a generation is created in AMCGenCreate(), its forwarding buffer has a null p field, indicating that the forwarding buffer has no generation to allocate in. The collector will assert out (in AMCBufferFill() where it checks that buffer->p is an AMCGen) if you try to forward an object out of such a generation.

.gen.forward.setup: All the generation’s forwarding buffer’s are associated with generations when the pool is created (just after the generations are created in AMCInitComm()).

21.3.10. Ramps

.ramp: Ramps usefully implement the begin/end mps_alloc_pattern_ramp() interface.

.gen.ramp: To implement ramping (request.dylan.170423), AMC uses a special “ramping mode”, where promotions are redirected. One generation is designated the “ramp generation” (amc->rampGen in the code).

.gen.ramp.ordinary: Ordinarily, that is whilst not ramping, objects are promoted into the ramp generation from younger generations and are promoted out to older generations. The generation that the ramp generation ordinarily promotes into is designated the “after-ramp generation” (amc->afterRampGen).

.gen.ramp.particular: the ramp generation is the second oldest generation and the after-ramp generation is the oldest generation.

.gen.ramp.possible: In alternative designs it might be possible to make the ramp generation a special generation that is only promoted into during ramping, however, this is not done.

.gen.ramp.ramping: The ramp generation is promoted into itself during ramping mode;

.gen.ramp.after: after this mode ends, the ramp generation is promoted into the after-ramp generation as usual.

.gen.ramp.after.once: Care is taken to ensure that there is at least one collection where stuff is promoted from the ramp generation to the after-ramp generation even if ramping mode is immediately re-entered.

.ramp.mode: This behaviour is controlled in a slightly convoluted manner by a state machine. The rampMode field of the pool forms an important part of the state of the machine.

There are five states: OUTSIDE, BEGIN, RAMPING, FINISH, and COLLECTING. These appear in the code as RampOUTSIDE and so on.

.ramp.state.cycle.usual: The usual progression of states is a cycle: OUTSIDE → BEGIN → RAMPING → FINISH → COLLECTING → OUTSIDE.

.ramp.count: The pool just counts the number of APs that have begun ramp mode (and not ended). No state changes occur unless this count goes from 0 to 1 (starting the first ramp) or from 1 to 0 (leaving the last ramp). In other words, all nested ramps are ignored (see code in AMCRampBegin() and AMCRampEnd()).

.ramp.state.invariant.count: In the OUTSIDE state the count must be zero. In the BEGIN and RAMPING states the count must be greater than zero. In the FINISH and COLLECTING states the count is not constrained.

.ramp.state.invariant.forward: When in OUTSIDE, BEGIN, or COLLECTING, the ramp generation forwards to the after-ramp generation. When in RAMPING or FINISH, the ramp generation forwards to itself.

.ramp.outside: The pool is initially in the OUTSIDE state. The only transition away from the OUTSIDE state is to the BEGIN state, when a ramp is entered.

.ramp.begin: When the count goes up from zero, the state moves from COLLECTING or OUTSIDE to BEGIN.

.ramp.begin.leave: We can leave the BEGIN state to either the OUTSIDE or the RAMPING state.

.ramp.begin.leave.outside: We go to OUTSIDE if the count drops to 0 before a collection starts. This shortcuts the usual cycle of states for small enough ramps.

.ramp.begin.leave.ramping: We enter the RAMPING state if a collection starts that condemns the ramp generation (pedantically when a new GC begins, and a segment in the ramp generation is condemned, we leave the BEGIN state, see AMCWhiten). At this point we switch the ramp generation to forward to itself (.gen.ramp.ramping).

.ramp.ramping.leave: We leave the RAMPING state and go to the FINISH state when the ramp count goes back to zero. Thus, the FINISH state indicates that we have started collecting the ramp generation while inside a ramp which we have subsequently finished.

.ramp.finish.remain: We remain in the FINISH state until we next start to collect the ramp generation (condemn it), regardless of entering or leaving any ramps. This ensures that the ramp generation will be collected to the after-ramp generation at least once.

