[imellag]

What

_LCosh and _LSinh reduce the argument with std::fma(-exponent, Ln2High, magnitude), where exponent is an int and the other two are long double. That picks the mixed-type std::fma template. In libc++ (macOS, and any clang build that uses libc++) that template computes the promoted type, long double, but then calls __builtin_fma, the double version (__math/fma.h, line 48 of the macOS SDK header; the other mixed-type math templates call __math:: and are fine). So the reduced argument is rounded to 53 bits before expl, and the last ~10 bits of the 64-bit result are wrong. libstdc++ promotes correctly, so Linux and the MinGW build are not affected.

The fix converts the negated exponent to long double once and passes that to both fma calls, so every compiler picks the long double overload. In the built object the two functions now call fmal four times and fma no longer.

Tested

macOS 27.2 (Apple M4 Pro, x86-64 under Rosetta 2), Apple clang 21, Release, guest_math_tests, on main plus the support commits needed to build on macOS.

• _LCosh(11357, 0x1p-16000) returns 0xc.e2773666cc8cb2dp+380; the test expects 0xc.e2773666cc8cc68p+380. The difference, 0x13b in the last place, matches the reduced argument rounded to 53 bits (0x136 predicted).
• On this machine guest_math first stops at fmodf (Apple's libm does not raise invalid for fmod(1, 0)), which fix(libc): set EDOM in fmod and fmodf where titles' libc.prx does #10345 fixes. With fix(libc): set EDOM in fmod and fmodf where titles' libc.prx does #10345 applied: without this change GuestMath.cpp stops at the _LCosh(11357.0L, 0x1p-16000L) check 3 out of 3 runs; with this change the whole guest_math passes 3 out of 3.
• No test change: the existing CheckLongDoubleHyperbolic checks already catch it on libc++.

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 under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none
• Unblocks: none
• GPUs tested: n/a
• AI-assisted: yes

[MarcZX10]

Reviewed — looks good

At head a7e91c26fdca4bbac6cb07817983fc1e1ce7594d, both long-double hyperbolic paths now pass long-double operands to std::fma, addressing the reported narrowing behavior.

What I ran

Reviewed head a7e91c26fdca4bbac6cb07817983fc1e1ce7594d against base 73f4f7e06e4f210a8c6570e335fd654dc0a86134; the merge preview was clean. Local environment: Linux x86_64, Python 3.11.2, GCC 12.2.0. I ran the no-build precheck and inspected both changed call sites and the existing long-double assertions.

python3 tools/pr_precheck.py --repo boykopovar/AnyPS5 --pr 13952
gh pr checks 13952 --repo boykopovar/AnyPS5

Results

Check	Result
Diff and conventions	Passed
Current-base merge preview	Clean
Hosted Windows job	Passed
Hosted Ubuntu Debug and Release register-context jobs	Passed
Hosted Linux full build	Skipped by workflow
Hosted macOS jobs	Pending at review time
Author-reported target test	On macOS 27.2 / Apple clang 21 / M4 Pro x86-64 under Rosetta, guest_math reportedly failed at the existing _LCosh assertion 3/3 times without this change and passed 3/3 with it, with #10345 applied. I did not run this test independently.
Additional local precheck heuristics	Unavailable in the isolated PR snapshot; no conclusion drawn from those checks

Unedited output

All conventions checks passed.
windows	pass	3m22s
register-context (ubuntu-latest, Release)	pass	25s
register-context (ubuntu-latest, Debug)	pass	33s
linux	skipping	0

Sources

The changed long-double reduction paths are in Math.cpp at this head. The existing long-double comparison helper and _LCosh/_LSinh edge assertions are in GuestMath.cpp at this head.

Not verified

I did not run a local build or test. The target-specific macOS result is author-reported; the macOS CI jobs were still pending, and the hosted Linux full build was skipped.

[github-actions]

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 6295	core/libs/prx/libc/src/Math.cpp
PR 10113	core/libs/prx/libc/src/Math.cpp
PR 10226	core/libs/prx/libc/src/Math.cpp
PR 10245	core/libs/prx/libc/src/Math.cpp
PR 10250	core/libs/prx/libc/src/Math.cpp
PR 10253	core/libs/prx/libc/src/Math.cpp
PR 10265	core/libs/prx/libc/src/Math.cpp
PR 10274	core/libs/prx/libc/src/Math.cpp
PR 10325	core/libs/prx/libc/src/Math.cpp
PR 10328	core/libs/prx/libc/src/Math.cpp
PR 10345	core/libs/prx/libc/src/Math.cpp
PR 10349	core/libs/prx/libc/src/Math.cpp
PR 10506	core/libs/prx/libc/src/Math.cpp
PR 10510	core/libs/prx/libc/src/Math.cpp
PR 10512	core/libs/prx/libc/src/Math.cpp
PR 10546	core/libs/prx/libc/src/Math.cpp
PR 10565	core/libs/prx/libc/src/Math.cpp
PR 11260	core/libs/prx/libc/src/Math.cpp
PR 11262	core/libs/prx/libc/src/Math.cpp
PR 11266	core/libs/prx/libc/src/Math.cpp
PR 11632	core/libs/prx/libc/src/Math.cpp
PR 11634	core/libs/prx/libc/src/Math.cpp
PR 12122	core/libs/prx/libc/src/Math.cpp

[imellag]

The two red write-watch jobs (macos-15 and macos-15-intel) are unrelated to this change. They stop at CMake configure with the duplicate write_watch_execute-fault / write_watch_read-fault registration that was on main when they ran; #13220 has since removed it. This PR only changes two std::fma calls in Math.cpp.

AI-assisted: yes