P3008 R2 Atomic floating-point min/max

[wg21bot]

P3008R0 Atomic floating-point min/max (Gonzalo Brito Gadeschi)

[cor3ntin]

@ogiroux , @mattkretz - can SG1 and SG6 please review and tag LEWG once forwarded

[ogiroux]

SG1 poll in Kona:

DIRECTION POLL:
1. The default fetch_min(T, mo=SC)->T, fetch_max(T, mo=SC)->T should exist
2. The default fetch_min, fetch_max should have these semantics:
  * Which value is stored in the atomic is unspecified if an input is a NaN
  * Implementations should treat -0 as less than +0 (use normative encouragement)
  * (Note: must _not_ require std::min/max)
3. Users need a way to specify semantics (up to LEWG how to achieve this):
  * Must-have: fminimum/fmaximum
  * Must-have: fminimum_num/fmaximum_num
  * Can have if LEWG desires it but we do not: std::min/max
SF F N A SA
 5 6 2 0 0
Unanimous consent

[inbal2l]

Needs to be rebased on P0493. (Resolves issues, see #866)

[wg21bot]

P3008R1 Atomic floating-point min/max (Gonzalo Brito Gadeschi, David Sankel)

[mattkretz]

P3008R1: Atomic floating-point min/max

2024-03-19 SG6 Minutes

Chair: Lisa Lippincott

Champion: Gonzalo Brito

Polls

POLL: Accept SG1 proposal as listed in P3008R1 with
"implementation-defined" instead of "unspecified" and forward to LEWG.

SF	F	N	A	SA
2	3	2	2	0

# of Authors: 2
Outcome: Consensus
Comments: Against because of a preference for only two function instead of
three.
Against because of a worry about implementation-defined.

Discussion about walking back from "implementation-defined". Take the above
poll with a grain of salt in light of:

POLL: If one of the inputs is a NaN then it is "unspecified" whether the
result value is either a NaN or the other value.

SF	F	N	A	SA
2	2	3	0	1

# of Authors: 2
Outcome: Consensus
Comments: SA: bad for teachability

Discussion about -0 / +0. Status quo is still normative encouragement (as in
the paper). No taste for a poll to change the status quo.

POLL: With respect to signaling NaNs, SG6 recommends:

• for fetch_fminimum/fminimum_num/fmaximum/fmaximum_num to be specified in
  terms of their corresponding C23 APIs (ignore signaling NaNs in the wording,
  refer to C23), and
• for fetch_min/max, to not explicitly specify signaling NaN behavior (ignore
  signaling NaNs in the wording)

SF	F	N	A	SA
4	3	2	0

# of Authors: 2
Outcome: Consensus

[wg21bot]

P3008R2 Atomic floating-point min/max (Gonzalo Brito Gadeschi, David Sankel)