[llvm] [InstCombine] Generalize folds for inversion of icmp operands (PR #74317)

[Yingwei Zheng via llvm-commits]

================
@@ -7037,32 +7037,18 @@ Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
       return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(X, Y),
                           Op1);
 
-    // Transform (~X ^ Y) s< ~Z  --> (X ^ Y) s> Z,
-    //           (~X ^ Y) s> ~Z  --> (X ^ Y) s< Z,
-    //           (~X ^ Y) s<= ~Z --> (X ^ Y) s>= Z,
-    //           (~X ^ Y) s>= ~Z --> (X ^ Y) s<= Z,
-    //           (~X ^ Y) u< ~Z  --> (X ^ Y) u< Z,
-    //           (~X ^ Y) u> ~Z  --> (X ^ Y) u< Z,
-    //           (~X ^ Y) u<= ~Z --> (X ^ Y) u>= Z,
-    //           (~X ^ Y) u>= ~Z --> (X ^ Y) u<= Z,
-    //           (~X ^ Y) == ~Z  --> (X ^ Y) == Z,
-    // and       (~X ^ Y) != ~Z  --> (X ^ Y) != Z,
-    if (match(&I, m_c_ICmp(Pred, m_c_Xor(m_Not(m_Value(X)), m_Value(Y)),
-                           m_Not(m_Value(Z)))) &&
-        (I.getOperand(0)->hasOneUse() || I.getOperand(1)->hasOneUse()))
-      return new ICmpInst(I.getSwappedPredicate(Pred), Builder.CreateXor(X, Y),
-                          Z);
-
-    // ~X < ~Y --> Y < X
-    // ~X < C -->  X > ~C
-    if (match(Op0, m_Not(m_Value(X)))) {
-      if (match(Op1, m_Not(m_Value(Y))))
-        return new ICmpInst(I.getPredicate(), Y, X);
-
-      const APInt *C;
-      if (match(Op1, m_APInt(C)))
-        return new ICmpInst(I.getSwappedPredicate(), X,
-                            ConstantInt::get(Op1->getType(), ~(*C)));
+    // Op0 pred Op1 -> ~Op1 pred ~Op0, if this allows us to drop an instruction.
----------------
dtcxzyw wrote:

Can we remove the logic here?

https://github.com/llvm/llvm-project/blob/cdc37325669c0321328a7245083c427b229e79e9/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp#L3237-L3250



https://github.com/llvm/llvm-project/pull/74317