feat: macOS host support (Apple Silicon via Rosetta 2, Vulkan on MoltenVK) by iamthehimansh · Pull Request #8658 · boykopovar/AnyPS5

iamthehimansh · 2026-10-20T20:13:04Z

What

macOS host support for AnyPS5 on Apple Silicon: the emulator runs as x86_64 under Rosetta 2 and renders through Vulkan on MoltenVK. The relinker and the web converter build natively.

Every platform-specific change is behind __APPLE__ / elseif(APPLE) or sits in specifics/macos/. Before every commit I compared the Linux and Windows preprocessed views of each changed file against main; they are unchanged.

Graphics driver address

macOS reserves the AGC driver's fixed 2 MB window at 0xfe0000000 for itself; mmap there fails with EPERM even in an empty process, natively and under Rosetta.
On macOS the driver memory moves to 0xee0000000, and the converted shaders are retargeted to that address.

Kernel, libc and networking

Directory listing and directory-relative paths.
Disk-space queries, rename and same-file checks.
Memory locking and unlocking, and write watching.
Per-thread CPU time from the Mach thread counter.
nanosleep errno, precise waits, and kqueue file and socket events.
Socket options, SO_ACCEPTCONN, the UDP pending-bytes count and the stream-send error after a peer close.
EPROTOTYPE → NET_EPROTONOSUPPORT in sceNetSocket, which is what FreeBSD reports.
IPv4 parsing that rejects leading zeros.
acosf precision under Rosetta.
The guest TLS entry points, and guest unwinding over Mach-O __eh_frame.

Graphics and video output

The macOS Vulkan loader name and portability enumeration.
Integer colour targets on Metal, which need uint/int fragment outputs.
A shutdown-order fix: the registration-preparation lifecycle mutex outlives static destruction.

Relinker

1 KB of header room in the Mach-O image, so code signing does not overlap the entry stub.

Build

macOS-only fixes for headers and linker flags that only exist on Linux, such as link.h, if_link.h, AF_PACKET and --exclude-libs.
A few tests use macOS names where the macOS SDK defines a macro with the same name (SIGUSR1).

Tested

Machine: macOS 26 (Darwin 25.6, arm64) on an Apple M2 Pro. The emulator ran as x86_64 under Rosetta 2, with the Vulkan SDK 1.4.363.0 and MoltenVK 1.4.2. To test the GPU path, source setup-env.sh from the SDK and point VK_DRIVER_FILES at MoltenVK_icd.json. The SDK also installs the arm64-only KosmicKrisp ICD, which an x86_64 process cannot load; with VK_DRIVER_FILES set, it no longer logs that.

Builds and conventions:

cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=ON with the full build: 0 failed objects.
The relinker-only build: 230/230 tests pass.
The web converter build: tools/test_web.py passes.
python3 tools/check_conventions.py --base origin/main: all checks pass.

Full ctest on this machine (M2 Pro, macOS 26.6, at 51a06b126): 1,804 tests; 90 are skipped by the tests themselves, and 19 fail. Main does not build on macOS, so before this branch none of these ran. Full runs on an M1 (@Mauro42K) and an M4 (@prosperkartik) are in the thread.

16 of the 19 failures are graphics features that Apple GPUs or MoltenVK do not provide, or graphics paths still open:

clip and cull distances (4)
denormal and tininess handling (2)
sampler min/max reduction, multisample storage images, robust reads past a pooled buffer's end
pixel interlock, per-sample masks and sample iteration
agc_driver_sample_order, which passes with feat(recompiler): read I/J from barycentric built-ins on MoltenVK #9193
image-table descriptor reuse: MoltenVK's image limits keep the table in the shared pools, which are not kept
tessellation of a merged LS+HS hull program, and indexed indirect rect-list draws (both still open)

The other 3:

guest_floating_point: Rosetta ignores the MXCSR exception mask bits.
guest_math_exact: Rosetta canonicalises x87 pseudo-denormals.
agc_write_tracking: on macOS a host import opens its whole range for writing, so an untouched block in the import window is reported as possibly written by the CPU.

