P0876 R18 fiber_context - fibers without scheduler

[jensmaurer]

P0876R5 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[jensmaurer]

SG1 in Kona

[wg21bot]

P0876R6 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[ogiroux]

When a non-empty fiber_context is destroyed...

...we like the previous version of the paper (R5+fixes?). SF F N A SA 0 4 9 7 1

...we like the paper as proposed (R6). SF F N A SA 1 0 3 10 9

...we like the paper changed to UB (so the user has to do it all). SF F N A SA 1 12 6 3 1

...we like the paper changed to invoke a function (in that fiber). SF F N A SA 6 9 9 0 0 - Increases consensus

The cancellation function should have a default in the TS SF F N A SA 0 2 4 17 2

[ogiroux]

After the following changes are applied...

• change ~fiber_context and operator=() to have "requires: is empty" instead of cancelling
• add cancel() and cancel_from_any_thread() that calls the cancellation function if not empty
• replace thread termination when returning an empty fiber_context, with UB ...the SG1 design of fiber_context is ready for a TS vehicle SF F N A SA 3 7 0 0 0

We need more SG1 wording review before merging into a TS SF F N A SA

• Unanimous consent

[wg21bot]

P0876R8 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[wg21bot]

P0876R9 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[theres-waldo]

Based on discussion with JF, this should get language review, and it should start with EWG-I.

[theres-waldo]

EWG-I in Belfast

P0876R9 is ready to bring to EWG as it stands. SF / F / N / A / SA 2 / 8 / 2 / 1 / 0

[theres-waldo]

(EWG-I did suggest adding a section to the paper to discuss some of the issues that came up, such as TLS and and saving the floating-point state. However, given the poll, I forwarded the paper to EWG without blocking on that revision.)

[jfbastien]

This was seen by EWG in Belfast, will see again.

[wg21bot]

P0876R10 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[jfbastien]

EWG Prague, Wednesday afternoon:

Needs to be updated to list the questions which the TS is going to answer.

• We’d like a timeline of future work, to understand how this interacts with other language facilities, and what library facilities would be useful to add.
• We want to understand how blocking primitives such as static initializers would need to change for fiber.
• Does this require a language change that inhibit optimizations, e.g. around function calls or TLS.
• Do we need some way to control how the stack is allocated?
• The problem of cancellation.
• Should fibers be allowed to resume on different a thread than the one it was suspended on?
• Is this onerous for implementers to provide?
• What are the different implementation choices on other operating systems / architectures, what tradeoffs are made?
• Why do you need a separate API if you need a separate API?

Please gather whatever other questions need to be answered for a TS, both from Library and Language perspectives.

We’d like to see this again for consideration into a TS.	SF	F	N	A	SA
7	11	6	1	0

[wg21bot]

P0876R11 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[jensmaurer]

This has a library-looking interface; adding LEWG.

[brycelelbach]

Removing Library Evolution. There are still open directional questions from Evolution. Those need to be settled first.

[erichkeane]

Forward P0876R11 to LEWG for consideration for C++26.

SF	F	N	A	SA
2	3	3	0	0

Result: No strong consensus, but no objection. Forward to LEWG, with no recommendations for/against ship vehicle.

[RobertLeahy]

2023-02-09 15:30 to 17:30 Issaquah Library Evolution Meeting

P0876R12: fiber_context

2023-02-09 15:30 to 17:30 UTC-8 Issaquah Library Evolution Minutes

Champion: Nathaniel Goodspeed (IP)

Chair: Robert Leahy (IP) & Bryce Adelstein Lelbach (IP)

Minute Taker: Steve Downey (IP)

__POLL: We are willing to proceed with fiber_context even if it cannot be resumed on multiple threads.__

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
6	10	2	0	0

Attendance: 22 (in person) + 7 (remote)

# of Authors: 1

Author Position: SF

Outcome: Strong consensus in favor

POLL: fiber_context should be default constructible.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
4	12	1	1	1

Attendance: 22 (in person) + 7 (remote)

# of Authors: 1

Author Position: SF

Outcome: Consensus in favor

POLL: fiber_context should have explicit operator bool.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
2	8	3	6	1

Attendance: (in person) + (remote)

# of Authors: 1

Author Position: SF

Outcome: No consensus.

Robert Leahy to support Nathaniel Goodspeed on wording et cetera.

Next Steps

Return to Library Evolution with wording updated for additional precision and clarity on:

• Exception safety and guarantees (particularly that resumption always ensures fiber_context is empty even if an exception is thrown)
• Decay-copy of entry
• Lifetime of entry

[wg21bot]

P0876R12 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[wg21bot]

P0876R13 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[nliber]

2023-06-12 Varna Library Evolution Meeting polls

P0876R13 fiber_context - fibers without scheduler

2023-06-12 Library Evolution Minutes

Chair: Nevin ":-)" Liber, Fabio Fracassi

Champion: Nat Goodspeed

Minute Taker: Guy Davidson

Summary

POLL: From P0876R13 we want to change can_resume() to a different name before forwarding to LWG.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
0	5	11	7	3

