[llvm] [KnownBits] Make `avg{Ceil,Floor}S` and `{s,u}{add,sub}_sat` optimal (PR #110329)

[Yingwei Zheng via llvm-commits]

================
@@ -766,12 +775,44 @@ KnownBits KnownBits::usub_sat(const KnownBits &LHS, const KnownBits &RHS) {
 static KnownBits avgCompute(KnownBits LHS, KnownBits RHS, bool IsCeil,
                             bool IsSigned) {
   unsigned BitWidth = LHS.getBitWidth();
-  LHS = IsSigned ? LHS.sext(BitWidth + 1) : LHS.zext(BitWidth + 1);
-  RHS = IsSigned ? RHS.sext(BitWidth + 1) : RHS.zext(BitWidth + 1);
-  LHS =
-      computeForAddCarry(LHS, RHS, /*CarryZero*/ !IsCeil, /*CarryOne*/ IsCeil);
-  LHS = LHS.extractBits(BitWidth, 1);
-  return LHS;
+  KnownBits ExtLHS = IsSigned ? LHS.sext(BitWidth + 1) : LHS.zext(BitWidth + 1);
+  KnownBits ExtRHS = IsSigned ? RHS.sext(BitWidth + 1) : RHS.zext(BitWidth + 1);
+  KnownBits Res = computeForAddCarry(ExtLHS, ExtRHS, /*CarryZero*/ !IsCeil,
+                                     /*CarryOne*/ IsCeil);
+  Res = Res.extractBits(BitWidth, 1);
+
+  auto SignBitKnown = [BitWidth](KnownBits KB) {
----------------
dtcxzyw wrote:

We already have `KnownBits::isSignUnknown`.


https://github.com/llvm/llvm-project/pull/110329