On the M4, agc_driver_ds_permute also fails. SPIRV-Cross forwards the recompiler's unconditional OpGroupNonUniformShuffle into a short-circuit &&/?: operand, which makes the shuffle conditional in MSL; @prosperkartik reduced and filed it as KhronosGroup/SPIRV-Cross#2722.

On the M1 (@Mauro42K, macOS 27.0.1), agc_driver_pass_input_hazard and agc_driver_uint8x4_color_target also fail; both pass on the M2 Pro and the M4 with the same MoltenVK, and the cause is still open.

I have not tested on Linux or Windows. For every changed C/C++ file I compare its Linux and Windows preprocessor views (unifdef with __APPLE__ undefined) with main. 14 files are not macOS-only, and they change what Linux and Windows compile:

Whitespace or a comment only: ExportMacros.hpp, DisplayWindow.cpp, ApplicationProcessParameters.cpp.
New headers that are empty off macOS: MainThread.hpp, DriverMemory.hpp.
The same behaviour, written portably:
- FileStream.hpp calls feof_unlocked/ferror_unlocked unqualified, so macOS can supply them.
- CxxAbiSupport.cpp names the __cxa_thread_atexit lambda.
- Jit.cpp adds #include <filesystem>.
- QueuedIndexReadOrder.cpp changes 4 reinterpret_casts to static_cast.
- GuestGenericCategory.cpp adds const_cast before the vtable casts.
- Formatting.cpp adds <cstdarg> and a clang __builtin_sysv_va_list alias.
- GuestRaiseException.cpp renames its SIGUSR1 constant to GuestSigUsr1 and uses the shared assembly macros.
New macOS target in shared converter code: the relinker's macOS image writer (MacOsImage.cpp) and the web converter's "macos" target option (wasm/Bindings.cpp).

Every other changed file has identical Linux and Windows views.

Homebrew: ccleste (PPSA59322) converts, gets signed, starts, reads its settings and assets and runs into gameplay. It stops at its first mesh-path draw whose pixel shader reads barycentrics, because MoltenVK exposes neither VK_EXT_mesh_shader nor a usable VK_KHR_fragment_shader_barycentric. That needs a non-mesh fallback and is left for a follow-up PR.

Checklist

cc @boykopovar @oneandonlydean @adriadam10. It would help if one of you could start the full Linux/Windows build for this PR (gh workflow run build.yml --repo boykopovar/AnyPS5 --ref main --field pr=8658).

macOS owns 0xfc0000000-0x1000000000 with maximum protection none, so the AGC driver's fixed 2 MiB block at 0xfe0000000 can never be mapped there, even in an empty process. Every driver test stopped with "the AGC driver memory was not mapped at its fixed address". On macOS the block sits at 0xee0000000, lower in the same system-reserved guest range that games are never given. It replaces only the guest arena's own inaccessible reservation; anything else mapped there is reported instead of overwritten. Guest compute shaders build the FS table pointer from two s_mov_b32 immediates. FindTableCalls records which instructions write it, and on macOS LinkUserDataCallees moves a pointer that falls inside the console block to the macOS block, both for the callee lookup and in the code the GPU runs. A move that cannot hold the new value is reported rather than left pointing at unmapped memory. DriverMemory.hpp holds both bases and the translation. Linux and Windows preprocessed sources are unchanged. Full macOS suite: 84 of 1,599 tests fail, down from 210; the 126 driver-memory failures pass and nothing newly fails. compute FS table calls tests pass.

No conflicts. One macOS build fix for new upstream code: - QueuedIndexReadOrder test: reinterpret_cast from std::uintptr_t to std::uint64_t is ill-formed where they are different types (macOS: unsigned long vs unsigned long long); static_cast gives the same value on every platform. Checked: the macOS build has 0 failed objects, 208/208 converter tests pass, the web converter builds and passes tools/test_web.py, and the new agc_driver_queued_index_read_order test passes under MoltenVK.

GuestMemoryLifetimeSupport, IndirectArgumentsPendingProducer and PerformanceOverlay opened the Linux loader name on every non-Windows host, so on macOS they failed with "Failed loading libvulkan.so.1". They now open libvulkan.1.dylib on Apple, like the driver and the other tests already do. The six agc_driver_guest_memory_lifetime* tests and agc_driver_indirect_arguments_pending_producer now pass under MoltenVK. Linux and Windows preprocessed sources are unchanged.

