Entry 72

racecheck: check whole steps between cars_update and the collision's return

Filescrates/hwtr-hle/src/bin/racecheck.rs

racecheck reported differences with steering and turbo scripts: 5, 12 and 32 steps on the drive, speed and air states. That held at the earlier commit 2b4ca54 too, so it was not a regression. All of them turned out to be the harness, not the port.

  1. Steps cut by the vertical blank. Hle::frame stops at the vertical blank wherever the main loop is. A race step can therefore straddle two frames: at steps 85, 86 and 100 the step counter moved but the collision stages ran in the next frame. racecheck had compared a half-done original step with the port's whole one, and the original's wheel contacts and ground looked one step behind. racecheck now only checks a frame in which the cars' update (0x8004064c) began and the collision update (0x8004de6c) returned.
  2. The step's tail. The previous step's places, wrong-way watch, camera and clock tick could run at the start of the next frame. The start state is now memory as cars_update begins, and the end state is memory as the collision update returns, both captured by hooks, instead of the frame's edges. This fixed contact_time off by one clock tick.
  3. The wrong-way watch. 0x8005c9b4 runs inside cars_update, after the places. racecheck now runs the port's laps::watch_way after the car update.
  4. Port events in the diff. The Pending event queues the port keeps for the race (sounds, lines and others) printed in the debug diff and misaligned it. They are cleared before diffing.
  5. Resets. A reset step is counted, not compared: racecheck does not model the reset path.

On a difference, racecheck now also reports:

  • the car as the collision began against the port's car after its update
  • the whole world before the stages
  • each stage (zones, wheels, ground) run from the original's world against the original's result

One gotcha found on the way: Machine::check keeps one check per address, so a second hook on the same function replaces the first.

Result, with every whole step checked:

  • Script 0:x:600,100:right:40,250:left:60: 412 + 606 + 612 + 612 steps, all the same.
  • Script 0:x:900,50:left:120,300:square:60,400:right:200: 411 + 608 + 612 + 612, all the same.
  • Script 0:x:900,0:r1:900,...: everything the same except 17 resets, which are not checked.

The port's race step matches the original on every step checked, including today's computer-car walls stage.

Still unknown

The reset path is not checked by racecheck