Apple's open-source XNU kernel on Android phones: its own loader and drivers, Metal on the Adreno. Runs from RAM; a reboot gives the phone back to Android.
C
mk-mesa: this build is not reproducible, so a hash cannot verify it
vcvkkd2e09bf85202
diff --git a/tools/mk-mesa.sh b/tools/mk-mesa.shindex 4e15eaf..1cbe463 100755--- a/tools/mk-mesa.sh+++ b/tools/mk-mesa.sh@@ -20,6 +20,19 @@# gl.xml their prototypes (the Khronos registry as Mesa has it): for userland/mkeagl.py# Needs on this machine: meson, ninja, bison, flex, python3 with venv (Mako is installed into a# private environment), llvm-ar.+#+# This build is NOT reproducible byte for byte, so do not compare one of its outputs with another+# by hash. Measured on 20 October: two builds from the same source tree, the second with+# build-softpipe-freedreno deleted first, gave different libOSMesa.8.dylib and air_run -- same size+# to the byte, 9 132 774 of 14 855 048 bytes different in the library and 4 171 752 of 9 789 824 in+# air_run, with 5 of the 6 named text symbols at new addresses. So the layout moves between builds.+# What moves it was not established; it is not source filenames, which -Db_ndebug=true compiles out.+#+# The practical consequence: a hash answers whether this is the same FILE, never whether it holds+# the same CODE, and the size answers neither -- the one-line ir3 rounding patch+# (tools/mesa/ir3-round-default.patch) produced a library of exactly the same 14 855 048 bytes as+# the build without it. To tell whether a patch is in a build, read the source in+# xnu-deps/build/mesa/mesa-$VERSION, which is where this script unpacks and patches it.set -eVERSION=25.0.7SHA256=592272df3cf01e85e7db300c449df5061092574d099da275d19e97ef0510f8a6