The macOS acosf the guest acosf forwarded to is not correctly rounded under Rosetta: acosf(0.5f) returns 0x3f860a91, one ulp below the correctly rounded 0x3f860a92 that the console returns and guest_math checks (native arm64 macOS gets it right). On macOS the guest acosf now evaluates in double and rounds once, which is correctly rounded for every float input. guest_math now passes on macOS. Linux and Windows preprocessed sources are unchanged.

Darwin's nanosleep returns 0 after a full sleep but leaves errno set to ETIMEDOUT (60). The guest errno is the host errno, so a guest that slept successfully saw its errno changed, which guest_nanosleep checks (GuestNanosleep.cpp:82-84). On macOS the sleep loop now restores the caller's errno. guest_nanosleep passes on macOS. Linux and Windows preprocessed sources are unchanged.

FreeBSD and glibc inet_pton reject a dotted-quad component with a leading zero ("01.2.3.4"); macOS inet_pton accepts it, so sceNetInetPton returned 1 where the console returns 0 (GuestSceNet.cpp:738-739). On macOS AF_INET text with such a component is rejected before the host call; output and errno stay untouched. guest_sce_net now passes its inet_pton checks on macOS (it still stops later, at the unbound datagram shutdown check). Linux and Windows preprocessed sources are unchanged.

…nters Three macOS differences that stopped guest_sce_net: - Like Winsock, macOS neither wakes nor fails a receive blocked on an unbound datagram socket when the socket is shut down (shutdown even fails with ENOTCONN). The existing Windows wait for a local address now also applies on macOS, where an unbound socket reports port 0 from getsockname instead of failing, and it waits with poll(). - macOS answers an AF_INET6 lookup of an IPv4 literal with its IPv4-mapped form; FreeBSD and glibc, without AI_V4MAPPED, find no IPv6 address. The resolver skips IPv4-mapped answers on macOS. - sceNetCtlGetIfStat had no macOS byte counters and threw. It now sums the 64-bit if_data64 counters from the NET_RT_IFLIST2 routing list for interfaces that are up and not loopback. guest_sce_net and guest_net_ctl pass on macOS. Linux and Windows preprocessed sources are unchanged.

Two macOS differences that stopped guest_udp: - macOS refuses an address on a connected datagram socket with EISCONN; the console accepts it. When the address is the connected peer (same family, port and address), the datagram is now sent with send(), which goes to the same place. Any other address still fails with EISCONN. - macOS FIONREAD on a datagram socket also counts the queued address records (35 for one 19-byte datagram, 86 for two). The console and Linux report the next datagram's size, which macOS gives through SO_NREAD. FIONREAD, the poll readiness count and the stat-style queued byte count now use SO_NREAD for datagram sockets. guest_udp passes on macOS. guest_socket_read_write is flaky on macOS with and without this change (line 154, errno 54 or 35 instead of EPIPE; 2/30 runs fail on the parent commit). Linux and Windows preprocessed sources are unchanged.

After the peer closes, macOS can fail the next stream send with ECONNRESET where FreeBSD reports EPIPE, so guest_socket_read_write (GuestSocketReadWrite.cpp:151-154) failed intermittently with errno 54. On macOS SendError now maps ECONNRESET on a stream socket to EPIPE; the receive path is unchanged. guest_socket_read_write: 0 of 40 runs fail, against 2 to 5 of 30 before. Linux and Windows preprocessed sources are unchanged.

macOS refuses shutdown(SHUT_RD) with ENOTCONN once the peer has shut down its sending side, although the socket is still connected (getpeername succeeds) and queued data stays readable. The console accepts it, so guest_socket_pair stopped at GuestSocketPair.cpp:146-149. On macOS a SHUT_RD that fails with ENOTCONN on a socket that still has a peer is treated as done; the guest's receive-shut state is recorded as before. A socket that was never connected still fails with ENOTCONN. guest_socket_pair now passes that check (it next stops at the SOCK_SEQPACKET socket pair). Linux and Windows preprocessed sources are unchanged.

