[tgl][bisected][regression] GPU hang in The Witcher 3

System information

System:    Host: ccallawa-pnp-02 Kernel: 5.8.0-050800rc7-generic x86_64 bits: 64 compiler: N/A Desktop: Gnome 3.36.3
           wm: gnome-shell dm: GDM3 Distro: Ubuntu 20.04.1 LTS (Focal Fossa)
Graphics:  Device-1: Intel driver: i915 v: kernel bus ID: 00:02.0 chip ID: 8086:9a49
           Display: server: X.Org 1.20.8 driver: fbdev unloaded: modesetting,vesa compositor: gnome-shell
           resolution: 1920x1080~60Hz
           OpenGL: renderer: Mesa Intel Xe Graphics (TGL GT2) v: 4.6 Mesa 20.2.0-devel (git-12dd5455f4) direct render: Yes
Proton version: Proton 5.0-9

Description

The Witcher 3 fails to render its introductory cutscene, instead displaying a black screen. A GPU hang is logged in the kernel ring buffer:

[  133.197144] i915 0000:00:02.0: [drm] GPU HANG: ecode 12:1:859ffffb, in witcher3.exe [2433]
[  133.197146] GPU hangs can indicate a bug anywhere in the entire gfx stack, including userspace.
[  133.197146] Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/intel/issues/new.
[  133.197146] Please see https://gitlab.freedesktop.org/drm/intel/-/wikis/How-to-file-i915-bugs for details.
[  133.197147] drm/i915 developers can then reassign to the right component if it's not a kernel issue.
[  133.197147] The GPU crash dump is required to analyze GPU hangs, so please always attach it.
[  133.197147] GPU crash dump saved to /sys/class/drm/card0/error
[  133.298176] i915 0000:00:02.0: [drm] Resetting rcs0 for stopped heartbeat on rcs0
[  133.298192] i915 0000:00:02.0: [drm] witcher3.exe[2433] context reset due to GPU hang

witcher3-gpuhang-error-dump

I reproduced this issue on two platforms with packages from the Oibaf PPA as well as manual builds. I performed a manual bisect, and found that reverting de0ebaf0 resolves the hang.

While bisecting, I found it necessary to reboot between build+deploy of a new driver rev; simply copying the new binaries over the old ones isn't enough to trigger the behavior.

@majanes is working to reproduce the issue independently.