[llvm] [SelectionDAG] Improve ucmp/scmp codegen (PR #190787)

[Takashi Idobe via llvm-commits]

================
@@ -11479,6 +11479,40 @@ SDValue TargetLowering::expandCMP(SDNode *Node, SelectionDAG &DAG) const {
   EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
   SDLoc dl(Node);
 
+  // fold scmp(x, 0) -> or(sra(x, bw-1), zext(setne(x, 0)))
+  // This is cheaper than the generic sub(IsGT, IsLT) expansion for wide
+  // integers: the sign mask comes from a single arithmetic shift and the
+  // nonzero test is a single comparison, with no flag-chain dependencies
+  // between them. Only apply when VT is a non-native integer type.
+  if (Opcode == ISD::SCMP && isNullConstant(RHS) && VT.isScalarInteger() &&
+      getTypeAction(*DAG.getContext(), VT) == TypeExpandInteger) {
+    SDValue ShiftAmt =
+        DAG.getShiftAmountConstant(VT.getScalarSizeInBits() - 1, VT, dl);
+    SDValue SignMask = DAG.getNode(ISD::SRA, dl, VT, LHS, ShiftAmt);
+    SDValue IsNZ = DAG.getSetCC(dl, BoolVT, LHS, RHS, ISD::SETNE);
+    SDValue IsNZExt = DAG.getZExtOrTrunc(IsNZ, dl, VT);
+    if (BoolVT.getScalarSizeInBits() != 1 &&
+        getBooleanContents(BoolVT) != ZeroOrOneBooleanContent)
+      IsNZExt =
+          DAG.getNode(ISD::AND, dl, VT, IsNZExt, DAG.getConstant(1, dl, VT));
+    SDValue Result = DAG.getNode(ISD::OR, dl, VT, SignMask, IsNZExt);
+    return DAG.getSExtOrTrunc(Result, dl, ResVT);
+  }
+
+  // fold ucmp(x, 0) -> zext(setne(x, 0))
+  // Since 0 is the minimum unsigned value, ucmp(x, 0) is never -1,
+  // returning 0 when x == 0 and 1 otherwise. This avoids the generic
+  // sub(IsGT, IsLT) expansion and its flag-chain dependencies.
----------------
Takashiidobe wrote:

I'll leave this up to you whether or not you want to see the change in FoldSetCC, it'll change a lot more tests so it's more code review burden but it does seem like a better place than for ucmp.

https://github.com/llvm/llvm-project/pull/190787