1
Fork 0
mirror of git://git.sv.gnu.org/emacs.git synced 2026-01-26 15:10:50 -08:00
Commit graph

108 commits

Author SHA1 Message Date
Gareth Rees
2087688964 Fix problems identified by rb in review <https://info.ravenbrook.com/mail/2014/06/05/14-58-24/0/>.
Copied from Perforce
 Change: 186451
 ServerID: perforce.ravenbrook.com
2014-06-08 22:10:21 +01:00
Gareth Rees
a0e076be57 Improvements following review.
Copied from Perforce
 Change: 186227
 ServerID: perforce.ravenbrook.com
2014-05-21 00:43:06 +01:00
Gareth Rees
dbfe3ca425 Fix compilation on windows.
Fix bug in ArenaDescribeTracts (only described the first tract).

Copied from Perforce
 Change: 186215
 ServerID: perforce.ravenbrook.com
2014-05-20 19:19:14 +01:00
Gareth Rees
2f0ef9355d Clarify tracefix logic by unwinding the nested conditions.
Change the arena's chunk tree from a splay tree to an ordinary tree (so that it's not possible to accidentally splay it and leave it unbalanced).
New function TreeFindNext allows us to implement TractFirst and TractNext without having to splay the tree.
Make sure all operations on the chunk tree leave it balanced. But don't balance the tree directly in ChunkFinish() because this is only ever called in a loop where multiple chunks are being deleted from the tre. Instead use the sequence TreeToVine -- iterate and delete -- TreeBalance. The new macro TREE_DESTROY assists with this.
No need any more for ArenaIsReservedAddr, CHUNK_OF_ADDR, TRACT_OF_ADDR.
Update design documentation.

Copied from Perforce
 Change: 186212
 ServerID: perforce.ravenbrook.com
2014-05-20 18:12:37 +01:00
Gareth Rees
47befaa46b Use treefind instead of splaytreefind to search the chunk tree.
Balance the chunk tree after insertion and deletion.
Avoid calling TractFirst and TractNext in ArenaDescribeTracts and PoolOfRange.

Copied from Perforce
 Change: 186199
 ServerID: perforce.ravenbrook.com
2014-05-19 20:19:50 +01:00
Gareth Rees
632ac692a5 Restore "avoid" mechanism in arenaallocpage.
Make sure that we can tear down the arena if ArenaCreate fails:
1. Don't set hasFreeCBS until the block pool has some pages.
2. Finish the CBS block pool in ArenaFinish, not ArenaDestroy.
3. Delete the chunk from the arena's free CBS before destroying the chunk, just in case the chunk contains pages from the CBS's block pool.

Copied from Perforce
 Change: 186177
 ServerID: perforce.ravenbrook.com
2014-05-19 12:24:11 +01:00
Gareth Rees
e2d346aa67 No need to store primary chunk in the closure: can get it via the arena.
Copied from Perforce
 Change: 186164
 ServerID: perforce.ravenbrook.com
2014-05-18 22:27:07 +01:00
Gareth Rees
0b0a465674 Replace the chunk ring with a chunk tree.
Fix bug in SplayFindNext (LESS and GREATER the wrong way round).

Copied from Perforce
 Change: 186157
 ServerID: perforce.ravenbrook.com
2014-05-17 17:05:42 +01:00
Gareth Rees
d3d2795fce Improve control over checking:
1. Where Type is a pointer type with a signature, replace CHECKL(TypeCheck(val)) with CHECKD(Type, val).
2. Where Type is a pointer type with no signature, replace CHECKL(TypeCheck(val)) with CHECKD_NOSIG(Type, val).
3. Where Type is a pointer type with a signature, but the structure is not visible at point of checking, replace CHECKL(TypeCheck(val)) with CHECKD_NOSIG(Type, val). Reference <design/check/#.hidden-type>
4. Make BTCheck extern and use it where possible.
5. Replace AVER(TypeCheck(val)) with AVERT(Type, val).

