* mps/code/trace.c (tracePropagateToLowerRanks): Move the hack from
traceFindGrey to here and simplify it a bit.
* (traceScanSegRes): Remove a debugging assertion that can't fail.
Disabling write barrier deferral in config.h, uncovered a second
problem: mps/test/function/eph1.c fails because segments aren't scanned
if the segment's summary doesn't intersect with the white set.
* mps/code/trace.c (traceFindGrey): Force rescanning by adding the white
set to the segments summary.
* mps/code/trace.c (TraceRankForAccess): While in the ephemeron band,
use RankFINAL for final segments and RankEXACT for the others.
* mps/test/function/eph5.c: New Test.
* mps/test/testsets/passing: Include eph5.c.
* mps/test/test/testlib/ephfmt.h (union header): Use a one-word header.
* mps/test/test/testlib/ephfmt.c: Update accordingly.
* mps/test/function/eph1.c: Update accordingly.
Ephemerons solve the key-in-value problem for weak key/value pairs. It
seems that it's not possible to solve that problem with weak pointers or
finalizers. Basically we want a weak key/value pair that implements the
rule:
The value is reachable, if the key is reachable.
The key-in-value problem occurs if all paths to the key go through the
value. In this situation, the key is considered reachable only if the
value is "externally reachable", i.e. the value is reachable through a
path other than through the weak pair.
Such weak pairs seem to need an extra phase in the GC algorithm. See
[1] for some variants of the algorithm.
https://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
* mps/code/mpmst.h (SegStruct): Add fields propagationNeeded,
marksChanged, and propagationFinished.
* mps/code/mpmtypes.h (RANK_LIST): Add EPHEMERON rank.
* mps/code/mps.s (mps_rank_ephemeron): New prototype.
* mps/code/mpscawl.h (mps_class_aeph, mps_fix_weak_pair)
(mps_fix_weak_or_pair, mps_fix_weak_and_pair): New prototypes.
* mps/code/mpsi.c (mps_rank_ephemeron): New function.
* mps/code/poolawl.c (AEPHSegStruct, AEPHPoolStruct): New. The
implementation the ephemeron pool.
* mps/code/seg.c (segAbsInit): Initialize propagationNeeded,
marksChanged, and propagationFinished.
* mps/code/trace.c (traceBandRetreat, tracePropagateToLowerRanks)
(moveSegToEndOfGreyRing): New helpers.
(TraceRankForAccess): Handle the RankEPHEMERON.
(traceFindGrey, traceScanSegRes): Add special handling for ephemeron
segments.
* mps/test/function/eph1.c: New test file.
* mps/test/function/eph2.c: New test file.
* mps/test/function/eph3.c: New test file.
* mps/test/function/eph4.c: New test file.
* mps/test/test/testlib/ephfmt.c: New file.
* mps/test/test/testlib/ephfmt.h: New file.
* mps/test/test/testlib/manifest: Add ephfmt.c.
* mps/test/testsets/passing: Include eph{1,2,3,4}.c.
This will allow us to reuse the scanning protocol with an arbitrary area
scanning function (replacing traceFormatScan) in order to implement
formatted object walking without an extra segment method.
Don't insist on scanning only grey segments: we want to be able to
reuse the scan protocol for walking, when the segments are black.
Clang 10 turns on -Wimplicit-int-float-conversion, and on 64-bit
platforms this issues a warning that implicit conversions to double
from Size and unsigned long "may lose precision".
This commit adds casts to (double) for all such conversions. The loss
of precision is either impossible in practice (because a double can
represent all integers up to 2**53, which is about 9 petabytes, well
beyond the addressing capabilities for current CPUs), or else
acceptable, because we are accumulating an approximate quantity like
"collection work" or "fill size" (that has to cope with loss of
precision due in any case), or computing a threshold like the spare
commit limit where it is acceptable for it to be somewhat approximate.
Publish minor changes to readme and configure files in main directory.
Main directory and code directory licence texts are now fully updated.
Copied from Perforce
Change: 196994
1. Being on the critical path is a property of the code location where the event is emitted, not of the event type, so indicate this using EVENT_CRITICAL macros rather than a field in the event table.
2. Compile out events on the critical path in hot varieties.
3. Remove TraceFixWhite event (redundant with TraceFixSeg).
4. No need for special handling of PoolFree event -- this are already skipped in hot varieties by dispatching directly to the pool class via PoolFreeMacro.
Copied from Perforce
Change: 195261
Split event ArenaAccess into ArenaAccessBegin and ArenaAccessEnd to avoid the need for the count field.
New events SegReclaim and SegScan.
Delete some redundant events: AMCFinish (PoolFinish), AMCFix, AMCFixForward, AMCFixInPlace (TraceFix), AMCGenCreate (GenInit), AMCGenDestroy (GenFinish), AMCInit (PoolInitAMC), AMCReclaim (SegReclaim), AMCScanBegin, AMCScanEnd (SegScan), ArenaWriteFaults (ArenaAccessBegin), PoolInitMV, TraceScanSeg (SegScan).
Add result code field to events ArenaAllocFail, CommitLimitSet, SegAllocFail.
Remove arena field from events PoolInitAMS, PoolInitMFS, PoolInitMVFF (already appeared in generic PoolInit event).
Copied from Perforce
Change: 195247
Eliminate duplication between policyCondemnChain and traceCondemnAll.
Compute better mortality estimate for collection of the world.
Copied from Perforce
Change: 194611
Add cross-reference from second-stage test in _mps_fix2 so that if the white set test is changed, the roots walking code can be changed to match.
In walkt0, add a smoke test for mps_arena_roots_walk.
Copied from Perforce
Change: 194507
Remove unused event AMCTraceEnd.
The fixEmergency method must have the same type as the fix method because of how they are used.
Copied from Perforce
Change: 192636
ServerID: perforce.ravenbrook.com
Place a cross-reference by each assertion the appears in the manual, to assist in keeping the manual up to date.
Copied from Perforce
Change: 192097
ServerID: perforce.ravenbrook.com
