[llvm] [InstCombine] Fold zext(X) + C2 pred C -> X + C3 pred C4 (PR #110511)

[Nikita Popov via llvm-commits]

================
@@ -3169,6 +3169,31 @@ Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
                         Builder.CreateAdd(X, ConstantInt::get(Ty, *C2 - C - 1)),
                         ConstantInt::get(Ty, ~C));
 
+  // zext(V) + C2 pred C -> V + C3 pred' C4
+  Value *V;
+  if (match(X, m_ZExt(m_Value(V)))) {
+    Type *NewCmpTy = V->getType();
+    unsigned CmpBW = Ty->getScalarSizeInBits();
+    unsigned NewCmpBW = NewCmpTy->getScalarSizeInBits();
+    if (shouldChangeType(Ty, NewCmpTy)) {
+      if (auto ZExtCR = CR.exactIntersectWith(ConstantRange(
+              APInt::getZero(CmpBW), APInt::getOneBitSet(CmpBW, NewCmpBW)))) {
----------------
nikic wrote:

Can this condition be written as `CR.getActiveBits() <= NewCmpBW`? The exact intersection here seems pretty roundabout.

https://github.com/llvm/llvm-project/pull/110511