[llvm] [SLP]Attempt to vectorize long stores, if short one failed. (PR #88563)

[Nikita Popov via llvm-commits]

nikic wrote:

I've reverted the new version again -- it looks like the pairlocalalign regressions have been fixed, but all the other ones are still there (https://llvm-compile-time-tracker.com/compare.php?from=693a458287d019c5c6a66fe3019d099df2978cdb&to=6f7160eedb2db02f37d4ffd52fff7b0cf88b3fdc&stat=instructions:u)

For example libclamav_nsis_LZMADecode.c from clamscan has a 5-6% regression, transform8x8.c from lencod as well, Ppmd7.c from 7zip has 7%.

https://github.com/llvm/llvm-project/pull/88563