Changes to address Sanjoy's comments.

@Eli
Thanks for the info!

Fixed max backedge count computation for loops with unknown stride by assuming a min stride of 1.

Hi all,

Thanks for the additional info!

I can't find anything wrong with the IR. I am awaiting your response for more information.

Please ignore my previous email. My analysis was incorrect. The stride cannot be negative here as it will lead to UB.

Hi Sanjoy,

Added more elaborate comments explaining why the change is correct.

Added bailout condition for predicated IVs.

• Added check to prove that the stride is truly unknown to ScalarEvolution so that the wrap flags are not propagated in edge cases.
• Removed unnecessary check for loopHasNoAbnormalExits() and loop entry guard.
• Fixed comments and the test.

Hi Sanjoy,

Hi Sanjoy,

Hi Sanjoy,

Added check for side effects in the loop.

Bumping this up again.

Hi Sanjoy,

Added check for loopHasNoAbnormalExits() and a comment about bailing out for do-while loops.

In D22377#484607, @eli.friedman wrote:

You have to be a bit more careful here... consider:

for (int i=0; i<16; i-=2) {
    if (i < -2000) break;
}

Here, the backedge is taken 1000 times. (There are actually two related cases here: one, the case where there's an explicit break, and the case where some call in the loop throws an exception.)

It might be sufficient to show that the loop only has one exit and loopHasNoAbnormalExits() is true; not 100% sure about that.