[ksco]

Registers updated inside a loop should be written back to xEmu when dirty, so the next iteration will see the correct values.

[ptitSeb]

Mmmm, the idea of preload was to load the regs before the loop, so the read is done only one time. I don't understand why this should be needed? or is it for another jump and not the main loop?
I don't understand this fix, as this is exactly what preload is supposed to prevent.

[ksco]

We have to store the dirty preloaded registers back so they stay clean at the loop head -- in case there is a call_c inside the loop; the stale values of an xmm register can be restored from stale memory.

Another (maybe better?) way to do this is to have the host_call mark and skip loops with hostcalls. This is how LoongArch handles this, IIRC.

[ptitSeb]

We have to store the dirty preloaded registers back so they stay clean at the loop head -- in case there is a call_c inside the loop; the stale values of an xmm register can be restored from stale memory.

Another (maybe better?) way to do this is to have the host_call mark and skip loops with hostcalls. This is how LoongArch handles this, IIRC.

Not ideal, but better yes.
The ideal way would be for call_c to not skip those preloaded regs, but that might be more tricky to track down I guess.

[ksco]

I can try.

[ptitSeb]

This might help #4065 also.