[llvm] [InstCombine] Fold ((X << nuw Z) binop nuw Y) >>u Z --> X binop nuw (Y >>u Z) (PR #88193)

[via llvm-commits]

================
@@ -1259,6 +1259,53 @@ Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
       match(Op1, m_SpecificIntAllowPoison(BitWidth - 1)))
     return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
 
+  // ((X << nuw Z) sub nuw Y) >>u exact Z --> X sub nuw (Y >>u exact Z),
+  // ONLY if I is exact, and both the shift and sub are nuw
+  Value *Y;
+  if (I.isExact() &&
+      match(Op0, m_OneUse(m_NUWSub(m_NUWShl(m_Value(X), m_Specific(Op1)),
+                                   m_Value(Y))))) {
+    Value *NewLshr = Builder.CreateLShr(Y, Op1, "", /* isExact */ true);
+    auto *NewSub = BinaryOperator::CreateNUWSub(X, NewLshr);
+    NewSub->setHasNoSignedWrap(
+        cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap());
+    return NewSub;
+  }
+
+  auto isSuitableBinOpcode = [](Instruction::BinaryOps BinOpcode) {
+    switch (BinOpcode) {
+    default:
+      return false;
+    case Instruction::Add:
+    case Instruction::Or:
+    case Instruction::Xor:
+      // And does not work here, and sub is handled separately.
+      return true;
+    }
+  };
+
+  // If both the binop and the shift are nuw, then:
+  // ((X << nuw Z) binop nuw Y) >>u Z --> X binop nuw (Y >>u Z)
----------------
goldsteinn wrote:

Although you can handle `And`, you just cant propagate `exact`: https://alive2.llvm.org/ce/z/JNaAvP

https://github.com/llvm/llvm-project/pull/88193