[llvm] [InstCombine] Add folds for (add/sub/disjoint_or/icmp C, (ctpop (not x))) (PR #77859)

[Nikita Popov via llvm-commits]

================
@@ -740,6 +740,93 @@ static Value *tryFactorization(BinaryOperator &I, const SimplifyQuery &SQ,
   return RetVal;
 }
 
+// If `I` has one Const operand and the other matches `(ctpop (not x))`,
+// replace `(ctpop (not x))` with `(sub nuw nsw BitWidth(x), (ctpop x))`.
+// This is only useful is the new subtract can fold so we only handle the
+// following cases:
+//    1) (add/sub/disjoint_or C, (ctpop (not x))
+//        -> (add/sub/disjoint_or C', (ctpop x))
+//    1) (cmp pred C, (ctpop (not x))
+//        -> (cmp pred C', (ctpop x))
+Instruction *InstCombinerImpl::tryFoldInstWithCtpopWithNot(Instruction *I) {
+  unsigned Opc = I->getOpcode();
+  unsigned ConstIdx = 1;
+  switch (Opc) {
+  default:
+    return nullptr;
+    // (ctpop (not x)) <-> (sub nuw nsw BitWidth(x) - (ctpop x))
+    // We can fold the BitWidth(x) with add/sub/icmp as long the other operand
+    // is constant.
+  case Instruction::Sub:
+    ConstIdx = 0;
+    break;
+  case Instruction::ICmp:
+    // Signed predicates aren't correct in some edge cases like for i2 types, as
+    // well since (ctpop x) is known [0, log2(BitWidth(x))] almost all signed
+    // comparisons against it are simplfied to unsigned.
+    if (cast<ICmpInst>(I)->isSigned())
+      return nullptr;
+    break;
+  case Instruction::Or:
+    if (!match(I, m_DisjointOr(m_Value(), m_Value())))
+      return nullptr;
+    [[fallthrough]];
+  case Instruction::Add:
+    break;
+  }
+
+  Value *Op;
+  // Find ctpop.
+  if (!match(I->getOperand(1 - ConstIdx),
+             m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(Op)))))
+    return nullptr;
+
+  Constant *C;
+  // Check other operand is ImmConstant.
+  if (!match(I->getOperand(ConstIdx), m_ImmConstant(C)))
+    return nullptr;
+
+  Type *Ty = Op->getType();
+  Constant *BitWidthC = ConstantInt::get(Ty, Ty->getScalarSizeInBits());
+  // Need extra check for icmp. Note if this check is it generally means the
----------------
nikic wrote:

I think something is missing after "if this check is"?

https://github.com/llvm/llvm-project/pull/77859