To:      J3                                             J3/14-158r1
From:    Reinhold Bader
Subject: atomic subroutine semantics and example for Annex
Date: 2014 June 19
References: N1999, N2007

Discussion:
~~~~~~~~~~~

Starting from my comments on section A.3.2.2 in N1999, this paper
proposes a rewording of the text that defines the atomic subroutine
semantics as well as addition of another example to A.3.2.2.

The text in N2007/[34:4-7] does not appear in section 7.2; conversely,
the changes made in 7.2 are not completely applied in section 8.8.
Fixes for both these issues are included.

EDITS to N2007:
~~~~~~~~~~~~~~~

[17:12-15] Replace "The effect ... any other." by

"For any two executions of atomic subroutines in unordered segments by
 different images on the same atomic object, the effect is as if one of
 the executions is performed before the other in a single segment on a
 separate image, without access to the object in either execution
 interleaving with access to the object in the other. Which is executed
 first is indeterminate."

[[Reword in terms of the simplest building block that provides full
 generality, based on a suggestion by John Reid. "processor-dependent"
 appears misleading since the ordering may change between subsequent
 program runs as well as appear to be different on different images.]]

"For an execution of an atomic subroutine in segment P_i, the effect
 on any image Q is as if it is executed on Q within a finite time after
 execution of the first segment Q_j that is unordered against P_i has
 started."

[[Prevent an implementation from infinitely delaying visibility of
  updates on other images while claiming conformity.]]


[17:16+] After "2010.", insert

"If two variables are updated by atomic memory operations in segments
 P_1 and P_2, and the changes to them are observed by atomic accesses
 from a  segment Q which is unordered relative to either P_1 or P_2,
 the changes  need not be observed in segment Q in the same order as
 they are made in  segments P_1 and P_2, even if segments P_1 and P_2
 are ordered."

[[This is a reworded version of the text from [34:4-7].
  Also, I've moved the text to an earlier location since this seems
  topically appropriate; in fact it may be desirable to turn [17:16-19]
  "For invocation ... atomic action." into a separate para.]]

[17:19+] Add
  "NOTE 7.1-
   These properties support the use of atomic subroutines for designing
   customized synchronization mechanisms. The programmer needs to
   account for all possible orderings of sequences of atomic
   subroutine executions that can arise as a consequence of the above
   rules; the orderings can turn out to be different on different
   images even in the same program run."

[[The above is intended to replace NOTE 13.1 of 10-007]]

[34:1-7] Replace para by

{In 13.1 Classes of intrinsic procedures, replace paragraph 3 by
  paragraph 1 [[and 2, see above]] of 7.2 Atomic subroutines,
  with these changes: Delete "of ISO/IEC 1539-1:2010"}

{In 13.1 Classes of intrinsic procedures, replace the text
 of NOTE 13.1 with that of NOTE 7.1- in 7.2 Atomic subroutines.}

[47:26+]
"Example 4:

Assuming the declarations

INTEGER(ATOMIC_INT_KIND) :: x[*]= 0, z = 0

the statements

CALL ATOMIC_ADD(x[1], 1)         ! (A)
IF (THIS_IMAGE() == 2) THEN
  wait : DO
     CALL ATOMIC_REF(z, x[1])    ! (B)
     IF (z == NUM_IMAGES()) EXIT wait
  END DO : wait                  ! (C)
END IF

will execute the "wait" loop on image 2 until all images have completed
statement (A). The updates of x[1] are performed by each image in
the same manner, but arbitrary order. Because the result from the
complete set of updates will eventually become visible by execution
of statement (B) for some loop iteration on image 2, the termination
condition is guaranteed to be eventually fulfilled, provided that no
image failure occurs, no image branches around the above code, and no
other code is executed in an unordered segment that performs updates
to x[1]. Furthermore, if two SYNC MEMORY statements are inserted in
the above code before statement (A) and after statement (C),
respectively, the segment started by the second SYNC MEMORY on image 2
is ordered after the segments on all images that end with the first
SYNC MEMORY."