fix(libkernel): use precise sleep for sub-millisecond host waits

[neuralsin]

What

Addresses the sub-millisecond sleep busy-spin issue documented in #1471:
On Windows, std::this_thread::sleep_for goes through winpthreads' nanosleep, which truncates sub-millisecond durations (tv_nsec / 1'000'000) to 0 ms and returns immediately without sleeping or yielding. As a result, polling loops requesting sub-millisecond pauses spin at 100% CPU.

This PR switches the sub-millisecond host waits in four areas to use PreciseSleepUs() (which uses a Windows high-resolution waitable timer CREATE_WAITABLE_TIMER_HIGH_RESOLUTION on Windows and standard microsecond sleep on POSIX):

• core/libs/prx/libSceAgcDriver/Execution/src/Pm4.cpp: WAIT_REG_MEM polling loop (50 µs)
• core/libs/prx/libkernel/Apr/src/Apr.cpp: _execute counter and address wait loop (50 µs)
• core/libs/prx/libkernel/Aio/src/Aio.cpp: sceKernelAioWaitRequests and sceKernelAioWaitRequest polling loops (10 µs)
• core/libs/prx/libSceFont/src/State.cpp: Font::Backoff spinlock backoff (30 µs)

Tested

• Linux / WSL 2 Ubuntu 24.04 (x86_64), GCC 15.2.0, Ninja 1.13.2
• ctest -R 'guest_aio|guest_ampr|guest_font': 4/4 passed (100% tests passed: guest_ampr_markers, guest_font, guest_font_writing_line, guest_aio)
• python tools/check_conventions.py --base origin/main: 100% passed

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes

[MarcZX10]

I found that the previous head of #1551 duplicated the same four production substitutions as this PR. I have re-scoped #1551 to add and register only a focused PreciseSleepUs CTest; it no longer changes those production files. The focused test passes locally on Linux, and fork CI is running on the updated head.

[MarcZX10]

The updated test-only head of #1551 has now passed its fork Build: 495/495 Linux tests, 486/486 Windows tests, and all three relinker jobs. Its diff contains only PreciseWait.cpp and CTest registration; no production files overlap this PR.

[popcoder565]

Windows numbers for this, since it's only been run on Linux so far. Setup: Windows 10 (19045), MinGW-w64 GCC 15.2.0 (WinLibs r7), with the 0.5 ms timer resolution libkernel sets.

wait	takes
std::this_thread::sleep_for for 10, 50 or 200 µs	0 ms (returns at once)
high-resolution waitable timer (PreciseSleepUs) for 10, 50 or 200 µs	about 0.50 ms each

So on Windows each of these polls goes from a busy spin to a 0.5 ms wait, 10–50 times the requested pause. On Linux they sleep roughly 60 µs.

• AIO: neither loop runs in practice. SubmitCommands performs the reads and sets the final state before it returns. On main, 2000 back-to-back 4 KB sceKernelAioSubmitReadCommands + sceKernelAioWaitRequest pairs take 3.2 ms in total, and every wait returns on its first check. The AIO hunk doesn't change anything measurable.
• Font backoff and APR waits: these do loop, so each contended backoff or unmet wait now costs about 0.5 ms instead of spinning. I haven't run a workload that hits them.

🤖 Addressed by Claude Code

[MarcZX10]

I have integrated the four production call-site changes from #1576 into my #1551, alongside its focused PreciseSleepUs regression test and Windows test setup; the commit credits you as co-author. The combined current head passes the seven focused Linux Release tests and Conventions. #1551 is now the more complete canonical PR for this change; I have asked an upstream collaborator to run its required full Build under #2113. Please close #1576 as superseded once that current-head Build is confirmed, so the same call-site fix is reviewed and merged once.

[MarcZX10]

Update on the overlapping wait changes

I have rebased #1551 onto current main at 8b393b6e (head 46a77cca) and rerun its seven focused Linux tests; all pass. It contains the same four production call-site changes as #1576, credits you as co-author, and also carries the focused PreciseSleepUs regression test. For that reason #1551 is the more complete canonical PR for this behavior.

#1576 currently has no checks reported, while #1551 still needs the repository-required upstream Build/relinker run. Please close #1576 as superseded once that required validation is recorded on #1551, so reviewers only have to land one copy of the call-site changes.

[MarcZX10]

Consolidation status

I refreshed Marc's existing PR #1551 onto upstream main 0df342c6 (head e777d26b) and reran focused Linux validation. It still carries the four precise-wait call-site changes from this PR, credits Shaan as co-author, and also includes the dedicated helper regression test.

PR	Scope	Current status
#1551	Four production call sites plus helper regression test	Canonical carrier; Linux focused tests pass on the current head; upstream Build/relinker matrix still needs a collaborator dispatch
#1576	Same four production call sites	Duplicate behavior already integrated into #1551 with attribution retained

Could you close #1576 so the same production change is not reviewed twice? Please keep this PR open if you have additional behavior or evidence that is not covered by #1551; we can review that separately.

[adriadam10]

Closing as a duplicate: #1551 now carries the same four call-site changes (Pm4.cpp, Apr.cpp, Aio.cpp, State.cpp), credits you as co-author, and adds a regression test for PreciseSleepUs, so the change should be reviewed and merged once, there.

Two points from @popcoder565's Windows measurements that are worth addressing in #1551: each wait becomes ~0.5 ms instead of 10-50 µs, and the AIO loops don't run in practice, so that hunk changes nothing measurable.

Thanks for the patch; feel free to reopen if #1551 doesn't move forward.

[MarcZX10]

Coordination requested after the #1551 review

The reviewer on #1551 confirmed that both open PRs currently contain the same four production call-site changes and asked us to choose one carrier. I have preserved your call-site work in #1551 with you credited as co-author; #1551 also contains the focused helper regression test. I prefer #1551 as the single carrier.

Could you confirm whether you agree to close #1576 as superseded once its changes are retained in #1551? If you prefer #1576 to carry the production changes, please say so and I will remove those call-site edits from #1551. I will not ask for both PRs to merge.