[PATCH] D44102: Teach CorrelatedValuePropagation to reduce the width of udiv/urem instructions.

[Sanjay Patel via Phabricator via llvm-commits]

spatel added reviewers: anna, davide, reames.
spatel added a comment.

I think this is the right approach, but I don't know much about CVP, so adding more potential reviewers.

Context: This is part of improving udiv/urem IR as discussed in the latest comments on https://reviews.llvm.org/D37121. 
Motivation: Narrowing the width of these ops improves bit-tracking / analysis and potentially enables further IR transforms. 
Bonus: It also aligns with codegen transforms such as the magic number division mentioned by Justin and improves perf for targets that have faster narrow div/rem instructions.

We have basic div/rem narrowing folds in instcombine, but we want to handle cases like this where edge info and/or computeKnownBits would also allow narrowing. Is it preferred to use ValueTracking here to get the single BB case or should that be added to InstCombine?


https://reviews.llvm.org/D44102