Green Lifter (Claw Guys) Faster in DOS than in Rebirth

[Mako88]

I tested this on D1 level 4 on Hotshot. In the original DOS flying backwards away from one of these bots without any sliding is BARELY enough to get away from them. In Rebirth, it's very easy to pull away while only using reverse with no sliding.

[vLKp]

Was this correct in 0.58.1-d1x?

[Mako88]

Yes, it was correct.

[Mako88]

I've realized this is also true of other robots as well. Either the ship is moving faster, or the robots are moving slower. (This seems to be the case across all difficulty levels, but is the most apparent on Hotshot and above).

[Mako88]

This happened sometime after commit 93888b4 (A build made at that commit is still correct). The more I compare, the more I don't think it's so much a max speed issue as much as a behavior issue. In older builds (and the original) they bots would constantly chase you, whereas now they seem to randomly slow down or become...disinterested.

[vLKp]

That is good news. That commit is only 1230 commits back from tip. Since our history is mostly linear, git bisect should be able to find the first bad commit in ~11 steps.

[Mako88]

According to git bisect, e6887bb is the first bad commit. (Although that could be slightly off since this is a relatively difficult thing to test accurately).

[vLKp]

That is probably not the first bad commit. It should have no effect on the generated code, since it is just the choice between passing a pointer and a reference. However, that gives a good starting point for further analysis.

[Mako88]

I tried to do another git bisect, but I got a commit that was over 1,000 commits away from the one I posted above (e40172f). This seems like it's going to be tough to debug... :/

[Mako88]

For completeness sake, here's how I'm testing:

I fly into the main room on level 4 of D1 on insane. I then get a claw guy a ship length or so away from me where I can see him and fly backwards using only reverse, while turning in a circle and keeping away from the walls (this is about as difficult as it sounds :P ). In 0.58.1 I can do this as long as I want...I'll very VERY slowly creep further away from him, but he'll constantly chase me. In the latest code he'll randomly break to the inside of the circle and stop chasing me for a moment, making it much easier to escape.

Also on Hotshot it seems as if their max speed is less in the current build than in 0.58.1.

[Mako88]

Haha! I found the commit: 2a19da8

I discovered it was MUCH easier to see the bug on Hotshot rather than on Insane, and this is the commit I got. It's actually from the same batch of commits as my first bisect (which was done with a much less reliable way of determining it), so I'm confident this is it.

[vLKp]

That fits. That commit accidentally changed the order of operations. Difficulty levels below Insane are affected more severely.

[Mako88]

Confirmed fixed. Closing.

[vLKp]

Based on how the values were handled, Ace (difficulty=3) would work correctly. Insane would be close, but not right. Lower difficulties would be more wrong, with Hotshot being the most affected.

In the good build, we compute result = (int)((value * (difficulty + 5)) / 4). In the bad build, we compute result = value * (int)((difficulty + 5) / 4).

For Hotshot (difficulty=2):
Good: (int)((value * 7) / 4) = (int)(value * 1.75)
Bad: (int)(value * (int)1.75) = (int)(value * 1)

For Ace (difficulty=3):
Good: (int)((value * 8) / 4) = (int)(value * 2)
Bad: (int)(value * (int)2.00) = (int)(value * 2)

For Insane (difficulty=4):
Good: (int)((value * 9) / 4) = (int)(value * 2.25)
Bad: (int)(value * (int)2.25) = (int)(value * 2)

That is why it seemed more difficult to spot when using Insane.