To work around an erratum, it looks like move elimination has been disabled in Ice Lake, at least as far back as the November microcode update (surprised nobody noticed until now).

I guess this might regress some workloads by a few %.

If anyone has TGL or any other post-ICL hardware, you could test this out by running uarch-bench like so:


./uarch-bench.sh --test-name=misc/mov-elim*
./uarch-bench.sh --timer=perf --extra-events=uops_executed.core --test-name=misc/mov-elim*


The second version uses perf and gives extra-definitive confirmation (but requires perf paranoid to be set up and rdpmc access, so sometimes is a bit trickier: try running as root).

Results like so indicate no elimination (smaller numbers indicate elimination):


** Running group misc : Miscellaneous tests **
Benchmark Cycles uops_e
8 chained moves 7.99 9.00
8 chained movs and inc 8.00 9.00