[X86] combineConcatVectorOps - recursively call combineConcatVectorOps instead of predicting when ops will freely concat (#130275)

The IsConcatFree helper is limited to estimates on where concatenating the subvector operands is beneficial, this patch replaces FADD/FSUB/FMUL concatenation checks with a recursive call to combineConcatVectorOps to see if it will profitably concatenate further up the chain.

Other opcodes can be moved to using the CombineSubOperand helper in future patches.

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -41903,7 +41903,8 @@ static SDValue canonicalizeLaneShuffleWithRepeatedOps(SDValue V,
 static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
                                       ArrayRef<SDValue> Ops, SelectionDAG &DAG,
                                       TargetLowering::DAGCombinerInfo &DCI,
-                                      const X86Subtarget &Subtarget);
+                                      const X86Subtarget &Subtarget,
+                                      unsigned Depth = 0);
 
 /// Try to combine x86 target specific shuffles.
 static SDValue combineTargetShuffle(SDValue N, const SDLoc &DL,
@@ -57791,7 +57792,8 @@ CastIntSETCCtoFP(MVT VT, ISD::CondCode CC, unsigned NumSignificantBitsLHS,
 static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
                                       ArrayRef<SDValue> Ops, SelectionDAG &DAG,
                                       TargetLowering::DAGCombinerInfo &DCI,
-                                      const X86Subtarget &Subtarget) {
+                                      const X86Subtarget &Subtarget,
+                                      unsigned Depth) {
   assert(Subtarget.hasAVX() && "AVX assumed for concat_vectors");
   unsigned EltSizeInBits = VT.getScalarSizeInBits();
 
@@ -57803,6 +57805,9 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
       }))
     return getZeroVector(VT, Subtarget, DAG, DL);
 
+  if (Depth >= SelectionDAG::MaxRecursionDepth)
+    return SDValue(); // Limit search depth.
+
   SDValue Op0 = Ops[0];
   bool IsSplat = llvm::all_equal(Ops);
   unsigned NumOps = Ops.size();
@@ -57933,6 +57938,20 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
       }
       return AllConstants || AllSubs;
     };
+    auto CombineSubOperand = [&](MVT VT, ArrayRef<SDValue> SubOps, unsigned I) {
+      bool AllConstants = true;
+      SmallVector<SDValue> Subs;
+      for (SDValue SubOp : SubOps) {
+        SDValue BC = peekThroughBitcasts(SubOp.getOperand(I));
+        AllConstants &= ISD::isBuildVectorOfConstantSDNodes(BC.getNode()) ||
+                        ISD::isBuildVectorOfConstantFPSDNodes(BC.getNode());
+        Subs.push_back(SubOp.getOperand(I));
+      }
+      if (AllConstants)
+        return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Subs);
+      return combineConcatVectorOps(DL, VT, Subs, DAG, DCI, Subtarget,
+                                    Depth + 1);
+    };
 
     switch (Op0.getOpcode()) {
     case ISD::VECTOR_SHUFFLE: {
@@ -58354,14 +58373,17 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
     case ISD::FADD:
     case ISD::FSUB:
     case ISD::FMUL:
-      if (!IsSplat && (IsConcatFree(VT, Ops, 0) || IsConcatFree(VT, Ops, 1)) &&
-          (VT.is256BitVector() ||
-           (VT.is512BitVector() && Subtarget.useAVX512Regs()))) {
-        return DAG.getNode(Op0.getOpcode(), DL, VT,
-                           ConcatSubOperand(VT, Ops, 0),
-                           ConcatSubOperand(VT, Ops, 1));
+      if (!IsSplat && (VT.is256BitVector() ||
+                       (VT.is512BitVector() && Subtarget.useAVX512Regs()))) {
+        SDValue Concat0 = CombineSubOperand(VT, Ops, 0);
+        SDValue Concat1 = CombineSubOperand(VT, Ops, 1);
+        if (Concat0 || Concat1)
+          return DAG.getNode(Op0.getOpcode(), DL, VT,
+                             Concat0 ? Concat0 : ConcatSubOperand(VT, Ops, 0),
+                             Concat1 ? Concat1 : ConcatSubOperand(VT, Ops, 1));
       }
       break;
+    // Always prefer to concatenate high latency FDIV instructions.
     case ISD::FDIV:
       if (!IsSplat && (VT.is256BitVector() ||
                        (VT.is512BitVector() && Subtarget.useAVX512Regs()))) {