[llvm] [SelectionDAG] Optimize 32-bit udiv with 33-bit magic constants on 64-bit targets (PR #181288)

[MITSUNARI Shigeo via llvm-commits]

================
@@ -6864,6 +6894,46 @@ SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
     PostShift = PostShifts[0];
   }
 
+  // Our Approach: Use optimized 33-bit method for 32-bit division
+  if (Use33BitOptimization) {
+    // x is i32, MagicFactor is pre-shifted i64 constant
+    // Compute: (i64(x) * MagicFactor) >> 64
+    EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), 64);
+    SDValue X64 = DAG.getNode(ISD::ZERO_EXTEND, dl, WideVT, N0);
+
+    // Get the pre-shifted constant (it's already in MagicFactor as i64)
+    SDValue MagicFactor64 = isa<ConstantSDNode>(MagicFactor)
+        ? MagicFactor
+        : MagicFactors[0];
+
+    SDValue Result;
+    // Perform 64x64 -> 128 multiplication and extract high 64 bits
+    if (isOperationLegalOrCustom(ISD::MULHU, WideVT, IsAfterLegalization)) {
+      SDValue High = DAG.getNode(ISD::MULHU, dl, WideVT, X64, MagicFactor64);
+      Created.push_back(High.getNode());
+      // Truncate back to i32
+      Result = DAG.getNode(ISD::TRUNCATE, dl, VT, High);
+    } else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, WideVT, IsAfterLegalization)) {
+      SDValue LoHi = DAG.getNode(ISD::UMUL_LOHI, dl,
+                                  DAG.getVTList(WideVT, WideVT),
+                                  X64, MagicFactor64);
+      SDValue High = SDValue(LoHi.getNode(), 1);
+      Created.push_back(LoHi.getNode());
+      Result = DAG.getNode(ISD::TRUNCATE, dl, VT, High);
+    } else {
+      // Fallback to standard path if 64-bit MULHU is not available
----------------
herumi wrote:

Sorry, I forgot to remove the debugging code. I'll remove it.

https://github.com/llvm/llvm-project/pull/181288