This issue has been discussed by the authors at every recent Standards meetings, yet a full solution has been elusive despite helpful proposals. We believe that this proposal can fix this oft-encountered problem once and for all.
[P0528r0] details extensive background on this problem (not repeated here),
and proposed standardizing a trait, has_padding_bits
, and using it on compare_and_exchange_*
. [P0528r1] applied EWG guidance and simply added
wording directing implementations to ensure that the desired behavior occur. At
SG1’s request this paper follows EWG’s guidance but uses different wording.
1. Edit History
1.1. r2 → r3
In Rapperswil, CWG suggested various wording updates to the paper.
1.2. r1 → r2
In Jacksonville, SG1 supported the paper but suggested an alternate way to approach the wording than the one EWG proposed in Albuquerque: don’t talk about contents of the memory, but rather discuss the value representation to describe compare-and-exchange. This paper follows SG1’s guidance and offers different wording, with the intent that the semantics be equivalent. EWG reviewed the updated wording an voted to support it and forward to Core.
1.3. r0 → r1
In Albuquerque, EWG voted to make the padding bits of atomic
and the incoming
value of T
have a consistent value for the purposes of read/modify/write
atomic operations?
Purposefully not addressed in this paper:
-
union
with padding bits -
Types with trap representations
2. Proposed Wording
In Operations on atomic
types [atomics.types.operations], edit ❡17 and
onwards as follows:
bool compare_exchange_weak(T& expected, T desired, memory_order success, memory_order failure) volatile noexcept; bool compare_exchange_weak(T& expected, T desired, memory_order success, memory_order failure) noexcept; bool compare_exchange_strong(T& expected, T desired, memory_order success, memory_order failure) volatile noexcept; bool compare_exchange_strong(T& expected, T desired, memory_order success, memory_order failure) noexcept; bool compare_exchange_weak(T& expected, T desired, memory_order order = memory_order::seq_cst) volatile noexcept; bool compare_exchange_weak(T& expected, T desired, memory_order order = memory_order::seq_cst) noexcept; bool compare_exchange_strong(T& expected, T desired, memory_order order = memory_order::seq_cst) volatile noexcept; bool compare_exchange_strong(T& expected, T desired, memory_order order = memory_order::seq_cst) noexcept;
❡17:
Requires: The
failure
argument shall not bememory_order::release
normemory_order::acq_rel
.
❡18:
Effects: Retrieves the value in
expected
. It then atomically compares thecontents of the memoryvalue representation of the value pointed to bythis
for equality with that previously retrieved fromexpected
, and if true, replaces thecontents of the memoryvalue pointed to bythis
with that indesired
. If and only if the comparison is true, memory is affected according to the value ofsuccess
, and if the comparison is false, memory is affected according to the value offailure
. When only onememory_order
argument is supplied, the value ofsuccess
isorder
, and the value offailure
isorder
except that a value ofmemory_order::acq_rel
shall be replaced by the valuememory_order::acquire
and a value ofmemory_order::release
shall be replaced by the valuememory_order::relaxed
. If and only if the comparison is false then, after the atomic operation, thecontents of the memoryvalue inexpected
areis replaced by the valueread from the memorypointed to bythis
during the atomic comparison. If the operation returnstrue
, these operations are atomic read-modify-write operations on the memory pointed to bythis
. Otherwise, these operations are atomic load operations on that memory.
❡19:
Returns: The result of the comparison.
❡20:
[Note:
For example, the effect of
compare_exchange_strong
on objects without padding bits isif (memcmp(this, &expected, sizeof(*this)) == 0) memcpy(this, &desired, sizeof(*this)); else memcpy(expected, this, sizeof(*this));—end note]
[Example:
The expected use of the compare-and-exchange operations is as follows. The compare-and-exchange operations will update
expected
when another iteration of the loop is needed.expected = current.load(); do { desired = function(expected); } while (!current.compare_exchange_weak(expected, desired));—end example]
[Example:
Because the expected value is updated only on failure, code releasing the memory containing the
expected
value on success will work. E.g. list head insertion will act atomically and would not introduce a data race in the following code:do { p->next = head; // make new list node point to the current head } while (!head.compare_exchange_weak(p->next, p)); // try to insert—end example]
❡21:
Implementations should ensure that weak compare-and-exchange operations do not consistently return
false
unless either the atomic object has value different fromexpected
or there are concurrent modifications to the atomic object.
❡22:
Remarks: A weak compare-and-exchange operation may fail spuriously. That is, even when the contents of memory referred to by
expected
andthis
are equal, it may returnfalse
and store back toexpected
the same memory contents that were originally there.[Note:
This spurious failure enables implementation of compare-and-exchange on a broader class of machines, e.g., load-locked store-conditional machines. A consequence of spurious failure is that nearly all uses of weak compare-and-exchange will be in a loop. When a compare-and-exchange is in a loop, the weak version will yield better performance on some platforms. When a weak compare-and-exchange would require a loop and a strong one would not, the strong one is preferable.
—end note]
❡23:
[Note:
Under cases where theThememcpy
andmemcmp
semantics of the compare-and-exchange operations apply, the outcome might bemay result infailed comparisons for values that compare equal withoperator==
if the underlying type haspadding bits,trap bits,or alternate representations of the same value. Notably, on implementations conforming to ISO/IEC/IEEE 60559, floating-point-0.0
and+0.0
will not compare equal withmemcmp
but will compare equal withoperator==
, and NaNs with the same payload will compare equal withmemcmp
but will not compare equal withoperator==
.—end note]
[Note:
Because compare-and-exchange acts on an object’s value representation, padding bits that never participate in the object’s value representation are ignored.
As a consequence, the following code is guaranteed to avoid spurious failure:
struct padded { char clank = 0x42; // Padding here. unsigned biff = 0xC0DEFEFE; }; atomic<padded> pad = ATOMIC_VAR_INIT({}); bool zap() { padded expected, desired { 0, 0 }; return pad.compare_exchange_strong(expected, desired); }—end note]
[Note:
For a union with bits that participate in the value representation of some members but not others, compare-and-exchange might always fail. This is because such padding bits have an indeteminate value when they do not participate in the value representation of the active member.
As a consequence, the following code is not guaranteed to ever succeed:
union pony { double celestia = 0.; short luna; // padded }; atomic<pony> princesses = ATOMIC_VAR_INIT({}); bool party(pony desired) { pony expected; return princesses.compare_exchange_strong(expected, desired); }—end note]