Interpreter-JIT discrepancies

[wheremyfoodat]

Some ROMs show different behavior between the 2 CPU cores, and are even broken in the latter Example 1: Relevant files (reposting with the author's permission): https://cdn.discordapp.com/attachments/466353899670863873/806240805073190973/psx.bin https://cdn.discordapp.com/attachments/466353899670863873/806460090336149514/psx.cue

On the interpreter, this shows an anime girl listening to music, as it should On the JIT, there's no anime girl nor music to be found.

Example 2: The BIOS shell (with my memcard) is colored differently between the 2 CPU cores Picture 1 is interpreter, picture 2 is JIT

My memory card: memcard1.zip

[nicolasnoble]

The bios' color is supposed to be random, isn't it?

[wheremyfoodat]

JIT: Tekken 2 has broken graphics. I'd initially thought this would be due to the game's weird handling of GPU DMAs (being the icon of infinite linked list DMAs), but this works in the interpreter and should be before it issues any weird DMA

[wheremyfoodat]

The bios' color is supposed to be random, isn't it?

I believe it's generated based on memcard contents? Not sure

[nicolasnoble]

Not sure where it's taking its entropy from.

[wheremyfoodat]

I'll change my memcard contents as an experiment and see if the color changes

[wheremyfoodat]

After changing my memory card contents, the screen seems to stay the same Not sure on the entropy source now.

[nicolasnoble]

To be fair, the only "good" source of srand would be root counters. JIT and Interpreter have different cycle bias in general due to differences in memory access functions. This might be what's happening here. A good test would be to fuzz the delay of the cdrom interrupts, which would definitely influence the value of the root counters when the shell decides to pop into the menu. There's interrupt scalers for this.

[wheremyfoodat]

I'll make sure tomorrow but yeah seems to be timing-dependent, will have to look deeper for the other ones though.

There's also the "JIT literally segfaults on Amidog's CPU test" thing which makes it harder to debug, regrettably.

[wheremyfoodat]

JIT jumps to this point in the BIOS (system exit) in Gran Turismo, after several failed CDFileQuery calls, and hangs.

Screen where it hangs

Last CD-ROM activity before hanging (same sequence of commands gets repeated a few times)

[wheremyfoodat]

Gran Turismo works on x86-64 JIT

[wheremyfoodat]

JIT: Tekken 2 has broken graphics. I'd initially thought this would be due to the game's weird handling of GPU DMAs (being the icon of infinite linked list DMAs), but this works in the interpreter and should be before it issues any weird DMA

This relies on GTE->GPR loads (MFC2/CFC2) being delayed

[wheremyfoodat]

Tekken has been fixed on the x64 backend in #744

[wheremyfoodat]

Cross post from #850: Disney's Tarzan reportedly works on x64, not on x86-32

[StanleyKid-22]

Cross post from #850: Disney's Tarzan reportedly works on x64, not on x86-32

I clarify issue: it was about ROM's version 1.1. 1.0 works great on x86-32.

[StanleyKid-22]

x86-32 Gekido: Urban Fighters and Metal Slug X shutdown emulator altogether. Mortal Kombat 4 works, but rarely crashes on intro movies.

[wheremyfoodat]

Oh, nice catch, thanks