In exploring freebsd build server build failures and personal build failures for lang/ghc (and lang/ghc810 now) I discovered that the the process size limitations are different for armv7-on-aarch64 chroot's/jail's vs. on the likes of an Orange Pi+ 2Ed (so directly booted as armv7). The test context uses llvm12's llc on a file that fails in both contexts, allowing top to show an approximation to how bug the llc process is allowed to get:
OPi+2e: top shows SIZEs like 2489Mi somewhat before the failure. (It only has 2 GiBytes of RAM, so RES stays below that.) armv7 on Cortex-A72 system: top shows SIZEs like 2046Mi in the time frame of the failure, never showing 2048+ Mi. This explains some examples of ports not building on the aarch64 based build systems --ports that can build via using a direct armv7 boot environment. In some other cases it explains why the failure points vary for various builds. (Note: I've locally adjusted top to display more digits and be explicit about powers of 1024, among other things.) === Mark Millard marklmi at yahoo.com
