[QRMOE]

What

Accept the complete signed 64-bit result range in llroundl. On main, llroundl(9223372036854775807.0L) returns INT64_MIN and raises FE_INVALID, although the input is exactly representable in the guest's x87 extended format and the correct rounded result is INT64_MAX.

The shared rounding helper uses an exclusive upper bound of static_cast<TFloat>(INT64_MAX). That bound becomes 2^63 for float and double but stays INT64_MAX for long double, excluding the largest valid integer. Use the exactly representable 2^63 bound in every specialization. The existing float/double range checks and invalid-input handling retain their behavior.

The valid result agrees with FreeBSD's s_llroundl.c and its s_lround.c, which includes the maximum integer within the long-double input range.

Tested

Windows, MinGW GCC 15.2, Release, BUILD_TESTING=ON; negative control on main 7e181fcc3, final branch rebased onto 88107b7de:

• The new test fails on unchanged main: expected 9223372036854775807, got -9223372036854775808.
• The fixed regression checks both integer limits, adjacent integral and half-integral inputs, overflow on either side, infinities and NaN under all four rounding modes. It checks FE_INVALID and preserves the caller's errno. Existing float/double boundary cases also pass.
• Explicit libs build and ten libc/runtime/loading checks passed after the rebase. Conventions and whitespace checks passed.
• A full Windows build was attempted on the previous base. All remaining components built, but Windows denied recreation of eight unrelated test outputs (guest_sce_net, guest_poll, guest_socket_read_write, guest_tcp, guest_socket_pair, guest_dup, guest_socket_events and guest_fstatfs). Avast's log records automatic sandboxing of the test programs. The local full-build limitation remains; local full-build success is not claimed.
• No title playback or console measurement was performed.

Review validation

• xXPBSXx's Windows report: Reports the relevant Windows test passing in a combined merge; the PR was not tested alone and no negative control was performed in that report.
• nicolasfalesy's Linux report: Linux relevant runtime checks passed in a combined merge, and failed when this PR's source change was reverted while retaining the test.

Ready for maintainer review. Full Linux/Windows CI requires a maintainer's manual dispatch; marking this PR ready does not claim those jobs 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 under Silent stubs in their library's technical debt file
• 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
• Unblocks: none
• AI-assisted: yes

[xXPBSXx]

Windows test of this PR.

Environment: Windows 10 (19045), MinGW-w64 GCC 15.2.0 (WinLibs r7), Release, BUILD_TESTING=ON, -j2, libusb-cmake from a local checkout. Built from main ea2bbefb2 with this PR and six other open PRs (#4958, #4956, #4954, #4953, #4946, #4942, #4927) merged together; the merges had no conflict.

Result: guest_math (1 of 1; the test file is shared with #4946, both merged together) pass.

Not checked: whether the test fails without the change in the library code; the PR was not tested alone, only in the combined merge (the tests of the other PRs in the batch also pass); no title run; Linux not tested.

AI-assisted (Claude Code).

[nicolasfalesy]

Linux test (Arch, GCC 16.2.1, glibc, Release), this PR merged onto main 7e181fc together with #4963, #4953, #4946 and #4942 (no conflicts): guest_math passes. With this PR's Math.cpp change reverted it fails on Linux too, so the check guards the change on both systems.