[all-commits] [llvm/llvm-project] 96e45a: [flang] Use full result range for clock_gettime im...

vdonaldson via All-commits all-commits at lists.llvm.org
Thu Apr 14 13:01:55 PDT 2022


  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 96e45a8958cbfb5906f5b8b3c2423ad5dd884963
      https://github.com/llvm/llvm-project/commit/96e45a8958cbfb5906f5b8b3c2423ad5dd884963
  Author: V Donaldson <vdonaldson at nvidia.com>
  Date:   2022-04-14 (Thu, 14 Apr 2022)

  Changed paths:
    M flang/docs/Extensions.md
    M flang/runtime/time-intrinsic.cpp

  Log Message:
  -----------
  [flang] Use full result range for clock_gettime implementation of SYSTEM_CLOCK

Update the primary clock_gettime implementation of SYSTEM_CLOCK to use
the full range of values, dependent on the type kind of the requested
result.  Counts/sec and count max for supported kinds become:

 kind          counts/sec             count max

    1                  10                   127
    2                1000                 32767
    4                1000            2147483647
    8          1000000000   9223372036854775807
   16          1000000000   9223372036854775807

The secondary "fallback" implementation is not changed.

Real valued COUNT_RATE arguments are not changed.

The test program below has calls for kinds 1, 2, 4, 8, 16.  Support for
these types varies by compiler.  The code as given can be restricted to
accommodate these variations, with results shown below.

subroutine c
  integer(1) c1, r1, m1
  integer(2) c2, r2, m2
  integer(4) c4, r4, m4
  integer(8) c8, r8, m8
  integer(16) c16, r16, m16

  print*
  print '(a5,3a22)', 'kind', 'counts/sec', 'count max', 'count'
  print*

  call system_clock(c1, r1, m1)
  print '(i5,3i22)', 1, r1, m1, c1

  call system_clock(c2, r2, m2)
  print '(i5,3i22)', 2, r2, m2, c2

  call system_clock(c4, r4, m4)
  print '(i5,3i22)', 4, r4, m4, c4

  call system_clock(c8, r8, m8)
  print '(i5,3i22)', 8, r8, m8, c8

  call system_clock(c16, r16, m16)
  print '(i5,3i22)', 16, r16, m16, c16
end

subroutine k(j)
  j = 0
  do i=1,1000000000
    j = j + i
  enddo
end

program p
  do i=1,1 ! increase loop count to check for (kind=1) wraparound
    call k(j)
    call c
  enddo
end

=== flang output without change (last column counts vary per run) ===

 kind          counts/sec             count max                 count

    1                 -24                   127                    83
    2                1000                   290                   211
    4                1000                   290                   211
    8          1000000000             290448383             211631452
   16          1000000000             290448383             211633853

=== flang output with change (last column counts vary per run) ===

    1                  10                   127                    21
    2                1000                 32767                  2100
    4                1000            2147483647                  2100
    8          1000000000   9223372036854775807            2100183374
   16          1000000000   9223372036854775807            2100185353

Other compilers; kind support varies (last column counts vary per run).
Test and ouput modified to avoid crashes and normalize results.
Some negative values indicate unsupported kinds; others are bugs.

 kind          counts/sec             count max                 count

    1                   0                     0                  -127
    2                   0                     0                -32767
    4                1000            2147483647              69271692
    8          1000000000   9223372036854775807        69271692353290
   16          1000000000   9223372036854775807        69271692354794

=======

    1                  10                   127                     0
    2                1000                 32767                     0
    4             1000000            2147483647                     0
    8            10000000   9223372036854775807                     9

=======

    1                   0                     0                  -127
    2                1000                 32767                  3263
    4               10000            2147483647            1788192630
    8             1000000   9223372036854775807      1649443459263095

=======

    1                 -24                    -1                    36
    2                1000                    -1                -10716
    4                1000            2147483647             176018980
    8                1000   9223372036854775807         1649443460644

=======

    2                 100                 28799                 23080
    4                 100               8639999               4285480
    8                 100               8639999               4285480
   16                 100               8639999               4285480

=======

    1                 -24                    -1                     4
    2                1000                 23551                -26108
    4                1000              86399999              67541508
    8             1000000   9223372036854775807      1649443541508087




More information about the All-commits mailing list