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

[Sanjoy Das via llvm-commits]

sanjoy added a comment.

Hi Pankaj,

In https://reviews.llvm.org/D22377#493355, @pankajchawla wrote:

>




> Regarding the loop entry guard check:




> Please note that I haven't changed this check and I believe the

>  original check is correct. The loop entry guard tells us whether the


Yes, for what it is doing it is correct.

> loop will be executed at least once. It cannot tell us whether the

>  loop will be executed more than once. Please also note that the

>  backedge taken count for the loop mentioned in the description is:

>  ((-1 + %n) /u %s)




> A more generic formula with unknown initial value is: ((-1 + %n -

>  %init) /u %s)

> 

> Now lets go through your examples:

> 

> > The example I was thinking of was:

> 

> > 

> 

> >   for (unsigned i = 1; i < 5; i += (-2)) {

> 

> >   }

> 

> >    

> 

> > 

> 

> > ==

> 

> > 

> 

> >   i = 1;

> 

> >   if (i < 5) {

> 

> >     do {

> 

> >       i += (-2); // NUW

> 

> >     } while (i u< 5);

> 

> >   }

> 

> >    

> 

> > 

> 

> > then the pre-increment IV is {1,+,-2}.  It does not unsigned-overflow on the first increment (1 + (-2) == -1), and then it exits the loop.

> 

> 

> The backedge taken count for this loop according to the formula above will be:

>  (5 -1 -1) /u (-2)  ==> 0

> 

> So it is correct.


I was trying to come up with examples where (based on the change
description):

- The add recurrence, AR, being compared to in the loop exit check is legitimately nuw or nsw
- The loop decrement amount (stride) is negative
- The loop is guarded by "AR->getStart() - AR->getStride()"

If I understand you correctly, you're saying that even if the stride
is negative the expression that computes the backedge taken count is
valid.  That may be correct (I'll re-review with that understanding),
but then the change description needs to say that instead of saying
"then we can assume the stride to be positive."  Is that a correct
assessment?

Btw, if you think a real-time chat can help and are in PDT, we can
talk about this on the #llvm IRC channel on Monday.

> Here's your second example-

> 

> > 

> 

> > 

> 

> >   i = INT_MIN

> 

> >   if (i s< 5) {

> 

> >     do {

> 

> >       i += (-1);

> 

> >     } while (i s< 5);

> 

> >   }

> 

> >    

> 

> 

> I believe this loop has undefined behavior since adding INT_MIN

>  and -1 leads to signed underflow.


I was using that as a shorthand for LLVM IR. :) You could have
something like:

  i = INT_MIN
  if (i s< 5) {
    do {
      i = wrapping_add(i, -1);
    } while (i s< 5);
  }

with

  int wrapping_add(int a, int b) {
    return (int)(((unsigned)a) + ((unsigned)b));
  }



> I guess my point is that once we have checked the loop entry guard, single trip loops will be handled just as well as multi-trip loops.

>  Please let me know if I am still missing anything.

> 

> Please note that the nsw marking on the IV AddRec is quite conservative. If I initialize the IV to anything else than 0 in the original loop, we lose the marking and the backedge taken count is not computed.





https://reviews.llvm.org/D22377