Yet another possibility is a call variant that implicitly discards all the belt. If there were any belt operands actually live over the call then they would have to be explicitly spilled. While loops generally have loop-carried live operands, it’s fairly common in open code for a call to consume everything live. Worth doing? No idea; measurement needed.
Reply To: spiller work optimisation Ivan Godard 2016-05-03T12:39:16-07:00