Fixed the meso2 and romeric5 maps on Amiga

[BSzili]

On Amiga there was a race condition in the order the entities get to think first when SV_SpawnServer is run. It was caused by the double to float conversion of host_frametime in SV_Physics_Pusher and later sv.time + host_frametime in SV_RunThink. Because of this some doors could be used or touched before their InitDoor function is executed causing a program error. I tried multiple ways to fix it, but it quickly became a game of whack-a-mole, so I ended up adding small constants to these checks to give them a bit of tolerance.

[sezero]

I wonder, wouldn't it be safe to always use those epsilons and not just for aos3.. (No, haven't yet tested myself.)

[BSzili]

Since it reinforces the existing behavior it probably wouldn't break other platforms, but this whole thing is quite fragile, so I wouldn't want to risk it.

[sezero]

SV_RunThink hexenwordl server doesn't need the same workaround?

[BSzili]

Oddly it didn't. There are some differences in how host_frametime is handled there, so it worked with only the SV_Physics_Pusher fix.

[sezero]

Asked a friend about his opinion on making the solution available to all platforms.

[sezero]

Asked a friend about his opinion on making the solution available to all platforms.

He says, in essence:

• The solution should be the same for all platforms. (Not all math processors of all cpus work the same way - even with the same C code, something crashing with x86 may not do so with ppc, and vice versa. But having the same code is best for debugging, even though the results are never 100% same.)

• However: as time progresses, i.e. as the float/double time value increases, the epsilon incrementation won't happen, the floating point value will stay the same.

The second part of what he says is what I'm somewhat worried about.

[BSzili]

If other platforms are not affected then it's probably better to keep this Amiga-only. The second part is no problem there, as the workaround only needed for the initial SV_Physics calls in SV_SpawnServer, where host_frametime is always initialized to a known value.

[sezero]

OK, will probably merge this as is tomorrow. Two questions until then:

• does gcc2.95 builds have the same issue?
• is it possible that -ffast-math is causing an adverse effect resulting in the issue?

[BSzili]

I couldn't test with gcc2 for now, but disabling -ffast-math had no effect on it.

[sezero]

OK, this is in.