100 Commits

This Branch

This Branch

All Branches

Author	SHA1	Message	Date
andrew	76a7e17b29	Update readme benchmarks ... I was incorrectly linking to an old build of conflict set, before the conflict-set name change. I don't know what the regression is from, but update the README for transparency now.	2024-03-27 16:12:45 -07:00
andrew	5cf43d1bfa	Add weaselab namespace	2024-03-27 16:07:05 -07:00
andrew	25cc427ec5	Assert safe_free size is correct in debug builds ... Closes #16	2024-03-27 15:34:24 -07:00
andrew	c15c2e7b44	Remove redundant static assert ... Closes #14	2024-03-27 12:41:45 -07:00
andrew	a4d1f91670	Update README benchmark after adding getBytes	2024-03-20 16:26:05 -07:00
andrew	b7cdecaf71	Document thread-safety in terms of constness	2024-03-20 16:16:15 -07:00
andrew	cda28643a6	Fix ConflictSet_getBytes	2024-03-20 16:15:43 -07:00
andrew	cdb5360b9a	Allow _GLOBAL_OFFSET_TABLE_ usage	2024-03-20 12:39:18 -07:00
andrew	ef224a60f4	Allow use of __tls_get_addr	2024-03-20 12:29:18 -07:00
andrew	6222b74787	Fix tests ... Add the new symbol, and update the valgrind client request so that Node::partialKeyCapacity is defined.	2024-03-20 12:21:59 -07:00
andrew	19edc6f78f	Interface change! Add ConflictSet::getBytes ... Closes #12	2024-03-20 12:11:34 -07:00
andrew	3f9d01c46a	Users will now do find_package(conflict-set)	2024-03-19 19:05:54 -07:00
andrew	db03c6f901	Fix invalid package name for debian	2024-03-19 17:43:58 -07:00
andrew	c1698b040b	Disable tsan for debug builds ... Too slow	2024-03-19 16:54:29 -07:00
andrew	2e08b54785	Update README.md with 1.0 benchmarks ... Closes #7	2024-03-19 16:41:59 -07:00
andrew	aa6f237d50	Document and test thread safety properties ... Closes #2	2024-03-19 16:27:24 -07:00
andrew	becfd25139	De-templatize kUseFreeList ... This results in smaller code. This is part of the "let the compiler be in charge of inlining decisions" theme.	2024-03-19 15:12:31 -07:00
andrew	d78b36821b	Remove more redundant nullptr checks	2024-03-19 15:05:34 -07:00
andrew	ce79b47fbe	Revert 303b368fc5 ... This is a code-size / speed tradeoff. Maybe it's a good idea here, but it's a bit weird to do this in some places and not others (there are many places we can avoid switching on type this way). The compiler can inline and then dead code eliminate to achieve the same effect, so we'll just let the compiler be in charge of inlining decisions.	2024-03-19 15:01:13 -07:00
andrew	727b7e642a	Save more redundant nullptr checks	2024-03-19 14:34:59 -07:00
andrew	cb4c2b7e1e	Avoid redundant null check in some cases	2024-03-19 14:29:13 -07:00
andrew	ef9b789745	Remove unused code	2024-03-19 11:12:03 -07:00
andrew	edd7bcaa1e	Check ConflictSet preconditions in script_test	2024-03-19 10:31:04 -07:00
andrew	be8ac879c5	Update README benchmarks	2024-03-18 17:24:26 -07:00
andrew	83c7f66d67	Remove some redundant nullptr checks	2024-03-18 16:22:24 -07:00
andrew	a5710b8282	Remove performance-only code from debug build for increased coverage	2024-03-18 16:21:55 -07:00
andrew	c31eebd5de	No caller of CheckRangeRightSide::downLeftSpine has null n	2024-03-18 16:03:41 -07:00
andrew	ddeb059968	Remove more dead code	2024-03-18 16:01:56 -07:00
andrew	5a0bcf9a5a	Strengthen precondition to checkRangeStartsWith ... and remove resulting dead code	2024-03-18 15:44:44 -07:00
andrew	97717cec86	Remove suspected dead code ... Removing it is definitely safe. I suspect that any way to get here would have already returned from checkRangeRead during the search for the common prefix	2024-03-18 15:30:15 -07:00
andrew	6a13c43a78	Remove mistakenly checked in printfs	2024-03-18 14:03:09 -07:00
andrew	c6c438bae2	Address more missing coverage ... cc #3. Found some non-trivial dead code this time!	2024-03-18 14:01:30 -07:00
andrew	7d4f832b43	Address some missing coverage ... cc #3	2024-03-18 13:36:35 -07:00
andrew	5b0c3c2428	Exclude more from coverage ... Only count ConflictSet.cpp in jenkins now, and also add some more excludes for unreachable lines	2024-03-18 12:05:52 -07:00
