On Thu, May 20, 2021 at 08:37:24PM +0200, Stefan Schulze Frielinghaus wrote: [...] > > but we won't ever arrive here because of the niters condition. But > > yes, doing the pattern matching in the innermost loop processing code > > looks good to me - for the specific case it would be > > > > /* Don't distribute loop if niters is unknown. */ > > tree niters = number_of_latch_executions (loop); > > if (niters == NULL_TREE || niters == chrec_dont_know) > > ---> here? > > continue; > > Right, please find attached a new version of the patch where everything > is included in the loop distribution pass. I will do a bootstrap and > regtest on IBM Z over night. If you give me green light I will also do > the same on x86_64.
Meanwhile I gave it a shot on x86_64 where the testsuite runs fine (at least the ldist-strlen testcase). If you are Ok with the patch, then I would rebase and run the testsuites again and post a patch series including the rawmemchr implementation for IBM Z. Cheers, Stefan