Copied from Perforce
 Change: 185263
 ServerID: perforce.ravenbrook.com
2014-04-06 22:51:05 +01:00
Gareth Rees
9bb6b2f8be Tidy-up of attributes and pool classes:
* Bring design up to date.
* New function PoolHasAttr encapsulates attribute checking.
* Abstract classes are abstract and mustn't be checked.
* The dummy pool class in fotest needs a size.
* Abstract pool classes null out methods that they can't provide a generic implementation for, to force subclasses to provide one.
* New function PoolTrivFramePopPending provides a generic implementation of that method.
* Rename PoolNoFreeWalk to PoolTrivFreeWalk since it has NOOP rather than NOTREACHED.
* Check that AttrMOVINGGC implies AttrGC.
* Remove unimplemented attributes (BUF_RESERVE, BUF_ALLOC, INCR_RB, INCR_WB, PM)
* AMC now inherits from AMCZ instead of the other way round. This is simpler: AMC adds features to AMCZ rather than AMCZ taking features away (and not quite getting it right).
* Similarly, LO inherits from AbstractSegBufPoolClass + PoolClassMixInCollect so that it doesn't have to clear AttrSCAN and the scan methods.
* Fix bug in MFSCheck -- mustn't check unroundedUnitSize >= UNIT_MIN since small unit sizes are rounded up to UNIT_MIN.
* Don't see AttrFREE in MRG (since no free method is supplied).
* Check AttrSCAN systematically (in PoolScan and SegCheck) rather than opportunistically in TraceStart and TraceQuantum.

Copied from Perforce
 Change: 185231
 ServerID: perforce.ravenbrook.com
2014-04-04 17:05:08 +01:00
Gareth Rees
673eea41a0 Check classes after defining them.
Copied from Perforce
 Change: 185228
 ServerID: perforce.ravenbrook.com
2014-04-04 12:51:07 +01:00
Gareth Rees
feaa22a74f Remove unnecessary calls to mps_args_done -- these crept back in via the merge of the cbs-tract-alloc branch.
Copied from Perforce
 Change: 185143
 ServerID: perforce.ravenbrook.com
2014-04-01 22:43:58 +01:00
Gareth Rees
9d5266c0cc Improved support for pelles c:
* Refactor nmake files so that a compiler-specific makefile is included; move Microsoft Visual C-specific options to mv.nmk.
* Add nmake files for Pelles (w3i3pc.nmk and pc.nmk).
* Rename spw3i3mv.c to spw3i3.c and spw3i6mv.c to spw3i6.c since these are also used by Pelles C.
* Make reasonable changes to the source code to avoid warnings from Pelles C:
** check results of function calls;
** avoid useless return values;
** undef max before defining it;
** ensure printf formats are checkable;
** move notreached() assertions to the end of blocks;
** suppress warnings in cases where the code shouldn't be changed ("Unreachable code", "Inline assembly code is not portable", "Structured Exception Handling is not portable").

Copied from Perforce
 Change: 184977
 ServerID: perforce.ravenbrook.com
2014-03-24 18:23:29 +00:00
Richard Brooksby
061eba523c Reponding to code review <https://info.ravenbrook.com/mail/2014/03/10/17-01-58/0/>.
Renaming the arena’s “zoned CBS” back to earlier “free CBS”, especially since it might not be zoned.

Copied from Perforce
 Change: 184738
 ServerID: perforce.ravenbrook.com
2014-03-11 17:27:57 +00:00
Richard Brooksby
2ce13f82dc Incomplete response to code review. see <https://info.ravenbrook.com/mail/2014/03/07/11-21-03/0/>.
Copied from Perforce
 Change: 184735
 ServerID: perforce.ravenbrook.com
2014-03-11 16:39:11 +00:00
Richard Brooksby
2b9719e4bd Catch-up merge from master.
Copied from Perforce
 Change: 184732
 ServerID: perforce.ravenbrook.com
2014-03-11 16:30:47 +00:00
Richard Brooksby
1956b0cee1 Restoring error path for arenaalloc.
Copied from Perforce
 Change: 184604
 ServerID: perforce.ravenbrook.com
2014-02-28 13:48:56 +00:00
Richard Brooksby
fcd691aaa3 Adding unzoned arena tests back, now that there is a method of asking for them.
Copied from Perforce
 Change: 184602
 ServerID: perforce.ravenbrook.com
2014-02-28 13:46:36 +00:00
Richard Brooksby
53bdfaab23 More closely duplicating the allocation policy from the old vmarena, and documenting why.
Copied from Perforce
 Change: 184601
 ServerID: perforce.ravenbrook.com
2014-02-28 13:35:51 +00:00
Richard Brooksby
4eec2493b7 Adding a keyword to the arena to disable zoned allocation.
Copied from Perforce
 Change: 184599
 ServerID: perforce.ravenbrook.com
2014-02-28 13:20:06 +00:00
Richard Brooksby
887c6e3845 Tidying up interfaces to cbsinit and cbsfindinzones.
Documenting assumptions about CBS block allocation used by the arena.

