N3786: Prohibiting "out of thin air" results in C++14

We argued in N3710 that the current attempt prohibit out-of-thin-air results in 29.3p9 [atomics.order] is both insufficient, in that it leaves it largely impossible to reason about programs with memory_order_relaxed, and seriously harmful, in that it arguably disallows all reasonable implementations of memory_order_relaced on architectures like ARM and POWER. This proposal attempts to address the latter "seriously harmful" part for C++14, without addressing the "insufficient" part, which would still need to be addressed in C++17, possibly along the lines of N3710.

This proposal thus eliminates some amount of complicated standardese that we actually want vendors like ARM and IBM to ignore.

There is no consensus in SG1 for a more extensive proposal, such as that in N3710, at this time, i.e. in time for C++14.

Proposed wording for 29.3

This removes the current badly formulated requirement, and instead promotes the existing note discouraging "out-of-thin-air" results to normative encouragment.

The core problem here is that we do not know (after 10 years of failed attempts, originally in a Java context) how to precisely state this requirement without strengthening it so that it impacts code performance. (See N3710 for details.) Thus it remains of little use to anyone trying to formulate a precise correctness argument. It is roughly analogous to our current statements about progress guarantees, which are also phrased as normative encouragement in large part because they would be extremely difficult to make precise while remaining implementable.

Proposed wording

Change 29.3p9-11 [atomics.order] as follows

An atomic store shall only store a value that has been computed from constants and program input values by a finite sequence of program evaluations, such that each evaluation observes the values of variables as computed by the last prior assignment in the sequence. The ordering of evaluations in this sequence shall be such that:

• if an evaluation B observes a value computed by A in a different thread, then B does not happen before A, and
• if an evaluation A is included in the sequence, then every evaluation that assigns to the same variable and happens before A is included.

Implementations should ensure that no "out-of-thin-air" values are computed that circularly depend on their own computation.

[ Note: The second requirement disallows "out-of-thin-air" or "speculative" stores of atomics when relaxed atomics are used. Since unordered operations are involved, evaluations may appear in this sequence out of thread order. For example, with x and y initially zero,

// Thread 1:
r1 = y.load(memory_order_relaxed);
x.store(r1, memory_order_relaxed);
// Thread 2:
r2 = x.load(memory_order_relaxed);
y.store(42, memory_order_relaxed);

is allowed to produce r1 = r2 = 42. The sequence of evaluations justifying this consists of:

y.store(42, memory_order_relaxed);
r1 = y.load(memory_order_relaxed);
x.store(r1, memory_order_relaxed);
r2 = x.load(memory_order_relaxed);

On the other hand, [Note: For example, with x and y initially zero,

// Thread 1:
r1 = y.load(memory_order_relaxed);
x.store(r1, memory_order_relaxed);
// Thread 2:
r2 = x.load(memory_order_relaxed);
y.store(r2, memory_order_relaxed);

may should not produce r1 == r2 == 42, since the store of 42 to y is only possible if the store to x stores 42, which circularly depends on the store to y storing 42. Note that without this restriction, such an execution is possible. there is no sequence of evaluations that results in the computation of 42. In the absence of "relaxed" operations and read-modify-write operations with weaker than memory_order_acq_rel ordering, the second requirement has no impact. -end note ]

[ Note: The requirements do allow The recommendation similarly disallows r1 == r2 == 42 in the following example, with x and y again initially zero:

// Thread 1:
r1 = x.load(memory_order_relaxed);
if (r1 == 42) y.store(r142, memory_order_relaxed);
// Thread 2:
r2 = y.load(memory_order_relaxed);
if (r2 == 42) x.store(42, memory_order_relaxed);

However, implementations should not allow such behavior. -end note ]