[llvm] [InstSimplify] Discard unnecessary `select` when the conditions are `ICmps` with EQ (PR #179183)

[Yingwei Zheng via llvm-commits]

================
@@ -5142,6 +5167,36 @@ static Value *simplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
       return ConstantVector::get(NewC);
   }
 
+  CmpPredicate Pred1, Pred2;
+  Value *V1, *V2, *EQV;
+  if (match(Cond,
+            m_And(m_SpecificICmp(ICmpInst::ICMP_EQ, m_Value(V1), m_Value(EQV)),
+                  m_SpecificICmp(ICmpInst::ICMP_EQ, m_Value(V2),
+                                 m_Deferred(EQV)))))
+    if (Value *V = simplifySelectWithEquivalence(
+            {{V2, EQV}, {V1, EQV}}, TrueVal, FalseVal, Q, MaxRecurse)) {
+      // We should only throw away a select if we get a select in place of that
+      // because of the poison barrier property.
+      SelectInst *VasSI = dyn_cast<SelectInst>(V);
+      // Also we have to check if the removing of the current select doesn't
+      // introduce any new poison. We should only remove the current select if
+      // its condition is poison in every case where the new select condition is
+      // poison.
+      if (VasSI && impliesPoison(VasSI->getCondition(), Cond))
----------------
dtcxzyw wrote:

I don't think the refinement check should be placed here. `simplifySelectWithEquivalence` should not refine the result, as it passes `AllowRefinement=false` when simplifying the false arm.
I guess there is something wrong with `simplifyWithOpsReplaced`.

https://github.com/llvm/llvm-project/pull/179183