[PATCH] D46760: [InstCombine] Enhance narrowUDivURem.

[Bixia Zheng via Phabricator via llvm-commits]

bixia added a comment.

Instcombine has udiv/urem narrowing without increasing the total number of ZExt instructions. Correlated-Propagation has udiv/urem narrowing that can increasing the total number of ZExt instruction. These are existing duplicated functionalities.  I have considered two possible changes to this: (1) allow instcombine to increase total number of ZExt instructions (by simplify removing the m_oneuse check), this will make the same transformation in correlated-propagation redundant (can be removed). (2) as shown in this change, allow instcombine to handle one more case without increasing the total number of ZExt instructions. 
I looked at other similar narrowing transformations in instcombine, and found most of them have the check to avoid increasing the total number of ZExt/SExt instructions (lshr narrowing for example), but some of them don't have such a check (shl narrowing for example). I am not sure whether these are by design or due to something else such as ignorance.  Do you know about this?

Now you are proposing a third solution: (3) Add some more duplicated functionalities between instcombine and correlated-propagation, that is to make them do lshr-narrowing and and-narrowing to make my original motivated test case "work" (see my Note below}.
Can you explain why you believe approach (3) is the best among the three solutions I described?

Note:  for my motivated test case, without my change here, instcombine can't narrow these code mostly due to the fact that  "tmp2" have two uses:

  %tmp2 = zext i32 %1 to i64
  %lane_id = and i64 %tmp2, 31
  %2 = lshr i64 %tmp2, 2
   


Repository:
  rL LLVM

https://reviews.llvm.org/D46760