macOS has no AF_UNIX SOCK_SEQPACKET (socketpair fails with EPROTONOSUPPORT), so guest_socket_pair stopped at GuestSocketPair.cpp:177. A connected AF_UNIX datagram pair keeps the same record boundaries and reports MSG_TRUNC for a short receive, which is what the guest relies on; on macOS a SEQPACKET pair is now backed by one. The guest socket keeps its SEQPACKET type. guest_socket_pair passes on macOS (3 of 3 runs). Linux and Windows preprocessed sources are unchanged.

macOS poll() never reports POLLPRI (or POLLRDBAND) for pending TCP urgent data and does not wake for it, so guest_poll's PollsUrgentData check (GuestPoll.cpp:202-208) failed. A MSG_OOB | MSG_PEEK check is not usable on macOS: the peek uses the urgent byte up, so the guest's later recv(MSG_OOB) fails with EINVAL. On macOS the guest poll now asks SIOCATMARK, which reports the urgent byte without consuming it once it is the next byte to read, for every socket that asked for POLLPRI, and ORs POLLPRI into its revents. While such a socket waits, the host wait is cut into 10 ms slices so a timed or infinite poll still returns when urgent data arrives. Other polls wait exactly as before. guest_poll passes on macOS (3 of 3 runs), and 48 of 50 networking and poll tests pass (the two SSL transport tests fail for another reason). Linux and Windows preprocessed sources are unchanged.

macOS timer coalescing wakes a nanosleep about 25% late on this machine (1.4 ms takes about 1.76 ms, natively and under Rosetta), so the fixed 100 us spin tail overshot: PreciseSleepUs(1500) took a median of 1.76 ms against precise_wait's 1.65 ms bound (PreciseWait.cpp:125-131). On macOS the sleep part now covers 70% of the wait before the spin tail, so the late wake-up falls inside the spin and the wait ends at the deadline. precise_wait passes on macOS (8 of 8 runs); the other 44 timing and sleep tests are unchanged (guest_itimer and guest_thread_cpu_clock fail as before). Linux and Windows preprocessed sources are unchanged.

Under Rosetta, clock_gettime(CLOCK_THREAD_CPUTIME_ID) badly under-counts a thread's CPU time: about 5 ms for 200 ms of busy CPU on this machine, while thread_info(THREAD_BASIC_INFO) reports 199 ms. The guest's per-thread CPU clocks already read the Mach counter, so the two clocks disagreed and guest_thread_cpu_clock failed (GuestThreadCpuClock.cpp: 75-76). On macOS the guest CLOCK_THREAD_CPUTIME_ID now reads the calling thread's THREAD_BASIC_INFO user and system time too. guest_thread_cpu_clock passes on macOS (3 of 3 runs); 46 of 47 timing and CPU-time tests pass (guest_itimer fails as before). Linux and Windows preprocessed sources are unchanged.

One conflict, in libSceVideoOut DisplayWindow.cpp: upstream's new present telemetry (boykopovar#7298) and the macOS main-thread window-title update are both kept. Checked: the macOS build has 0 failed objects, 209/209 converter tests pass, and the web converter builds and passes tools/test_web.py.

No conflicts. Checked: the macOS build has 0 failed objects, 209/209 converter tests pass, and the web converter builds and passes tools/test_web.py.

The guest exception runtime (libc/src/exception/Unwind.cpp) walks frames through __eh_frame, but Apple clang emits compact unwind and ld64 keeps only __unwind_info, so on macOS a guest throw could not step out of libc's own frames: guest_bad_optional_access died with "[libc] unhandled exception 'St19bad_optional_access'". ld64's -keep_dwarf_unwind is obsolete and has no effect. configure_windows_unwind already marks every image the guest unwinder must walk; on Apple it now compiles them with -femit-dwarf-unwind=always and links with -no_compact_unwind, so __eh_frame is kept. Full macOS suite: 66 of 1,624 fail, down from 69; guest_bad_optional_ access, unwind_callback_abi_backtrace and unwind_callback_abi_forced_ unwind now pass, nothing newly fails. Linux and Windows builds are unchanged (the new branch is elseif(APPLE)).

