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

[Jay Foad via llvm-commits]

================
@@ -5071,8 +5071,36 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
   LLT ScalarShiftAmtTy = ShiftAmtTy.getScalarType();
   auto &MIB = Builder;
 
+  bool UseSRL = false;
   bool UseNPQ = false;
   SmallVector<Register, 16> PreShifts, PostShifts, MagicFactors, NPQFactors;
+  SmallVector<Register, 16> Shifts, Factors;
+  auto *RHSDefInstr = cast<GenericMachineInstr>(getDefIgnoringCopies(RHS, MRI));
+  bool IsSplat = getIConstantSplatVal(*RHSDefInstr, MRI).has_value();
+
+  auto BuildExactUDIVPattern = [&](const Constant *C) {
+    // Don't recompute inverses for each splat element.
+    if (IsSplat && !Factors.empty()) {
+      Shifts.push_back(Shifts[0]);
+      Factors.push_back(Factors[0]);
+      return true;
+    }
+
+    auto *CI = cast<ConstantInt>(C);
+    APInt Divisor = CI->getValue();
+    unsigned Shift = Divisor.countr_zero();
+    if (Shift) {
+      Divisor.lshrInPlace(Shift);
+      UseSRL = true;
+    }
+
+    // Calculate the multiplicative inverse modulo BW.
+    // 2^W requires W + 1 bits, so we have to extend and then truncate.
----------------
jayfoad wrote:

Remove the comment about W + 1 bits.

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