[SCEV] Apply loop guard rewrites to mul expressions. (#243622)
Also rewrite SCEVMulExpr in SCEVLoopGuardRewriter.
This mostly improves constant BTCs, leading to additional unrolling in a
number of cases end-to-end, as well as additional simplifications, like
removing minimum iteration checks created by LoopVectorize:
https://github.com/dtcxzyw/llvm-opt-benchmark-nightly/pull/2067
In a few cases (linux/scatterlist, ibwebp/pnmdec, abc/cadical_kitten)
there are minor regressions of SCEV expansions.
After recent SCEV improvements, compile-time geomean increases are in
the noise. ClamAV compile-time increases by 0.03% - 0.04%Per workload:
https://llvm-compile-time-tracker.com/compare.php?from=4d669bea8878d02f94e427f1e41014f6480e4f23&to=1b9138a3b36ce7c369f70e12d8cafce76a0ba614&stat=instructions:u
PR: https://github.com/llvm/llvm-project/pull/243622