riscv64: forward-port RISC-V 64-bit host support onto Wine 11.15 (arm64ec)

[Lfan-ke]

Forward-ports André Zwing's RISC-V 64-bit host support from the frozen riscv64 branch (Wine 9.0) onto the arm64ec base (Wine 11.15), after two years of drift, so RISC-V hosts build against current Wine again.

Contents

The riscv64 host arch support rebased commit-by-commit onto 11.15 (André's authorship preserved, forward-port adaptations signed off): configure host detection, winnt.h RISCV64 CONTEXT, winebuild target, the ntdll host backend signal_riscv64.c realigned to 11.15's signal_arm64.c thread_data/syscall/dispatcher API (most of the drift), the PE-side signal_riscv64.c and dwarf.h, and the per-arch switches across ntdll, server, dbghelp, gdi32, msvcrt, kernel32, kernelbase, wow64, rpcrt4, oleaut32, setupapi, winspool, localspl, wineboot, winecrt0.

On top of the rebase

• rpcrt4: re-add the riscv64 thunks that 11.15 moved into thunks.c (call_stubless_func, call_server_func), plus the four Ndr*ClientCall vararg thunks in ndr_stubless.c.
• ntdll syscall frame: drop a copied arm64 NEON field and fix the size, then add an f[32] area and save/restore f0-f31 in the dispatcher so FP state and CONTEXT_FLOATING_POINT survive a syscall.
• segv_handler: switch to the new virtual_handle_fault() signature and drop a debug abort that killed every fault.
• handle_syscall_fault: implement, so a fault inside a syscall returns a status instead of crashing.
• trap_handler: report EXCEPTION_BREAKPOINT for ebreak.
• RtlCaptureContext: fix the F0-F31 stores, which were four bytes low.

Not included / known limits

• No x86 emulation backend, so x86 Windows programs don't run on a RISC-V host. That needs a riscv64 Windows PE, which no toolchain produces (clang emits ELF for riscv64-*-windows, and there is no riscv64-w64-mingw32).
• The dispatcher captures f0-f31 after NtCurrentTeb(), so the caller-saved fp registers may hold post-call values; the callee-saved fs0-fs11 that must survive a syscall are correct.
• except_riscv64.c still has a local __CxxFrameHandler that will clash with the shared except.c once riscv64 becomes an enabled PE target.

[AndreRH]

Welcome and thanks for your contribution!
I had a quick glance at it and it seems you forgot the assembler stuff in wineboot and rpcrt4.
It also would be good to preserve the asm comments in wineboot and update the comments in ntdll to the new syscall frame situation.
For verification I suggest to run at least winecfg, wine --version is too low level.

[Lfan-ke]

Thanks for the review. Pushed three commits:

rpcrt4: 11.15 moved the stubless/vtbl thunks, call_stubless_func and call_server_func into dlls/rpcrt4/thunks.c, so I re-added riscv64 there (call_stubless_func, call_server_func via args_stack_to_regs, and the two THUNK_ENTRY macros) plus the four NdrClientCall2 / NdrAsyncClientCall / NdrClientCall3 / Ndr64AsyncClientCall vararg-spill thunks in ndr_stubless.c, since the base has no C fallback for those.

wineboot: the only asm site is read_tsc_frequency, and RISC-V has no userspace equivalent of aarch64's CNTFRQ_EL0 (the timebase frequency comes from the device tree, not a CSR), so it stays on the return-0 path and falls back to the power-info MaxMhz, like the earlier riscv64 branch. Let me know if you had a specific wineboot asm in mind.

ntdll: fixed the frame description - dropped the arm64 NEON leftover and the 0x330 size, enabled the size assert at the real 0x118 and matched the dispatcher/callback asm. I also moved segv_handler to the new virtual_handle_fault(thread_data, EXCEPTION_RECORD *, stack) signature and dropped the debug ERR/abort_thread that was short-circuiting all fault handling.

Verification: the riscv64 host side builds and runs under qemu-riscv64.

$ QEMU_LD_PREFIX=/usr/riscv64-linux-gnu qemu-riscv64-static server/wineserver --version
Wine 11.15
$ QEMU_LD_PREFIX=/usr/riscv64-linux-gnu qemu-riscv64-static loader/wine --version
wine-8.14-18485-ge0d21bb2e (Hangover)

winecfg I can't drive from this tree yet - the amd64 guest backend (dlls/xtajit64) is still the arm64ec carryover stub and this build is headless (--without-x), so a guest PE can't execute or show a window. Happy to work toward that separately.

One note in passing: a few pre-existing riscv64 gaps remain (handle_syscall_fault NYI, no FP save area in the syscall frame, RtlCaptureContext storing the F-registers 4 bytes low, and except_riscv64.c colliding with the shared __CxxFrameHandler once riscv64 is a PE arch) - happy to send follow-ups if useful.