[llvm] [llvm] Add KnownBits implementations for avgFloor and avgCeil (PR #86445)

[via llvm-commits]

================
@@ -762,6 +762,43 @@ KnownBits KnownBits::usub_sat(const KnownBits &LHS, const KnownBits &RHS) {
   return computeForSatAddSub(/*Add*/ false, /*Signed*/ false, LHS, RHS);
 }
 
+KnownBits KnownBits::avgFloorS(const KnownBits &LHS, const KnownBits &RHS) {
+  // (C1 & C2) + (C1 ^ C2).ashr(1)
+  KnownBits andResult = LHS & RHS;
+  KnownBits xorResult = LHS ^ RHS;
+  xorResult.Zero.ashrInPlace(1);
+  xorResult.One.ashrInPlace(1);
+  return computeForAddSub(/*Add*/ true, /*Signed*/ true, andResult, xorResult);
+}
+
+KnownBits KnownBits::avgFloorU(const KnownBits &LHS, const KnownBits &RHS) {
+  // (C1 & C2) + (C1 ^ C2).lshr(1)
+  KnownBits andResult = LHS & RHS;
+  KnownBits xorResult = LHS ^ RHS;
+  xorResult.Zero.lshrInPlace(1);
+  xorResult.One.lshrInPlace(1);
+  return computeForAddSub(/*Add*/ true, /*Signed*/ false, andResult, xorResult);
+}
+
+KnownBits KnownBits::avgCeilS(const KnownBits &LHS, const KnownBits &RHS) {
+  // (C1 | C2) - (C1 ^ C2).ashr(1)
+  KnownBits andResult = LHS & RHS;
+  KnownBits xorResult = LHS ^ RHS;
+  xorResult.Zero.ashrInPlace(1);
+  xorResult.One.ashrInPlace(1);
+  return computeForAddSub(/*Add*/ false, /*Signed*/ true, andResult, xorResult);
+}
+
+KnownBits KnownBits::avgCeilU(const KnownBits &LHS, const KnownBits &RHS) {
+  // (C1 | C2) - (C1 ^ C2).lshr(1)
+  KnownBits andResult = LHS & RHS;
+  KnownBits xorResult = LHS ^ RHS;
+  xorResult.Zero.lshrInPlace(1);
+  xorResult.One.lshrInPlace(1);
+  return computeForAddSub(/*Add*/ false, /*Signed*/ false, andResult,
+                          xorResult);
----------------
goldsteinn wrote:

`nuw` means that when we do `X + Y` it won't have unsigned overflow. So `X nuw+ Y >= X && X nuw+ Y >= Y` must hold. `nsw` is the same but for the signbit.
So with 8bit, `128 + 128` violates both nuw/nsw. 

https://github.com/llvm/llvm-project/pull/86445