[PATCH] D22377: [SCEV] trip count calculation for loops with unknown stride

[Pankaj Chawla via llvm-commits]

pankajchawla added a comment.

F2275185: indvars.ll <https://reviews.llvm.org/F2275185>

Hi Sanjoy,

This seems to be an issue with indvars pass.

I have uploaded a smaller version of the reproducer (extracted from the original). The bug can be reproduced using this command-

$ opt -indvars indvars.ll -S

Before indvars, the loop for.body's IV is like this-

  for.body:                                         ; preds = %for.body.lr.ph, %for.body
    %index_Q16.0928 = phi i32 [ 0, %for.body.lr.ph ], [ %add187, %for.body ]
    ...
    %add187 = add nsw i32 %index_Q16.0928, %index_increment_Q16
    %cmp = icmp slt i32 %add187, %max_index_Q16
    br i1 %cmp, label %for.body, label %sw.epilog.loopexit

The SCEV form of post-incremented IV %add187 is this:

  {%index_increment_Q16,+,%index_increment_Q16}<nsw><%for.body>

The backedge taken count of the loop is like this:

  ((-1 + %max_index_Q16) /u %index_increment_Q16)
  


After indvars, the IV is extended to 64 bits but the problem is that the stride is zero-extended instead of getting sign-extended. This is incorrect as the stride can be negative. Here's the IR after modification-

  for.body.lr.ph:                                   ; preds = %for.cond.preheader
  %0 = zext i32 %index_increment_Q16 to i64
   br label %for.body
  
  for.body:
    %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.lr.ph ]
    ...
    %indvars.iv.next = add nuw nsw i64 %indvars.iv, %0
    br i1 %cmp, label %for.body, label %sw.epilog.loopexit

I am not familiar with indvars code. Can you please help?

Thanks,
Pankaj


Repository:
  rL LLVM

https://reviews.llvm.org/D22377