andrew
04ac41a7e7
Add memory usage and some "check" metrics
2024-07-12 14:22:16 -07:00
andrew
3288c583e4
Make metrics thread-safe
...
Even concurrently with calling non-const methods on the associated
ConflictSet
2024-07-12 13:22:02 -07:00
andrew
3ac16bc966
Propose a metrics interface
2024-07-12 12:00:46 -07:00
andrew
4182d904c5
Fix an issue where gc wasn't outpacing writes
2024-07-11 16:40:18 -07:00
andrew
65f8462e88
Remove dead code
2024-07-08 11:14:30 -07:00
andrew
46e01af027
Specialize scan16 for Node16 in checkMaxBetweenExclusive
2024-07-08 11:01:17 -07:00
andrew
c9d0d72684
Remove some branches for Node3 in checkMaxBetweenExclusive
2024-07-07 20:45:20 -07:00
andrew
9046dc5a8f
Only switch on type once in checkMaxBetweenExclusive
2024-07-07 17:28:45 -07:00
andrew
e2927bf0fa
Simplify condition for TooOld in check
2024-07-03 16:39:22 -07:00
andrew
75a2b8d06c
Remove bogus assert
...
It's bogus since we mess with oldestVersionFullPrecision in addWrites
2024-07-03 14:47:50 -07:00
andrew
76df63a9d7
Allow writeVersion, oldestVersion, and readVersion to span 2e9
2024-07-03 14:19:50 -07:00
andrew
3db3d975fc
Interface change! Allow decreasing setOldestVersion
2024-07-03 10:31:57 -07:00
andrew
982b31af34
Explicitly convert uint32x4_t to int32x4_t
2024-07-03 10:01:00 -07:00
andrew
cc716ef16b
Attempt to fix memory leak
2024-07-03 09:48:02 -07:00
andrew
3e6be6bd83
Allow write version to jump by more than 2e9
2024-07-03 09:07:09 -07:00
andrew
3e2c8310bb
Fix several bugs related to extant versions too old
2024-07-03 07:53:03 -07:00
andrew
8264f1342d
Fix some precision issues
2024-07-02 13:55:05 -07:00
andrew
5d7e9c6f85
Compare against oldestVersionFullPrecision at full precision
2024-07-02 13:53:56 -07:00
andrew
cdf42fcb34
Add gcScanStep
...
Also remove oldestVersion arg in write call tree
2024-07-02 13:06:53 -07:00
andrew
a04e81b3ff
Add version window constants
2024-07-02 12:45:35 -07:00
andrew
68ab9a9f08
Commit to 32-bit versions
2024-07-02 10:30:22 -07:00
andrew
01488880ef
Fix outdated comment
2024-07-01 17:27:04 -07:00
andrew
bb84792cff
Use avx512f across entire checkRangeRead call tree
...
Keeping the avx fallback
2024-07-01 13:07:34 -07:00
andrew
66bd799f05
Fix hasty change about end invariants
...
`end` is allowed to be <= 0 at the beginning of
checkMaxBetweenExclusive, but not later after checking for the first
range version.
2024-06-30 21:20:05 -07:00
andrew
2646d5eaf1
Get back to 100% coverage
...
Closes #30
This is achieved by running libfuzzer with USE_32_BIT_VERSIONS={OFF,ON},
and then combining the corpora. I suspect that the problem earlier was
that we only had the 32 bit corpus but were measuring coverage for 64
bit in jenkins.
2024-06-30 21:15:42 -07:00
andrew
0367ba9856
Fast path for prefix reads
2024-06-30 20:33:54 -07:00
andrew
a68ad5dd17
Interface change! Return TooOld after 2e9 versions
...
Event if setOldestVersion wasn't called
2024-06-30 15:28:51 -07:00
andrew
8e3eacb54f
Apply function multi versioning higher in call stack to save branches
2024-06-30 13:30:44 -07:00
andrew
c52d50f4f9
Remove bounds from forEachInRange
2024-06-29 22:53:51 -07:00
andrew
447da11d59
Remove obsolete optimizations
...
These are a relic of when we used forEachInRange for checkMaxBetween
2024-06-29 22:47:42 -07:00
andrew
daa8e02d4f
Fixes from testing on an avx512f-capable machine
2024-06-29 22:41:39 -07:00
andrew
fd3ea2c2a8
clang-format fixes
2024-06-29 22:21:50 -07:00
andrew
0b839b9d7e
Fixes for symbol multi-versioning with avx512f
2024-06-29 22:20:50 -07:00
andrew
11a022dcf7
Attempt at avx512f 32bit compare
2024-06-29 21:56:21 -07:00
andrew
94da4c72a5
Fix clang-format
2024-06-29 15:11:40 -07:00
andrew
461e07822a
32-bit x86 simd for the other scan16 too
2024-06-29 15:10:36 -07:00
andrew
75499543e7
Fix clang-format
2024-06-29 15:03:44 -07:00
andrew
81f44d352f
SIMD scan16 for x86 + 32-bit versions
2024-06-29 15:01:18 -07:00
andrew
45da8fb996
Use the faster unvectorized implementation for Node3
2024-06-28 22:52:39 -07:00
andrew
4958a4cced
Make always_inline function inline
...
Try to fix warning in jenkins
2024-06-28 22:42:01 -07:00
andrew
648b0b9238
Add an always_inline, with explanatory comment
2024-06-28 19:55:33 -07:00
andrew
d3f4afa167
More SIMD for scanning Node256 with 32-bit versions
2024-06-28 19:48:06 -07:00
andrew
f762add4d6
Write vectorized 32-bit compare by hand for arm in scan16
2024-06-28 17:28:59 -07:00
andrew
b311e5f1f0
Add an experimental, disabled 32 bit internal version
...
I think it's only missing detection for full-precision versions more
than 2e9 apart
2024-06-28 15:53:35 -07:00
andrew
ff81890921
Rename MaxVersionT to InternalVersionT
2024-06-28 14:34:48 -07:00
andrew
0e96177f5c
Allow to easily experiment with 32 bit "max version" type
2024-06-28 13:54:12 -07:00
andrew
efb0e52a0a
SIMD implementation of scan16 for x86
...
Closes #29
2024-06-27 22:21:41 -07:00
andrew
2df7000090
Remove switch on phase from Stepwise left/right step
2024-06-27 20:51:35 -07:00
andrew
5378a06c39
Vectorize all bounds checks for Node256 scan
2024-06-27 17:40:23 -07:00
andrew
12c6ed2568
Reorganize to prepare for better vectorized first/last page
2024-06-27 17:21:41 -07:00
andrew