OpenRCT2: the big-endian port, milestone 1 (2026-10-19)

The port followed the estimate of the same day. Upstream refuses big-endian on purpose (a static_assert in Platform.h), so this is our own out-of-tree series.

Starting point. The closed PR #27114, “AmigaOS 3.x (m68k, big-endian) port [RFC]” (Thomas Luebker, 43 commits, 2026-09-10/12, head a5dbca57c5be, closed unmerged as too invasive and AI-generated), contributed to OpenRCT2 and so GPL-3.0-or-later like the rest of the tree. Its endian core was applied on v0.5.6 (82b30979) with 2 conflicts resolved (DataSerialiserTraits.h, SawyerCoding.cpp) and cleaned of the Amiga parts. The work was done in a cross build (g++-16 for powerpc64 against a ppc64 sysroot) with the ppc64 binaries run under qemu-ppc64-static, next to a native amd64 build of the same tree.

Bugs found beyond the PR (each fixed in the series):

• the PR’s RCT12ResearchItem: RawValue swapped but the byte members read in little-endian order, so the research lists were garbage on big-endian;
• the PR’s maze-only swap of RCT12 track bytes 5/6: the importer reads Colour/Sequence, which alias MazeEntry, so the track’s flags2 differed on mazes (bpb.sv6, BigMapTest.sv6); fixed by host-ordered members and an unconditional swap;
• the PR’s park TILES swap was Amiga-only (#ifdef __amigaos__) and keyed mazes on a rideType heuristic: replaced by TileElement::LittleEndianToHost with host-ordered union bytes (path additionStatus, track sequence);
• CSStringConverter: the bsearch key was read as the first 2 bytes of a wchar_t (the high half on big-endian), so RCT2’s accented characters were not converted;
• TD6/TD4/TD7 track designs, scores.dat, the file index header, the sprite tool (the g2.dat build), the chunk writer header, the Sawyer and RCT1 checksums, the RLE magnify row offsets, the PNG importer’s offsets, the Ogg/FLAC formats, the 16/8-bit texture path, and coordinate structs written raw by the DataSerialiser (the track design actions);
• the build: DISABLE_HTTP misses <future>; ScriptingTests were built without scripting.

Checks, big-endian (ppc64 under qemu) against little-endian (amd64), the same source:

• gtest 275/275 on both, including the 19 ReplayTests recorded on x86 (entity checksums tick by tick);
• g2.dat, fonts.dat, palettes.dat and tracks.dat built by the big-endian sprite build: byte-identical;
• the oracle: every RCT2 scenario (26 SC6 in our rct2-data) and 9 test parks (SV6 and .park): load, simulate, save an uncompressed .park, compare the entity checksum and every chunk except authoring: 100 ticks 35/35, 5000 ticks 35/35, 10000 ticks with the park opened (guests in 28 of them, up to 2782) 35/35. Determinism needs OPENRCT2_FIXED_SEED (ScenarioBegin seeds from the clock) and SOURCE_DATE_EPOCH (authoring timestamps), both hooks in patch 0003;
• the renderer: 210/210 screenshots identical (giant zoom 2, rotations 0-3; 1920x1080 at zoom 0 rotation 0 and zoom 1 rotation 2) of the 35 parks after 10000 open-park ticks;
• track designs: all 169 TD6 in rct2-data imported, serialised (the network and replay form) and exported back to TD6 on both hosts: identical. This found patch 0006: the object entry checksum went into the serialiser in host order, and coordinate structs carried their (uninitialised) padding;
• a DWARF check of the S6/S4 swap tables: a dev-only test runs ByteSwapS6Data/ByteSwapS4Data on a pattern buffer (byte i = i*7+3, so 2-, 4- and 8-byte reversals are unambiguous), gdb dumps the layouts from DWARF, a script compares. S6: 342159 multi-byte fields outside unions, none missing or of the wrong width, no stray swaps (not swapped by design: the header’s magic, object count and version, swapped earlier, and Rides[].pad002, never read). S4: none missing. The unions were reviewed against what the importers read;
• the package: the debs installed in a ppc64 chroot under qemu, Electric Fields after 3000 open-park ticks gives the amd64 park chunk for chunk (only authoring differs: the version string), checksum 20994dc4a8b051a1.

Not covered: the zoomed-in (magnify) RLE path (the CLI screenshot can’t take a negative zoom), SDL texture formats other than 32-bit, audio (no device in the container: the Ogg/FLAC formats were fixed by reading; the sound is for the user’s ear on the PS3), RCT1 data (S4/TD4 checked only by the DWARF swap check), the network, scripting (built off).

RAM on the PC (the amd64 build of the same tree under Xvfb at 1920x1080, the SDL software renderer with SDL_FRAMEBUFFER_ACCELERATION=0, RSS median over 120 s after 60 s, the open-park saves after 10000 ticks):

With SDL’s GL renderer on llvmpipe: +140 MB (402-451 MB; SDL3 under sdl2-compat otherwise puts llvmpipe behind the window surface). The smaps of the lean Magic Mountain run: heap 113 MB (tile elements, objects, strings), 70 MB in one block, GameState_t (73.6 MB: the EntityRegistry 57.7 MB = 65535 x Entity_t of 512 B plus per-entity arrays), 21 + 15 MB (object images, g1 data), file-backed 25 MB. The memory knobs, biggest first: kMaxEntities 65535 to 16384 (about -43 MB; RCT2 parks hold up to 10000; needs the oracle re-run), the GL renderer out of the process (sdlsw), sound off (-10 to -35 MB), one thread and one malloc arena (-25 MB), 960x540 x2 (-9 to -11 MB), g1.dat mmap’d instead of read into the heap (17 MB), and smaller parks (Electric Fields -40 MB). The PS3 numbers are in the technical notes.