.ramp.finish.leave: When we next condemn the ramp genearation, we move to the COLLECTING state. At this point the forwarding generations are switched back so that the ramp generation promotes into the after-ramp generation on this collection.

.ramp.collecting.leave: We leave the COLLECTING state when the GC enters reclaim (specifically, when a segment in the ramp generation is reclaimed), or when we begin another ramp. Ordinarily we enter the OUTSIDE state, but if the client has started a ramp then we go directly to the BEGIN state.

.ramp.collect-all There used to be two flavours of ramps: the normal one and the collect-all flavour that triggered a full GC after the ramp end. This was a hack for producing certain Dylan statistics, and no longer has any effect (the flag is passed to AMCRampBegin(), but ignored there).

21.3.11. Headers

.header: AMC supports a fixed-size header on objects, with the client pointers pointing after the header, rather than the base of the memory block. See format documentation for details of the interface.

.header.client: The code mostly deals in client pointers, only computing the base and limit of a block when these are needed (such as when an object is copied). In several places, the code gets a block of some sort, a segment or a buffer, and creates a client pointer by adding the header length (pool->format->headerLength).

.header.fix: There are two versions of the fix method, due to its criticality, with (AMCHeaderFix()) and without (AMCFix()) headers. The correct one is selected in AMCInitComm(), and placed in the pool’s fix field. This is the main reason why fix methods dispatch through the instance, rather than the class like all other methods.

21.3.12. Old and aging notes below here

void AMCFinish(Pool pool)

.finish.forward: If the pool is being destroyed it is OK to destroy the forwarding buffers, as the condemned set is about to disappear.

void AMCBufferEmpty(Pool pool, Buffer buffer, Addr init, Addr limit)

.flush: Removes the connexion between a buffer and a group, so that the group is no longer buffered, and the buffer is reset and will cause a refill when next used.

.flush.pad: The group is padded out with a dummy object so that it appears full.

.flush.expose: The buffer needs exposing before writing the padding object onto it. If the buffer is being used for forwarding it might already be exposed, in this case the segment attached to it must be covered when it leaves the buffer. See .fill.expose.

.flush.cover: The buffer needs covering whether it was being used for forwarding or not. See .flush.expose.

Res AMCBufferFill(Addr *baseReturn, Addr *limitReturn, Pool pool, Buffer buffer, Size size, Bool withReservoirPermit)

.fill: Reserve was called on an allocation buffer which was reset, or there wasn’t enough room left in the buffer. Allocate a group for the new object and attach it to the buffer.

.fill.expose: If the buffer is being used for forwarding it may be exposed, in which case the group attached to it should be exposed. See .flush.cover.

Res AMCFix(Pool pool, ScanState ss, Seg seg, Ref *refIO)

.fix: Fix a reference to the pool.

Ambiguous references lock down an entire segment by removing it from old-space and also marking it grey for future scanning.

Exact, final, and weak references are merged because the action for an already forwarded object is the same in each case. After that situation is checked for, the code diverges.

Weak references are either snapped out or replaced with ss->weakSplat as appropriate.

Exact and final references cause the referenced object to be copied to new-space and the old copy to be forwarded (broken-heart installed) so that future references are fixed up to point at the new copy.

.fix.exact.expose: In order to allocate the new copy the forwarding buffer must be exposed. This might be done more efficiently outside the entire scan, since it’s likely to happen a lot.

.fix.exact.grey: The new copy must be at least as grey as the old as it may have been grey for some other collection.

Res AMCScan(Bool *totalReturn, ScanState ss, Pool pool, Seg seg)

.scan: Searches for a group which is grey for the trace and scans it. If there aren’t any, it sets the finished flag to true.

void AMCReclaim(Pool pool, Trace trace, Seg seg)

.reclaim: After a trace, destroy any groups which are still condemned for the trace, because they must be dead.

.reclaim.grey: Note that this might delete things which are grey for other collections. This is OK, because we have conclusively proved that they are dead – the other collection must have assumed they were alive. There might be a problem with the accounting of grey groups, however.

.reclaim.buf: If a condemned group still has a buffer attached, we can’t destroy it, even though we know that there are no live objects there. Even the object the mutator is allocating is dead, because the buffer is tripped.