1. 08 Aug, 2023 1 commit
  2. 07 Aug, 2023 2 commits
  3. 04 Aug, 2023 4 commits
    • Release the `DFATAL` pseudo-LogSeverity level · 70172ada
      `DFATAL` is defined as `FATAL` in debug mode, and as `ERROR`
      when `NDEBUG` is defined.
      
      Closes #1279
      
      PiperOrigin-RevId: 553904244
      Change-Id: Iaa207ee65b2a39b4b7f5da241208c3d39cd5da0e
      Derek Mauro committed
    • Import of CCTZ from GitHub. · 659b77b7
      PiperOrigin-RevId: 553878129
      Change-Id: I054a5bd4c9011155c9fe03df0f07803ad7d2ade3
      Abseil Team committed
    • Store infoz on the heap instead of inline and store it only when we are sampling… · 8f241e77
      Store infoz on the heap instead of inline and store it only when we are sampling the current allocation.
      
      PiperOrigin-RevId: 553847957
      Change-Id: Idd131d0362bf36bd22d9bd20df54bd9ae50f0e28
      Evan Brown committed
    • Optimize Swissmap Match on Arm. · 039d70f6
      Currently we require only a single bit to be set in each abstract bit for iterable bitmasks. However, in most cases, where we have a single match, or no matches in a group, iteration is not needed. We move the masking to the iteration function instead of having it as a requirement for iterable Bitmask construction.
      
      This is 4-8% faster for Find and Insert operations.
      
      This can hurt performance if we need to iterate many times (there are many matches in the same Group), however this is unlikely, even if we assume the table is completely full.
      
      If there are 0 or 1 matches in a group, or the first match is the correct item we are looking for, we save 1 instruction/cycle (most cases)
      If there are 2 matches in a group, and the first is a false positive, this is neutral (< 3%)
      If there are more than 2 matches in a group and the first two are false positives, this can be slower by 1 cycle/instruction per additional iteration (< 0.1%)
      
      No change to x86.
      
      PiperOrigin-RevId: 553831814
      Change-Id: I08620899847eaf0086da989d829a1029ea24173a
      Connal de Souza committed
  4. 03 Aug, 2023 1 commit
  5. 02 Aug, 2023 3 commits
  6. 01 Aug, 2023 8 commits
  7. 31 Jul, 2023 3 commits
  8. 27 Jul, 2023 2 commits
  9. 26 Jul, 2023 3 commits
  10. 25 Jul, 2023 2 commits
  11. 21 Jul, 2023 1 commit
  12. 20 Jul, 2023 1 commit
    • Add a special case for erase(begin(), end()) to reset the control bytes. The… · 9f1dcc70
      Add a special case for erase(begin(), end()) to reset the control bytes. The motivation is to avoid potentially expanding the table unnecessarily later.
      
      Note: I prefer doing this as a special case in erase(iterator, iterator) rather than special casing erase(iterator) for size==1 because IIUC that changes the time complexity of erase(iterator) from O(1) to O(N) and in pathological cases, it could change loops from O(N) to O(N^2).
      PiperOrigin-RevId: 549661855
      Change-Id: I8603324260f51a98809db32f840ff09f25cf2481
      Evan Brown committed
  13. 19 Jul, 2023 1 commit
  14. 18 Jul, 2023 1 commit
  15. 17 Jul, 2023 2 commits
  16. 14 Jul, 2023 1 commit
  17. 13 Jul, 2023 2 commits
  18. 11 Jul, 2023 2 commits