Turns out the squashfs mount has some non-trivial RAM overhead.
Copy the game to local storage and run from there to avoid the overhead.
This completely eliminates long 0 FPS periods when running the demo for PGO.
It may be possible to reduce the overhead by setting
CONFIG_SQUASHFS_FRAGMENT_CACHE_SIZE kernel setting to 1.
A PGO'd binary can be built as follows:
1. Build with `-DDEVILUTIONX_PROFILE_GENERATE=ON`.
3. Run the timedemo.
4. Build with `-DDEVILUTIONX_PROFILE_USE=ON`.
By default, the profile directory is at `${HOME}/devilutionx-profile`
Example for the RG99:
```bash
# Build the OPK for profiling data collection:
TOOLCHAIN=/opt/rs90-toolchain Packaging/OpenDingux/build.sh rg99 --profile-generate
# Copy the OPK to RG99:
scp -O build-rg99/devilutionx-rg99.opk rg99:/media/sdcard/apps
# Now, run the OPK. It will run the timedemo instead of the actual game and will take a couple of hours.
# ☕☕☕
# Copy the profiling data from RG99
scp -r -O rg99:/media/data/local/home/devilutionx-profile /tmp/devilutionx-profile
# Build the OPK use the collected profiling data:
TOOLCHAIN=/opt/rs90-toolchain Packaging/OpenDingux/build.sh rg99 --profile-use --profile-dir /tmp/devilutionx-profile
# Copy the resulting binary back to RG99
scp -O build-rg99/devilutionx-rg99.opk rg99:/media/sdcard/apps
```
Gleb Mazovetskiy