[Opvolger]

This adds support for 64-bit RISC-V Linux (RV64GC, so no vector extension needed). Half-Life 2 builds, runs and renders correctly on two RISC-V boards with an AMD GPU.

The approach follows the existing ARM support: wherever the code has an __arm__ / __aarch64__ case, riscv64 takes the same non-x86 path. For SSE, SIMDe is used in the same way that sse2neon.h is used on ARM.

How to build

On a riscv64 machine running Debian 13 (trixie):

scripts/build-debian-riscv64.sh

This works like the other scripts/build-*.sh scripts. The only new dependency is libsimde-dev (SIMDe headers; on Fedora simde-devel).

Changes

• public/tier0/platform.h: new PLATFORM_RISCV64, handled next to ARM for PLATFORM_64BITS, DebuggerBreak, the FPU control word (no-op), little-endian and Plat_Rdtsc (clock_gettime).
• SIMD: new common/sse2simde.h, the RISC-V counterpart of sse2neon.h. It enables SIMDe's native aliases, so the existing _mm_* SSE code compiles unchanged. It is included from ssemath.h, vector4d.h and mathlib/sse.cpp. sse.cpp takes the same non-inline-asm paths as ARM.
• public/mathlib/mathlib.h: RoundFloatToInt and RoundFloatToUnsignedLong use lrintf / llrintf (round to nearest even, like cvtss2si on x86).
• public/materialsystem/imesh.h: the intrinsics paths instead of the x86 MMX inline asm.
• public/tier0/tslist.h: the lock-free list needs a 128-bit CAS. RISC-V has no such instruction (no Zacas), so __atomic_compare_exchange_n is used, which libatomic provides (lock-based).
• tier0/cpu.cpp, tier0/cpu_posix.cpp, tier1/processor_detect_linux.cpp, tier1/reliabletimer.cpp, tier1/pathmatch.cpp, public/saverestoretypes.h, public/steam/steamtypes.h: no x86 cpuid, rdtsc or int3 on riscv. The arch name is riscv64 and the vendor RISC-V.
• wscript: -fsigned-char for riscv (char is unsigned there, like on ARM, and the code assumes x86). It also links libatomic, as LIB rather than a link flag, so it comes after the objects with --as-needed.
• scripts/build-debian-riscv64.sh: native build script for Debian 13.

Testing

Both boards run Debian 13 with an AMD Radeon R9 290 (radeonsi, desktop OpenGL) on PCIe and KDE Plasma. The test was a timedemo on a d1_canals_01 demo (5449 frames, 1280×720):

Board	SoC	Build	Result
Milk-V Jupiter	SpacemiT K1 (8 cores)	release	~52 fps
StarFive VisionFive 2	JH7110 (4× U74)	debug (from build-debian-riscv64.sh)	27 fps

The x86_64 build was checked again with these changes: it still builds and runs.

Under Wayland, start the game with SDL_VIDEODRIVER=x11 until #493 is merged. That crash is not RISC-V specific.

Not included

• No -march flags: the default rv64gc works on every RV64GC board. I also tried an RVV (vector) build on the Jupiter. It was not faster, because the game is limited by the main thread, so I left it out to keep this PR small.

How this was made

I made this together with Claude Code, Anthropic's AI coding assistant. I guided it throughout and corrected it where needed, so that the change fits the existing project structure:

• the same non-x86 paths as ARM
• a SIMDe header next to sse2neon.h
• the same kind of build script as the other scripts/build-*.sh, with no extra container or cross-compile tooling

I tested the result myself on both boards listed above.