Attendance: 22 + 7

# of Authors: 1

Author Position: WA

Outcome: No consensus for change

POLL: Move the discussion of uncaught_exceptions() and current_exception() from P0876R13 to a new paper and specify in P0876 that it is unspecified to call them when more than one fiber is running in the current thread.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
5	13	6	1	0

Attendance: 25 + 5

# of Authors: 1

Author Position: WF

Outcome: Consensus

WA: going with this approach undermines the design point of fibers

POLL: Modify P0876R13 with the following changes and then forward to LWG for C++26 (to be confirmed by electronic polling): • It is unspecified to call uncaught_exceptions() and current_exception() when more than one fiber is running in the current thread. • Make the three parameter constructor of fiber_context implicit. • Add to the fiber_context assignment operator the following effect: If empty() returns false, invoke terminate ([except.terminate]). • Strike the precondition from the fiber_context assignment operator. • Allow throwing system_error as well as bad_alloc. • Delete the remark under can_resume() and leave the note.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
4	13	1	1	0

Attendance: 29 + 3

# of Authors: 1

Author Position: SF

Outcome: Consensus

WA: If the first bullet point says it documents existing practice, then their vote changes to WF

Next Steps

The authors will apply the changes requested on the first and second polls, and then P0876R14 of the paper (including the changes) will be voted on in the next electronic poll.

[wg21bot]

P0876R14 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[JeffGarland]

This paper has core wording in it -- the author is going to need to get on the core agenda.

[jensmaurer]

CWG 2023-11-09: Reviewed (full minutes); Needs pervasive updates to properly disentangle "thread of execution" (is concurrent and has unique instance of thread-local storage) from "fiber" (has flow of control) throughout the standard.

[inbal2l]

P0876R14 was added to LEWG 2023 December electronic poll paper (P30532R0) (Please note that any design changes should be forwarded via LEWG, preferably, in a separate paper)

[jfbastien]

I'm told by Hubert that CWG asks for EWG input on the following:

As for the language evolution side, questions include whether a throw; lexically in the "body" of a handler should be allowed to rethrow an exception other than the one that became the currently handled exception when the handler became active. It seems LEWG was of the opinion that std::rethrow_exception(std::current_exception()) in a similar context should be allowed to rethrow such an other exception. The core language design and the library design require synchronization here. I am not aware of EWG having discussed this question. The last reference I found of exceptions in EWG discussion of fibers was a very long time ago. The more general question is whether there should be a per-fiber exception state.

See notes: https://wiki.edg.com/bin/view/Wg21kona2023/CoreWorkingGroup

[jensmaurer]

I agree this is an open question for EWG, although we might hit other corners that need EWG input once the paper has been properly drafted according to the feedback we gave. In other words, feel free to get an EWG response to this particular question at your earliest convenience, but don't be disappointed if the paper comes back to EWG anyway at some later point.

[inbal2l]

As P0876R14 is currently been polled in Dec 2023 LEWG electronic poll (to be forwarded to LWG), CCing my response from the LEWG mailing list (Dec 24th: https://lists.isocpp.org/lib-ext/2023/12/26286.php). To summarize from LEWG's perspective: I believe we can move forward with the current LEWG-requested design, and have EWG/Core give their input when possible. If EWG/Core gives contradicting input/brings up issues, LWG can send it back to us.

[inbal2l]

P0876R14 was forwarded to LWG on LEWG 2023 December electronic poll (P3054R0)

[wg21bot]

P0876R15 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[jensmaurer]

CWG 2024-03-18: Reviewed; needs updates.

[jensmaurer]

CWG 2024-03-21: Reviewed partially.

CWG asks EWG to reconsider the decision to permit implementations with per-thread (as opposed to per-fiber) exception state. Examples of surprise results have been added to the paper. Possible outcomes are to specify a per-fiber exception state, to require implementations to satisfy the existing rules around exceptions without actually requiring separate exception states, or to document the situations with bad outcomes as undefined behavior.

[wg21bot]

P0876R16 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[jfbastien]

Seen in EWG in St Louis:

Poll: P0876R16 — fiber_context - fibers without scheduler, change the paper to track exceptions on a per-fiber basis rather than leaving it implementation-defined.

| SF | F | N | A | SA | | 6 | 8 | 3 | 0 | 0 |

Consensus.

Poll: P0876R16 — fiber_context - fibers without scheduler, forward the paper, with modification to track exceptions on a per-fiber basis, to CWG for inclusion in C++26.

| SF | F | N | A | SA | | 3 | 2 | 5 | 6 | 3 |

No consensus, would like to see implementation experience before forwarding.

[wg21bot]

P0876R17 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[jensmaurer]

Unassigning CWG, since EWG wants to see implementation experience.

[wg21bot]

P0876R18 fiber_context - fibers without scheduler (Oliver Kowalke, Nat Goodspeed)

[JeffGarland]

LWG spent several sessions on this paper in St Louis (not to mention Tokyo and others). We are close to done with wording review, but now that the paper is going back to other groups we won't take our last pass on this until it's cleared those groups.

[jfbastien]

Seen in EWG on Thursday, Gor had concerns on Microsoft platforms because the exceptions live on the stack, Gor committed to review concerns internally to Microsoft and come back with feedback before the next meeting so that authors can know how to proceed.