For memory the guest registry does not know (the title's heap blocks, thread stacks), sceKernelVirtualQuery read /proc/self/maps on every non-Windows host. macOS has no /proc, so every such query failed with EACCES, and guest_system_configuration (KERN_PROC vmmap) and guest_kernel_is_stack failed. On macOS it now asks mach_vm_region for the region around the address and reports its extent and read/write/execute protection, like the Linux maps line. Unmapped or unreadable memory still fails with EACCES, as on the PS5. guest_system_configuration and guest_kernel_is_stack pass on macOS; 30 of 36 memory, stack and system tests pass, and the 6 that fail failed before. Linux and Windows preprocessed sources are unchanged.

No conflicts. One macOS build fix for new upstream code: - BdaResources test: memfd_create and MFD_CLOEXEC are Linux-only; on macOS the 16 KiB views block is backed by an unlinked POSIX shared memory object instead. Checked: the macOS build has 0 failed objects, 209/209 converter tests pass, the web converter builds and passes tools/test_web.py, and agc_driver_graphics_tests (which includes BdaResources.cpp) passes under MoltenVK.

Apple clang inlines the direct Descend call in Worker, so Descend's 4 KiB probe array lands in Worker's own frame (0x1028 bytes, via __chkstk_darwin), below the local the stack-size check measures from. The local then sat 4,208 bytes below the top of a 16 KiB guest stack against a 4,096-byte allowance, and the check at GuestPthreadStack.cpp:70 failed although the guest got its whole stack. Descend is marked noinline on Apple, so Worker's frame is 0x18 bytes again and the test measures the guest stack, as on Linux. guest_pthread_stack passes on macOS (5 of 5 runs). Linux and Windows preprocessed sources are unchanged.

The test read page state from /proc/self/maps and re-executed itself through /proc/self/exe; macOS has neither, so the child checks failed at line 54 and the parent at line 541. On macOS QueryPage now reads the Mach region (mach_vm_region) holding the address and encodes its permissions like the maps line, and the child is started through the path from _NSGetExecutablePath. guest_kernel_stack_protect passes on macOS (3 of 3 runs). The two main-stack-bottom child checks report that they skip, as they do whenever memory is mapped within 1 MiB of the main stack's bottom, and the unlimited-stack check skips because the hard stack limit cannot be raised; the thread-stack guard checks run. Linux and Windows preprocessed sources are unchanged.

On Linux fflush returns early for a stream that is not writing, so pushback from ungetc and buffered input survive. Darwin's fflush on a reading stream discards both, so after fgetc/ungetc('Z')/fflush the next fread returned "\n" instead of "Zbcdef\n" and guest_standard_streams aborted. On macOS fflush now likewise returns early unless the FILE's __SWR (writing) flag is set. guest_standard_streams, guest_fopen_errors and guest_file_timestamps pass on macOS. Linux and Windows preprocessed sources are unchanged.

Darwin's readv, writev, preadv and pwritev fail with EINVAL when the vector count is 0, while FreeBSD (and so the PS5) transfers nothing and returns 0. sceKernelReadv(file, reads, 0) therefore returned an error and guest_file_vector aborted at line 245. On macOS the four calls now return 0 for a zero count before reaching the host call; negative and oversized counts are still rejected by CheckIovecs as before. guest_file_vector, guest_kernel_sys, guest_standard_streams and guest_fopen_errors pass on macOS. Linux and Windows preprocessed sources are unchanged.

installProcessParameters() had only Windows (.procpar section) and Linux (patched PT_SCE_PROCPARAM program header) branches, so on macOS the scripting and lazy modes reached _init_env with no process parameters and aborted with "invalid or missing process parameters". On macOS libc finds them through dl_iterate_phdr, which reads the relinker's __APS5DATA,__meta section of the main executable. The test now places such a section in its own image and fills it with version-2 metadata holding a PT_LOAD and a 0x61000001 header that cover its fake parameter block, mirroring what the Linux branch does to its headers. All 9 application_heap tests pass on macOS, including scripting and lazy. Linux and Windows preprocessed sources are unchanged.