Copied from Perforce
 Change: 184593
 ServerID: perforce.ravenbrook.com
2014-02-28 10:30:32 +00:00
Richard Brooksby
14620f8c24 Simplifying the interface to cbsfindinzones.
Speeding up CBSFindInZones a bit.

Copied from Perforce
 Change: 184589
 ServerID: perforce.ravenbrook.com
2014-02-28 01:25:18 +00:00
Richard Brooksby
ddfec2c708 Eliminating complex zonedcbs adt and reintroducing a cbs with zone summaries. removing the default “high” segment preference, that caused the cbs trees to unbalance, and isn’t required outside of ep. considerable reduction in code and speed up, now at parity with master on test xc/release/gcbench -x 1234 amc
Copied from Perforce
 Change: 184587
 ServerID: perforce.ravenbrook.com
2014-02-28 00:35:05 +00:00
Richard Brooksby
7a33207595 Adding some missing headers revealed by separate compilation.
Copied from Perforce
 Change: 184584
 ServerID: perforce.ravenbrook.com
2014-02-27 18:01:05 +00:00
Richard Brooksby
2d685bee13 Adding check for unexpected case of zone barging.
Copied from Perforce
 Change: 184583
 ServerID: perforce.ravenbrook.com
2014-02-27 17:50:44 +00:00
Richard Brooksby
e6d1db24b6 Simplifying interface to high and low allocation in the zoned cbs.
Copied from Perforce
 Change: 184576
 ServerID: perforce.ravenbrook.com
2014-02-27 12:01:42 +00:00
Richard Brooksby
b160cf778a Separating arena cbs allocator into abstract zonedcbs adt. the number of chunks seems to blow up.
Copied from Perforce
 Change: 184575
 ServerID: perforce.ravenbrook.com
2014-02-27 04:50:52 +00:00
Richard Brooksby
3b3e01b10f Searching the freecbs now respects the “high” preference, though not yet within blocks.
Copied from Perforce
 Change: 184574
 ServerID: perforce.ravenbrook.com
2014-02-26 19:29:05 +00:00
Richard Brooksby
6b16beb0c6 Implemented “high” preference allocation for zone stripes.
Copied from Perforce
 Change: 184573
 ServerID: perforce.ravenbrook.com
2014-02-26 18:18:20 +00:00
Richard Brooksby
dc83b10c0d Fixing cbsfindfirstinzones to return a res, because it can fail when it can’t allocate a cbs block, unlike cbsfindfirst.
Fixing related corner case in arenaAllocFindInFreeCBS when it thought it couldn’t find a block when it was in fact running out of nodes.  This was revealed by “gcbench --npass 1 --arena-size 1M --seed 945632066 --niter 1 amc”.
Fixing draft ArenaFreeCBSDelete after exercising it.

Copied from Perforce
 Change: 184568
 ServerID: perforce.ravenbrook.com
2014-02-26 17:40:39 +00:00
Richard Brooksby
f215460eb1 Adding checking of the freecbs and zonecbss to arenacheck.
Copied from Perforce
 Change: 184563
 ServerID: perforce.ravenbrook.com
2014-02-26 14:25:01 +00:00
Gareth Rees
6971bee496 Make mps_args_done obsolete.
Copied from Perforce
 Change: 184543
 ServerID: perforce.ravenbrook.com
2014-02-26 12:47:27 +00:00
Richard Brooksby
3da21234fd Implementing arenafreecbsdelete that might work on chunks that have interesting patterns of zone allocation.
Further tidying up.

Copied from Perforce
 Change: 184531
 ServerID: perforce.ravenbrook.com
2014-02-25 19:19:09 +00:00
Richard Brooksby
d1b6ad25c7 Tidying up and wrapping arena cbs operations.
Copied from Perforce
 Change: 184530
 ServerID: perforce.ravenbrook.com
2014-02-25 16:58:23 +00:00
Richard Brooksby
460f988a46 Removing the freering allocation cache, which breaks the cbs abstraction and does not give any benefit once we’re allocating from zonecbss.
Copied from Perforce
 Change: 184529
 ServerID: perforce.ravenbrook.com
2014-02-25 11:35:18 +00:00
Richard Brooksby
5a03d625e2 Tidying up arena cbs mfs teardown.
Tidying up arena CBS initialisation and teardown.

