---
- branch: MAIN
  date: Tue Jul 31 06:34:11 UTC 2018
  files:
  - new: '1.4'
    old: '1.3'
    path: pkgsrc/math/djbsort/Makefile
    pathrev: pkgsrc/math/djbsort/Makefile@1.4
    type: modified
  - new: '1.3'
    old: '1.2'
    path: pkgsrc/math/djbsort/distinfo
    pathrev: pkgsrc/math/djbsort/distinfo@1.3
    type: modified
  - new: '1.2'
    old: '1.1'
    path: pkgsrc/math/djbsort/pseudo-PLIST
    pathrev: pkgsrc/math/djbsort/pseudo-PLIST@1.2
    type: modified
  id: 20180731T063411Z.10097fe01ff685c36c50295f9f1f0bbab1418bb9
  log: "Update to 20180729. From the changelog:\n\nAlgorithm\n\nRewrite of the core
    int32/avx2 implementation for (1) higher speed and\n(2) reduced memory consumption.
    Stack allocation is now at most a few\nkilobytes, even for gigantic arrays.\n\nInternally,
    the sorting algorithm is now mostly bitonic to simplify\nindexing, although odd-even
    speedups are still applied when\nconvenient. Lanes are complemented to take the
    down-up decision out of\nthe inner loops.\n\nAs in previous djbsort versions,
    data is sorted first in vector lanes\nand then transposed for final merges, reducing
    the overall number of\nvector permutations. Unlike previous versions, transposition
    is done\nin-place. The transposition in this version is bit-reversal on the outer\n6
    bits (bottom 3 bits and the top 3 bits), but leaves intermediate bits\nalone.
    Non-power-of-2 array sizes are handled by an extra, more\ntraditional, merge step.\n\nSizes
    2, 3, 4, 5, 6, 7, 8, 16, 32 are now special-cased. Non-power-of-2\nsizes below
    256 are padded to the next power of 2.\n\nPortable implementations: The out-of-place
    int32/portable1 and\nint32/portable2 implementations are now gone; the in-place\nint32/portable3
    and int32/portable4 implementations remain.\n\nC API\n\nfloat32_sort is now supported.
    The arithmetic in the reduction from\nfloat32 to int32 is int32 31-bit right shift,
    uint32 1-bit right shift,\nxor; this is slightly more efficient than the reduction
    from float32 to\nuint32 from 2001 Herf.\n\nCompiling\n\nTests now have more variation
    (without much slowdown): the uint32 test\ncases now deviate from int32 in more
    than the sign; float32 uses\nfloating-point numbers that aren't integers; int32
    does more loops for\nsmall cases, and some larger cases.\n\nInternals\n\nAPI for
    2-input sorting is now MINMAX macro operating on two\ninputs in place.\n\nBetter
    inline assembly from Jason Donenfeld for 2-input sorting: more\nflexibility in
    compiler's register allocation.\n\nThe package version number is now automatically
    copied to version.c as\nthe implementation version number for implementations
    that don't provide\nversion.c.\n\nVerification\n\nminmax now supports more peephole
    optimizations for complemented bitonic\nsorting and for padding: xor(s,xor(s,t))
    â\x87\x92 t; xor(-1,s) â\x87\x92 invert(s);\nReverse(Reverse(s)) â\x87\x92 s;
    signedmin(invert(s),invert(t)) â\x87\x92\ninvert(signedmax(s,t)); signedmax(invert(s),invert(t))
    â\x87\x92\ninvert(signedmin(s,t)); invert(s)[high:low] â\x87\x92 invert(s[high:low]);\ns[bits-1:0]
    â\x87\x92 s; s[high:low][high2:low2] â\x87\x92 s[high2+low:low2+low];\nConcat(...)[high:low]
    â\x87\x92 ...[high-pos:low-pos] when possible;\nReverse(s)[high:low] â\x87\x92
    Reverse(s[...]) when possible; eliminate\nsignedmin/signedmax when one input is
    the minimum or maximum constant.\n\nverifymany now includes the implementation
    version number on\nverified lines.\n"
  module: pkgsrc
  subject: 'CVS commit: pkgsrc/math/djbsort'
  unixtime: '1533018851'
  user: schmonz