On macOS guest mmap of any descriptor failed with ENOTSUP: the generic branch only accepted descriptors that IsHostSharedMemoryDescriptor recognised, and that had no macOS case, so it always said no. - IsHostSharedMemoryDescriptor now recognises the anonymous objects that CreateAnonymousFile makes on macOS: regular files whose last link was removed (st_nlink 0) and whose F_GETPATH name starts "anyps5-shm-". This also makes shm read/write/seek/truncate use the shared-memory rules (writes bounded by the object size), as on Linux. - mmap on macOS now takes the Linux branch: pin the descriptor, check its access mode against the requested protection, and accept shared-memory objects and regular files. guest_shm_open passes on macOS; guest_memory, guest_mmap_fixed_anonymous, guest_file_vector, guest_kernel_sys, guest_standard_streams and guest_file_timestamps still pass. Linux and Windows preprocessed sources are unchanged.

sceKernelJitMapSharedMemory with execute permission failed on macOS: mmap of a shared file mapping with PROT_EXEC returns EPERM there, so guest_jit_shared_memory aborted on its first read+execute view. On macOS DoMapShared now maps executable views in two steps: it maps the object read-write (shared) at any address, remaps those same pages onto the placed guest range with mach_vm_remap (VM_INHERIT_SHARE, no copy), drops the temporary view and sets the requested protection with mprotect. Writes through other views of the object stay visible to the executable view, and code written through a writable alias runs. Non-executable views keep the plain mmap path. guest_jit_shared_memory passes on macOS; guest_shm_open, guest_memory and guest_mmap_fixed_anonymous still pass. Linux and Windows preprocessed sources are unchanged.

guest_memory_lifetime's exit probe calls DoMprotect after the module's static destructors have run. DoMprotect asks GuestThreadStack whether the range is a thread stack, which locks the namespace-scope stackLock. With libc++ ~mutex calls pthread_mutex_destroy, so that lock threw "mutex lock failed: Invalid argument" (libstdc++'s mutex destructor is trivial, which is why Linux never hit it). On macOS stackLock, liveStacks and protectedStackPages are now allocated once and never destroyed, the same way DirectMemory and GuestAllocations already keep their state alive for the whole process. guest_memory_lifetime passes on macOS, and all 36 guest_pthread_*, guest_thread_*, guest_kernel_is_stack, guest_kernel_stack_protect and guest_memory tests pass. Linux and Windows preprocessed sources are unchanged.

ProcessParametersFixture::Install had only Windows and Linux branches and returned false on macOS, so guest_hmd2 aborted with "process parameters were not installed". On macOS it now fills a __APS5DATA,__meta section in the test image with version-2 relinker metadata (a PT_LOAD and a 0x61000001 header covering the fake parameter block), which libc's macOS dl_iterate_phdr reports to the process-parameter lookup — the same approach the application_heap test uses. guest_hmd2 passes on macOS and the 9 application_heap tests still pass. Linux and Windows preprocessed sources are unchanged.

libusb's configure check finds pipe2 in the macOS 27 SDK, but the SDK marks it available only from macOS 27, so on macOS 26 the weakly linked symbol is null and libusb_init_context jumped to address 0 from usbi_create_event. guest_usbd segfaulted in sceUsbdInit. On Apple HAVE_PIPE2 is now cleared before libusb's checks run, so libusb uses its pipe + fcntl(FD_CLOEXEC/O_NONBLOCK) fallback. The change sits in if(APPLE); Linux and Windows configure exactly as before. guest_usbd passes on macOS.

iamthehimansh · 2026-10-23T23:08:05Z

@krishalivaibhav Whenever you have a moment, the GuestAllocations.cpp branch is still ready to open as is: fix/macos-guest-allocations-dyld, one commit on main, which compiles on macOS. You can open it straight from that branch:

gh pr create --repo boykopovar/AnyPS5 --head iamthehimansh:fix/macos-guest-allocations-dyld

I've also sent the one-line depth-plane fix for #10369 as krishalivaibhav#2, based on your current head.

AI-assisted: yes (Cana)

iamthehimansh · 2026-10-23T23:38:19Z

@adriadam10 I closed this PR as you suggested. Split PR 1 is open as #10419: the shared-file macOS build fixes, 6 files, all macOS-only, rebased on current main, with the conventions check passing. The libc, libkernel, sockets, AGC and relinker split PRs follow one at a time from the same branch.

