Skip to content

PR.md

Latest commit

 

History

History
56 lines (40 loc) · 5.92 KB

PR.md

File metadata and controls

56 lines (40 loc) · 5.92 KB
Raw

Upstream pull requests

The Swift runtime fixes Vega sends upstream. Vega targets the stock Swift runtime. runtime/patches is an opt-in for servers that ship their own runtime, and each patch there should end up as a PR here.

Update the table whenever a PR changes state: opened, reviewed, CI run, merged, or released in a Swift version.

Open

PR Title Vega patch Opened Status
swiftlang/swift#93174 Concurrency: remove cancellation handlers without taking the status record lock 0002 2026-10-12 ktoso replied positively and started CI (Linux, Windows, macOS); pending on 2026-10-12
swiftlang/swift#93173 Runtime: load 16-byte atomics with one aligned vector load on x86-64 0005 2026-10-12 Reviewed 2026-10-12: mikeash suggested a three-state byte in place of the guarded static, rjmccall suggested skipping CPUID when built for __AVX__. Both done in 8df8d30, pushed 2026-10-13 (lit on Linux x86-64: 378 passed, 166 unsupported, unchanged). Reply posted 2026-10-13. CI failed on macOS arm64 and Windows: Compatibility56's ActiveTaskStatus has no default constructor, which the vector load assumed. Fixed in 8c4fbd6 with a bit cast (checked on Windows, and on the Mac by compiling the Compatibility56 and Concurrency sources for x86-64 and arm64); CI rerun requested 2026-10-15
swiftlang/swift#93179 Concurrency: tell TSan when a continuation is resumed before its await — (found while testing Vega's sync primitives) 2026-10-12 Awaiting review
swiftlang/swift#93180 Concurrency: Task.yield enqueues without copying the current executor 0004 2026-10-12 Awaiting review. Measured −35% per yield on a custom task executor (8 threads); no change on the default Dispatch executor

Merged

None yet.

Not yet proposed

From runtime/patches (see runtime/README.md for details). Each one needs a branch off swiftlang/swift main, a measurement, and a lit run, as #93174 had.

Patch Fixes
0001 getPreferredTaskExecutor takes the status-record lock and does two CASes on every enqueue. The patch makes it one acquire load. Not upstreamable as is: on main, stale AsyncTaskStealer and direct jobs call it (Actor.cpp, fromTaskExecutorPreference) while the task may be running on another thread, so a lock-free walk of the record chain can race the task pushing and popping records. Needs a different design, e.g. a separately published preference.
0003 Each task costs two mallocs, the task itself plus its allocator's first slab. The patch makes it one. Held back: every size tried loses somewhere (branch up-0003, not pushed; measured 2026-10-12 on the box, glibc). With a 512-byte inline slab (the patch as is), spawning onto other threads gets 22–30% faster, but tasks needing 480–1,000 bytes of task memory still take a second slab and get 3–7% slower (Vega chained_spawn). Filling the block to 1 KiB keeps that case level but cuts the spawn gain to 8–20%. With a full 1 KiB slab inline (the size the allocator would have mallocked), cross-thread spawns are 17–30% faster, but the 1,336-byte block leaves glibc's per-thread cache and tasks created and freed on one thread get 9–11% slower (Task.immediate). The result depends on the malloc in use, and it can't be measured on Darwin here. An allocator-independent fix looks like 0007's per-thread task-memory pool, which is a larger proposal.
0006 A continuation resume updates the escalation record under the status lock. The patch makes it one CAS. Not upstreamable as is: on main, flagAsAndEnqueueOnExecutor must decide under the lock whether to enqueue the task or an AsyncTaskStealer (swift_task_getSelfOrStealerForEnqueue), and with priority escalation enabled (Apple platforms) swift_executor_escalate on the generic executor now enqueues a stealer. So the record is no longer a no-op once the task is enqueued. Skipping its update would lose escalations. A version limited to builds without escalation (Linux, Windows) would also have to prove the self-or-stealer choice safe without the lock.
0007 Per-thread task-memory lists, handed between threads in batches of 32, so tasks freed on another thread don't take malloc's arena lock.

When a fix ships

Each fix lets Vega drop something, but only once the fix is in a released Swift toolchain that Vega requires. Merging into main isn't enough.

PR Remove from Vega
#93179 Tests/tsan.supp and its mention in README.md
#93174, #93173, #93180 Patches 0002, 0005 and 0004 from runtime/patches, and their rows in runtime/README.md

Working on a PR

  • Branches live in the WSL checkout ~/swiftrt/upstream/swift and are pushed to the fork grepjava/swift:

    Branch PR
    up-0002 #93174
    up-0005 #93173
    tsan-continuation-await #93179
    up-0004 #93180
    up-0003 — (held back, see above; local only)

  • Build and test with ~/swiftrt/upstream/up-switch.sh <branch> (incremental rebuild), then up-lit.sh <label> (test/Concurrency). The baseline on main is 378 passed and 166 unsupported.

  • TSan tests show as UNSUPPORTED in that build, because it has no tsan_runtime or libdispatch. Check them with the Swift 6.3.3 toolchain against a 6.3.3 runtime built with the fix, as was done for #93179.

  • Measuring: ~/swiftrt/upbench (bench.swift, snap.sh <label>, run-box.sh with LABELS=) measures main against a branch on the box. For changes aimed at custom executors, use Vega's vega-bench against a 6.3.3 runtime with only that patch (see #93180).

  • Status: gh pr view <n> --repo swiftlang/swift --comments and gh pr checks <n> --repo swiftlang/swift.

  • Commits and PR descriptions carry no attribution lines.