[llvm] [InstCombine] Simplify and/or of icmp eq with op replacement (PR #70335)

[Nikita Popov via llvm-commits]

================
@@ -2025,6 +2025,52 @@ static Value *simplifyAndOrOfCmps(const SimplifyQuery &Q, Value *Op0,
   return nullptr;
 }
 
+static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
+                                     const SimplifyQuery &Q,
+                                     bool AllowRefinement,
+                                     SmallVectorImpl<Instruction *> *DropFlags,
+                                     unsigned MaxRecurse);
+
+static Value *simplifyAndOrWithICmpEq(unsigned Opcode, Value *Op0, Value *Op1,
+                                      const SimplifyQuery &Q,
+                                      unsigned MaxRecurse) {
+  assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
+         "Must be and/or");
+  ICmpInst::Predicate Pred;
+  Value *A, *B;
+  if (!match(Op0, m_ICmp(Pred, m_Value(A), m_Value(B))) ||
+      !ICmpInst::isEquality(Pred) || !MaxRecurse--)
+    return nullptr;
+
+  auto Simplify = [&](Value *Res) -> Value * {
+    // and (icmp eq a, b), x implies (a==b) inside x.
+    // or (icmp ne a, b), x implies (a==b) inside x.
----------------
nikic wrote:

If we had a way to say "simplify x under the assumption that a != b", I would expect that to be beneficial. But we can only do this for equalities right now, as in that case we can just replace operands.

I think we need some generic way to say "evaluate this predicate under the assumption that v has range CR" -- some of @dtcxzyw's recent PRs go in that direction, but I feel like there must be a more generic solution to this, which integrates with existing range propagation logic, instead of reimplementing it in parts. This seems like something that LVI/CVP should be doing, but I don't see a good way to integrate it there.

https://github.com/llvm/llvm-project/pull/70335