andrew	f2b5e9b0bf	Change max key len to 8, update corpus ... Now that we don't have a fixed buffer reserved for partial key bytes, there's nothing (obvious) that makes testing short versus long keys much different. maybeDecreaseCapacity is an exception, and we'll write some tests covering that manually.	2024-03-18 11:55:43 -07:00
andrew	8e0e65dac6	Count implicit key byte for maybeDecreaseCapacity	2024-03-18 11:42:53 -07:00
andrew	5aab76847a	Add inner-full-word test	2024-03-17 14:42:11 -07:00
andrew	1a51aa00e5	Run scripted tests	2024-03-16 22:04:30 -07:00
andrew	3975bada0c	Add a simple language for scripting tests	2024-03-16 22:02:00 -07:00
andrew	a5330b6e23	Minor paper tweaks	2024-03-16 12:45:36 -07:00
andrew	2e246ec6a4	Cachegrind says this is fewer instructions	2024-03-16 08:46:41 -07:00
andrew	6d7e3c9849	Fix invalid offsetof warnings	2024-03-15 21:37:21 -07:00
andrew	671da5d096	Collect gcc warnings again in Jenkins	2024-03-15 19:00:01 -07:00
andrew	303b368fc5	Add Type template parameter to maybeDownsize to avoid branch	2024-03-15 18:05:32 -07:00
andrew	9f5a68e2c0	Use plain loop for Node3	2024-03-15 17:35:14 -07:00
andrew	dfbb3ce5f1	Use assume ... It works now that we fell back to the __builtin_unreachable based implementation for gcc.	2024-03-15 17:23:02 -07:00
andrew	e7719b6e0b	Disable sanitizers when cross-compiling	2024-03-15 17:04:49 -07:00
andrew	83fedf1f9e	Remove dead code	2024-03-15 17:00:20 -07:00
andrew	8556caf360	Cachegrind says memset uses fewer instructions here	2024-03-15 16:57:16 -07:00
andrew	9d13ca84f5	Fix bug spotted by hand. No test coverage there yet	2024-03-15 16:55:34 -07:00
andrew	a79436ee9b	Use statically-known numChildren in Node3 -> Node16	2024-03-15 16:52:40 -07:00
andrew	e9c8537cf2	Copy Node members and set children pointers in copyChildrenAndKeyFrom	2024-03-15 16:28:58 -07:00
andrew	5b988efe6f	Consolidate copyChildrenAndKeyFrom implementations	2024-03-15 16:12:44 -07:00
andrew	e35d698b21	Improve copyChildrenAndKeyFrom-related codegen	2024-03-15 08:19:16 -07:00
andrew	30496d14e7	Don't print out filename in TestDriver.cpp	2024-03-15 07:44:11 -07:00
andrew	eb93157ddf	Improve codegen in freeAndMakeCapacityAtLeast	2024-03-14 19:23:49 -07:00
andrew	9cafef8bbb	Update benchmarks after fixing skip list bug in a9b3d3d	2024-03-14 15:50:55 -07:00
andrew	6f81580953	Guard SHOW_MEMORY-only code	2024-03-14 15:40:09 -07:00
andrew	429fe5baed	Fix some uses of uninitialized memory ... Introduced in ee36bda8f8, unsurprisingly.	2024-03-14 15:39:37 -07:00
andrew	a0451e4423	Give radix tree setOldestVersion an extra 10 work per call ... It seemed to really be important for the skip list. I'm cargo culting this a little bit here.	2024-03-14 15:32:39 -07:00
andrew	a9b3d3d1c9	Show peak memory in skip list, and fix setOldestVersion bug ... It was not previously gc'ing faster than it was writing.	2024-03-14 15:31:29 -07:00
andrew	b817e3c749	Track malloc size with a header for SHOW_MEMORY	2024-03-14 15:30:46 -07:00
andrew	ee36bda8f8	Track initializedness of Node memory more precisely ... By not initializing Node members with dummy default values. This has performance/code size benefits, and improves debugging when running under valgrind. Unfortunately this also makes it easy to write code that uses uninitialized memory, so if valgrind doesn't have good coverage then we might let some uninit usages sneak through. We plan to have good coverage for valgrind, so I think it's ok. If writing correct code becomes too tedious then we can go back to initializing Node fields with dummy default values.	2024-03-14 14:58:20 -07:00
andrew	a8f4bd91c8	Use asan and ubsan for whitebox/fuzz tests	2024-03-14 13:47:21 -07:00
andrew	35086ee66a	Add GCOVR_EXCL_LINE to default labels too	2024-03-13 20:54:05 -07:00
andrew	0f795cf163	Set -DNVALGRIND for release artifacts	2024-03-13 20:45:59 -07:00
