1. For %fs/fsbase the patch follows the macOS logic with LDT descriptor registration and Linux with switching. One notable difference is that on 32->64 transition we set %fs to GSEL(GUFS32_SEL, SEL_UPL) before restoring fsbase, otherwise FreeBSD will just revert it by reloading the selector [at the first opportunity](https://github.com/freebsd/freebsd-src/blob/5673462af5330df20733008beae48a24...). GSEL(GUFS32_SEL, SEL_UPL) is the default %fs value on FreeBSD and is special-cased to save/restore actual fsbase value to/from PCB. 2. I was told we could get rid of fsbase glitches in signal handlers by blocking signals with [sigfastblock(2)](https://man.freebsd.org/cgi/man.cgi?query=sigfastblock) between %fs reset to the default value and fsbase reset to pthread_teb. This is currently a part of internal API for libthr, which could be exposed as pthread_signal_block_np for Wine. I'm on the fence whether it's worth it. 3. I fully admit I have no idea what registers are worth preserving around fallback sysarch(AMD64_SET_FSBASE) syscalls and whether it's appropriate to push those registers to stack. ("Kernel" stack should be fine, I assume?) Syscalls definitely clobber r8-r11. 4. For %ss see https://lkml.org/lkml/2015/4/24/216. FreeBSD doesn't have a similar workaround in the kernel, so it goes into Wine.
-- v5: ntdll: Unbreak new wow64 mode on FreeBSD.