AI-assisted: yes (Cana)

… expiry on macOS An ITIMER_REAL expiry with a guest SIGALRM handler throws, since guest handlers are never run from the timer thread. On Linux and Windows that throw goes through libc's exception runtime, so an uncaught one runs the guest's std::set_terminate handler. On macOS the host keeps libc++abi, so the throw ended in libc++abi's terminate and the guest handler never ran. Throw through the guest runtime there, and record the general macOS gap in the libc debt file. Ported from boykopovar#8658. Co-authored-by: Himansh Raj <iamthehimansh@gmail.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… expiry on macOS An ITIMER_REAL expiry with a guest SIGALRM handler throws, since guest handlers are never run from the timer thread. On Linux and Windows that throw goes through libc's exception runtime, so an uncaught one runs the guest's std::set_terminate handler. On macOS the host keeps libc++abi, so the throw ended in libc++abi's terminate and the guest handler never ran. Throw through the guest runtime there, and record the general macOS gap in the libc debt file. Ported from boykopovar#8658. Co-authored-by: Himansh Raj <67262056+iamthehimansh@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

RenamePath compares source and destination with fs::equivalent and treats ENOENT from it as "not the same file". libc++ reports errc::not_supported instead when either path does not resolve, so on macOS renaming a file over a dangling symbolic link failed with EIO (guest_filesystem). A path that does not resolve cannot be the same file as the other, so not_supported is cleared the same way on macOS. The same guard is part of #7201 and #8658. Co-Authored-By: marion909 <103169397+marion909@users.noreply.github.com> Co-Authored-By: iamthehimansh <67262056+iamthehimansh@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

NativeStat, SharedMemory, Stdio, DynamicLoader, ModuleInfo and Rtld used Linux-only APIs (st_atim, memfd_create, sys/vfs.h, link.h and dl_iterate_phdr), so libkernel did not compile on macOS. Add macOS paths behind __APPLE__: st_*timespec and a long-path fstatat walk, an unlinked temp file for shared memory, statfs, dlopen/dladdr and Mach-O segment walks for module and rtld queries. The Linux and Windows paths are unchanged. Extracted from feat/macos-port (boykopovar#8658). Co-authored-by: Himansh Raj <iamthehimansh@gmail.com> Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

iamthehimansh added 30 commits October 19, 2026 22:45

Merge origin/main into feat/macos-port

3adbae3

No conflicts. Checked: the macOS build has 0 failed objects, 209/209 converter tests pass, and the web converter builds and passes tools/test_web.py.

Merge origin/main into feat/macos-port

a1bfdac

iamthehimansh closed this Oct 23, 2026

iamthehimansh mentioned this pull request Oct 23, 2026

build: compile FileStream, error categories, thread priority and two tests on macOS #10419

Open

12 tasks done

maximilianfv mentioned this pull request Oct 24, 2026

fix(libkernel): reach the guest terminate handler from an ITIMER_REAL expiry on macOS #10505

Open

12 tasks done

krishalivaibhav mentioned this pull request Oct 24, 2026

fix(libkernel): build the file, module and rtld calls on macOS hosts #10552

Open

12 tasks done

This was referenced Oct 24, 2026

test(agc): skip cull distance cases on devices without shaderCullDistance #10612

Open

perf(shader): reuse a memoized result by patching its copied words #10423

Open

imellag mentioned this pull request Oct 24, 2026

fix(shader): translate f32 conversions in f32 denormal modes 1 and 2 #10595

Open

12 tasks done

Navigation Menu

Repository navigation

feat: macOS host support (Apple Silicon via Rosetta 2, Vulkan on MoltenVK) - #8658

feat: macOS host support (Apple Silicon via Rosetta 2, Vulkan on MoltenVK)#8658
iamthehimansh wants to merge 289 commits into
boykopovar:mainfrom
iamthehimansh:feat/macos-port

iamthehimansh commented Oct 20, 2026 •

edited

Loading

Uh oh!

iamthehimansh commented Oct 23, 2026

Uh oh!

iamthehimansh commented Oct 23, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

12 participants

Conversation

iamthehimansh commented Oct 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Tested

Checklist

Uh oh!

iamthehimansh commented Oct 23, 2026

Uh oh!

iamthehimansh commented Oct 23, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

12 participants

iamthehimansh commented Oct 20, 2026 •

edited

Loading