[PATCH] D89952: [SelectionDAG] Enable CTPOP optimization fine tuning

[David Zarzycki via Phabricator via llvm-commits]

davezarzycki updated this revision to Diff 303878.
davezarzycki added a comment.

Removed the comment about x86 bias now that more testing exists.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D89952/new/

https://reviews.llvm.org/D89952

Files:
  llvm/include/llvm/CodeGen/TargetLowering.h
  llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp


Index: llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -3403,15 +3403,23 @@
 
   // (ctpop x) u< 2 -> (x & x-1) == 0
   // (ctpop x) u> 1 -> (x & x-1) != 0
-  if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)) {
-    // This based on X86's custom lowering for vector CTPOP which produces more
-    // instructions than the expansion here.
+  if (Cond == ISD::SETULT || Cond == ISD::SETUGT) {
+    unsigned CostLimit = TLI.getCustomCtpopCost(CTVT, Cond);
+    if (C1.ugt(CostLimit + (Cond == ISD::SETULT)))
+      return SDValue();
+    if (C1 == 0 && (Cond == ISD::SETULT))
+      return SDValue(); // This is handled elsewhere.
+
+    unsigned Passes = C1.getLimitedValue() - (Cond == ISD::SETULT);
 
     SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT);
-    SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, CTOp, NegOne);
-    SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Add);
+    SDValue Result = CTOp;
+    for (unsigned i = 0; i < Passes; i++) {
+      SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, Result, NegOne);
+      Result = DAG.getNode(ISD::AND, dl, CTVT, Result, Add);
+    }
     ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
-    return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, dl, CTVT), CC);
+    return DAG.getSetCC(dl, VT, Result, DAG.getConstant(0, dl, CTVT), CC);
   }
 
   // If ctpop is not supported, expand a power-of-2 comparison based on it.
Index: llvm/include/llvm/CodeGen/TargetLowering.h
===================================================================
--- llvm/include/llvm/CodeGen/TargetLowering.h
+++ llvm/include/llvm/CodeGen/TargetLowering.h
@@ -605,6 +605,12 @@
     return false;
   }
 
+  /// Return the maximum number of "x & (x - 1)" operations that can be done
+  /// instead of deferring to a custom CTPOP.
+  virtual unsigned getCustomCtpopCost(EVT VT, ISD::CondCode Cond) const {
+    return 1;
+  }
+
   /// Return true if instruction generated for equality comparison is folded
   /// with instruction generated for signed comparison.
   virtual bool isEqualityCmpFoldedWithSignedCmp() const { return true; }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D89952.303878.patch
Type: text/x-patch
Size: 2294 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201109/093b8ca3/attachment.bin>