[llvm] [SelectionDAG] Emit a more efficient magic number multiplication for exact udivs (PR #87161)

[Jay Foad via llvm-commits]

================
@@ -6090,6 +6089,73 @@ static SDValue BuildExactSDIV(const TargetLowering &TLI, SDNode *N,
   return DAG.getNode(ISD::MUL, dl, VT, Res, Factor);
 }
 
+/// Given an exact UDIV by a constant, create a multiplication
+/// with the multiplicative inverse of the constant.
+static SDValue BuildExactUDIV(const TargetLowering &TLI, SDNode *N,
+                              const SDLoc &dl, SelectionDAG &DAG,
+                              SmallVectorImpl<SDNode *> &Created) {
+  SDValue Op0 = N->getOperand(0);
+  SDValue Op1 = N->getOperand(1);
+  EVT VT = N->getValueType(0);
+  EVT SVT = VT.getScalarType();
+  EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
+  EVT ShSVT = ShVT.getScalarType();
+
+  bool UseSRL = false;
+  SmallVector<SDValue, 16> Shifts, Factors;
+
+  auto BuildUDIVPattern = [&](ConstantSDNode *C) {
+    if (C->isZero())
+      return false;
+    APInt Divisor = C->getAPIntValue();
+    unsigned Shift = Divisor.countr_zero();
+    if (Shift) {
+      Divisor.lshrInPlace(Shift);
+      UseSRL = true;
+    }
+    // Calculate the multiplicative inverse, using Newton's method.
----------------
jayfoad wrote:

It would be nice to use `APInt::multiplicativeInverse` here but that seems awkward because you need to pass in 2^N as the modulus which requires an N+1-bit APInt.

https://github.com/llvm/llvm-project/pull/87161