[PATCH] Reassociate GEP operands for loop invariant code motion

[Sanjoy Das]

This looks like it is sharing intent with the loop strength reduction pass.  I don't know if that is intentional, just pointing it out.

E.g. `-loop-reduce` does this to `gep_with_const`:

  define void @gep_with_const(i32* %input, i32* %output, i64 %a) {
  entry:
    %0 = add i64 %a, 1234
    %scevgep = getelementptr i32, i32* %input, i64 %0
    br label %loop
  
  loop:                                             ; preds = %loop, %entry
    %lsr.iv1 = phi i32* [ %scevgep2, %loop ], [ %scevgep, %entry ]
    %lsr.iv = phi i64 [ %lsr.iv.next, %loop ], [ 1000000, %entry ]
    %1 = load i32, i32* %lsr.iv1
    store i32 %1, i32* %output
    %lsr.iv.next = add nsw i64 %lsr.iv, -1
    %scevgep2 = getelementptr i32, i32* %lsr.iv1, i64 1
    %exitcond = icmp eq i64 %lsr.iv.next, 0
    br i1 %exitcond, label %loop.exit, label %loop
  
  loop.exit:                                        ; preds = %loop
    ret void
  }


REPOSITORY
  rL LLVM

http://reviews.llvm.org/D9136

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/