[STUPIDCONTEM]

Evidence review (2026-10-09)

Game/title evidence has not been established. Validation below is host regression/build evidence only; it does not establish title compatibility. Older checked items and test results describe their stated revisions, not current merge readiness.

Game/title ID: none verified. No PS5 title execution is claimed.

What

_FSinh and _FCosh evaluated the unscaled float hyperbolic function before multiplication, so representable scaled results overflowed. For example, _FSinh(100.0f, 0x1p-40f) returned infinity instead of a finite result near 1.22e31.

Evaluate finite intermediates in double and narrow only after scaling. Handle zero, NaN and infinite arguments explicitly according to the scaled CRT routines, including signed-zero behavior, and report ERANGE for actual final float overflow. Tiny products that round to zero preserve errno.

Semantics are sourced from Microsoft's primary CRT implementations xsinh.cpp, _FSinh and xcosh.cpp, _FCosh, whose stated scale domain is |y| <= 1. No implementation code or polynomial tables were copied. PS5 hardware behavior has not been measured.

Tested

Windows 10 Pro 19045, WinLibs GCC 15.2.0 UCRT/posix/SEH r7, Ninja Release, BUILD_TESTING=ON, SPIRV-Tools disabled. Independent branch from f145139676a48306681d3a4d5fabf83938759daf. Open #1844 adds ordinary libm test coverage without changing these implementations; this regression has a separate test file and target.

• New guest_scaled_hyperbolic built and run against unchanged upstream: failed finite scaled-result, errno, and special-value assertions.
• Six finite reference cases (x=0.25, 1, 90, 100, 150, 192), all argument/scale signs, including subnormal scales. Fixed expected float values independently verified with 100-digit decimal arithmetic.
• Signed zero, tiny products, NaN/infinite arguments, zero scale with huge finite arguments, and actual overflow at x=100/y=1 and x=193/minimum-subnormal scale.
• Parent independently reviewed the implementation, corrected tiny-product errno behavior, and ran the new regression in the full project runtime.
• Full default and explicit libs builds passed.
• Full CTest suite: 489/489 passed, 26.87 s.
• Committed-head conventions, whitespace and test-registration checks passed.

Linux and PS5 hardware/titles were not run. Tests use the default floating-point rounding mode; alternate rounding modes were not validated.

Checklist

• Based on current main; no other open PR implements the same change (historical base; current readiness not asserted)
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw; no silent stubs added
• No notes, investigation files or images added to the repository
• New third-party code is a submodule built from source (none added)
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes — OpenAI Codex agents assisted with implementation; the parent agent independently reviewed and tested the final change.

[MarcZX10]

Reviewed the full diff and the referenced Microsoft CRT behavior. The implementation keeps the scale-before-narrowing behavior, handles the documented special cases, and adds direct regression coverage for finite results, signed zero, errno, and overflow. I built the exact head on Linux with GCC 12; guest_scaled_hyperbolic passed (1/1). Conventions and diff checks pass. Upstream Build, Conventions, and Progress workflows are still action_required, so they have not run on this PR.