[llvm] [DAG] Lower frem of power-2 using div/trunc/mul+sub (PR #91148)

[Matt Arsenault via llvm-commits]

================
@@ -17264,17 +17264,32 @@ SDValue DAGCombiner::visitFREM(SDNode *N) {
   EVT VT = N->getValueType(0);
   SDNodeFlags Flags = N->getFlags();
   SelectionDAG::FlagInserter FlagsInserter(DAG, N);
+  SDLoc DL(N);
 
   if (SDValue R = DAG.simplifyFPBinop(N->getOpcode(), N0, N1, Flags))
     return R;
 
   // fold (frem c1, c2) -> fmod(c1,c2)
-  if (SDValue C = DAG.FoldConstantArithmetic(ISD::FREM, SDLoc(N), VT, {N0, N1}))
+  if (SDValue C = DAG.FoldConstantArithmetic(ISD::FREM, DL, VT, {N0, N1}))
     return C;
 
   if (SDValue NewSel = foldBinOpIntoSelect(N))
     return NewSel;
 
+  // Lower frem N0, N1 => x - trunc(N0 / N1) * N1, providing N1 is an integer
+  // power of 2.
+  if (DAG.isKnownToBeAPowerOfTwoFP(N1) &&
+      (Flags.hasNoSignedZeros() || DAG.isKnownNonNegativeFP(N0)) &&
+      !TLI.isOperationLegal(ISD::FREM, VT) &&
----------------
arsenm wrote:

Check the legality stuff first 

https://github.com/llvm/llvm-project/pull/91148