andrew	a07c93ffff	New switch idiom ... Add -Wswitch-enum -Werror=switch-enum to enforce that we explicitly handle all cases, and add a default: __builtin_unreachable() to all switches to make un-enumerated values UB	2024-03-13 20:07:27 -07:00
andrew	c68f563017	"partial capacity bytes" -> "partial key capacity bytes"	2024-03-13 18:51:48 -07:00
andrew	6b6a9bace9	Fix SHOW_MEMORY for gcc and glibc on linux	2024-03-13 16:57:38 -07:00
andrew	3cb0765fdd	Rework SHOW_MEMORY ... closes #10	2024-03-13 16:48:28 -07:00
andrew	351ff3df3b	Cave in and just add the unreachable's gcc wants	2024-03-13 14:58:34 -07:00
andrew	e818648cdc	makeCapacityAtLeast -> freeAndMakeCapacityAtLeast	2024-03-13 14:01:47 -07:00
andrew	12540b8713	maybeDecreaseCapacity policy was too strong ... We can allow larger capacities and still pay for the key bytes	2024-03-13 14:00:01 -07:00
andrew	c2606cd26a	Add clang build in jenkins and record issues for clang ... I can't seem to get gcc to do what I want for this control flow warning thingy	2024-03-13 13:39:39 -07:00
andrew	4b72fc0b7b	Try -fstrict-enums ... https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87951 suggests this might make the control-flow warning go away	2024-03-13 13:28:09 -07:00
andrew	a9caa0249e	Use plain enum for type ... My understanding is that this it's now UB for any value that's not listed to inhabit Type, so the gcc warning about control reaching the end of a function with an exhaustive switch where every case returns should go away now.	2024-03-13 13:20:48 -07:00
andrew	08b2b7f41a	Move comment to above field it's commenting on	2024-03-13 12:38:16 -07:00
andrew	26bd8b94cc	Lower kBytesPerKey to 144 by changing Node4 to Node3	2024-03-13 12:29:01 -07:00
andrew	55eaef5b1d	Remove 16 bytes from Node0	2024-03-13 10:59:25 -07:00
andrew	797e6b4a3e	Use switch for type dispatch throughout	2024-03-13 10:59:19 -07:00
andrew	ee86b5289b	Rearrange induction inequalities ... This looks nicer IMO	2024-03-13 07:31:24 -07:00
andrew	b779c0f6f7	Fix induction (again). Now it's 176 bytes per key	2024-03-13 07:26:49 -07:00
andrew	ef802b8acd	Only skip free list from maybeDecreaseCapacity	2024-03-12 17:33:28 -07:00
andrew	5371c2bede	Change kMinChildrenNode4 to 2, fixing the induction	2024-03-12 16:34:20 -07:00
andrew	75c304bbe7	WIP my understanding of the memory bound induction was wrong	2024-03-12 15:56:34 -07:00
andrew	aefb83dbc6	Prepare for erase to invalidate children of parent	2024-03-12 15:54:21 -07:00
andrew	b0ac7e41b9	Update corpus	2024-03-12 14:52:43 -07:00
andrew	4b6b2747bf	Relax capacity property so that it _eventually_ needs to hold ... Doing otherwise turned out to be prohibitively inconvenient closes #9	2024-03-12 12:48:52 -07:00
andrew	1496aa106b	Avoid an unnecessary node0 to node4 transition	2024-03-12 11:44:58 -07:00
andrew	71e117965e	Fix issue with getSearchPath on nullptr on setOldestVersion	2024-03-12 11:02:51 -07:00
andrew	471b276947	Downsize nodes in erase	2024-03-12 10:31:36 -07:00
andrew	b721bc80a9	Document that erase may invalidate search path ... It doesn't yet, but it will when we downsize nodes.	2024-03-11 22:47:48 -07:00
andrew	5e4eab55fb	Avoid re-inserting begin if begin is not a prefix of end	2024-03-11 22:43:52 -07:00
andrew	1dcb380c73	Use getInTree in insert	2024-03-11 22:22:49 -07:00
andrew	87d650ff00	Change eraseChild to erase	2024-03-11 21:54:20 -07:00
andrew	b8f6a8edf2	sizeof(Node0) also needs to be < kBytesPerKey ... Also remove vestigial comment. CC #9	2024-03-11 18:21:00 -07:00
andrew	01f1d5850f	Create a Node0 when splitting existing partial key	2024-03-11 18:20:14 -07:00
andrew	cd567383c3	Only keep the assume's that actually improve codegen	2024-03-10 14:34:55 -07:00
andrew	53a442abf9	Use the assume attribute for gcc	2024-03-10 14:22:47 -07:00
andrew	6e212847ac	Add assume macro ... Validated with cachegrind that this reduces instructions executed	2024-03-09 19:45:54 -08:00