[PATCH] D71660: [ValueTracking] isKnownNonZero() should take non-null-ness assumptions into consideration (PR43267)

[Nikita Popov via Phabricator via llvm-commits]

nikic added a comment.

> I get that. The question is, given we have AssumptionCache abstraction, why do we want to avoid it.

As this code now handles more predicates than the other code-path, I'm okay with your approach.



================
Comment at: llvm/lib/Analysis/ValueTracking.cpp:623
+        return true;
+    }
+  }
----------------
lebedev.ri wrote:
> jdoerfert wrote:
> > < is also fine, however unlikely.
> Right, https://rise4fun.com/Alive/plslNp
As you went through the effort of supporting different predicates, possibly extend this to also handle non-zero constant? That is, not just `x > 0` but also `x > 100`. Something along the lines of...

```
if (!match(Cmp, m_ICmp(Pred, m_V, m_APInt(C))))
  return false;

// Possibly without the non-canonical predicates?
switch (Pred) {
case ICmpInst::ICMP_NE:
  return C->isNullValue();
case ICmpInst::ICMP_UGT:
  return true;
case ICmpInst::ICMP_UGE:
  return !C->isNullValue();
case ICmpInst::ICMP_SGT:
  return C->isNonNegative();
case ICmpInst::ICMP_SGE:
  return C->isStrictlyPositive();
case ICmpInst::ICMP_SLT:
  return !C->isStrictlyPositive();
case ICmpInst::ICMP_SLE:
  return C->isNegative();
}
```


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D71660/new/

https://reviews.llvm.org/D71660