@hrydgard sorry for mentioning you, just incase this issue slipped by you, it has been affecting me as well

Does this happen on desktop too (when using IR Interpreter)?

It'd help if someone has the game and wants to try debugging the IR:
https://github.com/hrydgard/ppsspp/wiki/How-to-debug-the-JIT

-[Unknown]

Not a regression, while unfortunate I'm bumping this to 1.8.0.

At least on Windows in the demo (using the workaround in #3365), this seems to be working correctly.

Is this still an issue in the latest git build? Unfortunately, iOS is generally problematic as far as builds and the latest OS, so it may be necessary to use Xcode to compile to test. Make sure the version says "v1.8.0-565-gd8db8ecfc" or higher.

-[Unknown]

Running PPSSPP 1.9.3 on iOS still exhibits this behavior with IR interpreter. Its a shame because the game runs beautifully otherwise. The character models remain glitched.

Just Switch To Interpreter It works for me

Just Switch To Interpreter It works for me

Too slow using Interpreter.

IR Interpreter is close to dynarec (jit) performance.

MGSPW_IR_INTERPRETER.ppdmp.zip

Maybe related to #12991

Disabled JIT Functionality VFPU_VEC can fix this issue using the latest build 🎉

This issue seems to be exist only on mobile platform (ARM/ARM64 only?) since i can't reproduce it on Windows but can reproduce it on Android.

Does IR have a separated codes for ARM and x86 just like JIT?

Edit: Apparently i can reproduce it on Ubuntu (WSL2), so i guess it wasn't specific to ARM/ARM64, may be because it use VM? not sure about a real linux box

Yeah this is some kind of bug in the IR interpreter / compiler, probably repros on all platforms with that.

I think I couldn't reproduce this in the demo, at least, on Windows. I don't think I tried Android or WSL2. Not sure why it'd be different, though...

-[Unknown]

VRot is VFPU_VEC, I suspect that's the most likely culprit here. ARM and x86 don't do IR differently (currently, eventually it'd be nice to have a jti backend from IR), but sin() and cos() are used, which might differ between those platforms.

Some things to try:

1. DISABLE only IRFrontend::Comp_VRot() and see if that makes this work.
2. Check for an overlap hazard, i.e. if anything in Handle vrot overlap and vscl/vmscl prefixes more accurately #16302 helps.

-[Unknown]

I've tried inserting DISABLE; on the first line of IRFrontend::Comp_VRot() function, but it didn't fix this issue.

Okay, disabling IRFrontend::Comp_VV2Op(...) alone seems to fix this issue, but i only tested this briefly on Ubuntu(WSL2), not sure about Android since i need a large free space to build Android version (i'm running out of free space)

To be more precise, disabling case 2: // vneg inside IRFrontend::Comp_VV2Op() function will also fixes this issue.

Okay, that's very interesting and pretty specific. I'll dig into vneg soon and see what could be wrong.

-[Unknown]

Had a little to check and cleaned up something that MIGHT cause an issue there and COULD explain a difference between Windows and Linux. Do these changes help at all?

master...unknownbrackets:ppsspp:ir-vneg

-[Unknown]

Had a little to check and cleaned up something that MIGHT cause an issue there and COULD explain a difference between Windows and Linux. Do these changes help at all?

master...unknownbrackets:ppsspp:ir-vneg

-[Unknown]

Yes, this fixes the issue on Ubuntu(WSL2), not sure about Android as there is no android artifact yet

Build should be available soon - will close if we can confirm this fixes it on Android or iOS too, but I expect it to.

-[Unknown]

Build should be available soon - will close if we can confirm this fixes it on Android or iOS too, but I expect it to.

-[Unknown]

I've tried the Android build and it also got fixed now on the latest dev build

Hurray, thanks for checking.

-[Unknown]