Copied from Perforce
 Change: 184523
 ServerID: perforce.ravenbrook.com
2014-02-25 11:16:12 +00:00
Richard Brooksby
8115dbd700 Quick hack at allocating from per-zone cbs. promising performance.
Copied from Perforce
 Change: 184507
 ServerID: perforce.ravenbrook.com
2014-02-25 01:51:27 +00:00
Richard Brooksby
c58eed5c6b Attempting to combine fast page allocation with cbs allocation. currently not working properly, as the cbs and free page rings overlap.
Copied from Perforce
 Change: 184471
 ServerID: perforce.ravenbrook.com
2014-02-19 11:04:18 +00:00
Richard Brooksby
5afbb9d290 Adding per-zone arena cbss, but they aren’t filled with any free space yet.
Copied from Perforce
 Change: 184470
 ServerID: perforce.ravenbrook.com
2014-02-18 13:53:56 +00:00
Richard Brooksby
f0b07f7bbe Added a pool argument to cbs creation so that cbss can share a pool. in particular, per-zone allocation cbss for the arena can share a single special block pool.
Copied from Perforce
 Change: 184469
 ServerID: perforce.ravenbrook.com
2014-02-18 13:53:56 +00:00
Richard Brooksby
2151d00c7f Catch-up merge from master, mainly to pick up spare-ring merge.
Copied from Perforce
 Change: 184365
 ServerID: perforce.ravenbrook.com
2014-02-18 13:47:49 +00:00
Richard Brooksby
3a1236d0ea Generic chunkinit and chunkdestroy now maintain their memory in the freecbs so that the client arena should also work.
ArenaDestroy fixed so that ChunkDestroy can go ahead even though the freeCBS has been torn down.  Allows completion of test programs.

Copied from Perforce
 Change: 184295
 ServerID: perforce.ravenbrook.com
2014-02-04 18:50:56 +00:00
Richard Brooksby
252383465b Abolishing vmarenaalloc and its horrible relatives.
Removing redundant SegPref structures from pools that only ever used the default.
Abolishing the VMNZ arena class, to be replaced with an option to ArenaCreate.

Copied from Perforce
 Change: 184293
 ServerID: perforce.ravenbrook.com
2014-02-04 18:10:53 +00:00
Richard Brooksby
2c3c67e73e Deleting a chunk’s address range from the arena’s freecbs when the chunk is destroyed. otherwise, the memory gets allocated and an assertion fires when that memory is not in any chunk.
Copied from Perforce
 Change: 184281
 ServerID: perforce.ravenbrook.com
2014-02-01 00:34:52 +00:00
Richard Brooksby
ae6362a7ce Adding arena extension back to the arena allocation policy.
Copied from Perforce
 Change: 184280
 ServerID: perforce.ravenbrook.com
2014-01-31 22:47:29 +00:00
Richard Brooksby
66a05e1f05 Changes from review comments made in <https://info.ravenbrook.com/mail/2014/01/31/13-40-25/0/>.
ArenaSpareCommitExceeded changed to ArenaPurgeSpare so that the arena can properly reduce the amount of spare memory when setting the commit limit etc.
pagesMarkAllocated promptly unmaps page table pages in its error path.
Miscellaneous clarifications and extra checking.

Copied from Perforce
 Change: 184267
 ServerID: perforce.ravenbrook.com
2014-01-31 15:21:21 +00:00
Richard Brooksby
ce329f192e Abolishing the spare pages bit table.
Not quite working yet. This code fails to unmap unused page table pages, and so triggers an assertion at chunk finish.

Copied from Perforce
 Change: 184246
 ServerID: perforce.ravenbrook.com
2014-01-30 19:58:36 +00:00
Richard Brooksby
7a73e98638 Refactored implementation of preferenced allocation policy from vm arena to general purpose arena.
Eliminating SegPrefExpress in most places, especially where there wasn’t really any preference.
Eliminating special case knowledge about garbage collection from the arena.

Copied from Perforce
 Change: 184279
 ServerID: perforce.ravenbrook.com
2014-01-30 18:08:10 +00:00
Richard Brooksby
d75cfe191c Incorporating early commit limit checking into arena allocation policy.
Copied from Perforce
 Change: 184240
 ServerID: perforce.ravenbrook.com
2014-01-30 15:44:46 +00:00
Richard Brooksby
973575140f Abstracting the allocation policy from arenaalloc in order to implement arena extension policy.
Copied from Perforce
 Change: 184239
 ServerID: perforce.ravenbrook.com
2014-01-30 15:44:46 +00:00