[X86][AVX] Pad small shuffle inputs in combineX86ShufflesRecursively

As detailed on PR45974 and D79987, getFauxShuffleMask is creating nodes on the fly to create shuffles with inputs the same size as the result, causing problems for hasOneUse() checks in later simplification stages.

Currently only combineX86ShufflesRecursively benefits from these widened inputs so I've begun moving the functionality there, and out of getFauxShuffleMask. This allows us to remove the widening from VBROADCAST and *EXTEND* faux shuffle cases.

This just leaves the INSERT_SUBVECTOR case in getFauxShuffleMask still creating nodes, which will require more extensive refactoring.

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -7662,20 +7662,8 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
   }
   case X86ISD::VBROADCAST: {
     SDValue Src = N.getOperand(0);
-    MVT SrcVT = Src.getSimpleValueType();
-    if (!SrcVT.isVector())
+    if (!Src.getSimpleValueType().isVector())
       return false;
-
-    if (NumSizeInBits != SrcVT.getSizeInBits()) {
-      assert((NumSizeInBits % SrcVT.getSizeInBits()) == 0 &&
-             "Illegal broadcast type");
-      SrcVT = MVT::getVectorVT(SrcVT.getScalarType(),
-                               NumSizeInBits / SrcVT.getScalarSizeInBits());
-      Src = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), SrcVT,
-                        DAG.getUNDEF(SrcVT), Src,
-                        DAG.getIntPtrConstant(0, SDLoc(N)));
-    }
-
     Ops.push_back(Src);
     Mask.append(NumElts, 0);
     return true;
@@ -7692,22 +7680,10 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
         (SrcVT.getScalarSizeInBits() % 8) != 0)
       return false;
 
-    unsigned NumSrcBitsPerElt = SrcVT.getScalarSizeInBits();
     bool IsAnyExtend =
         (ISD::ANY_EXTEND == Opcode || ISD::ANY_EXTEND_VECTOR_INREG == Opcode);
-    DecodeZeroExtendMask(NumSrcBitsPerElt, NumBitsPerElt, NumElts, IsAnyExtend,
-                         Mask);
-
-    if (NumSizeInBits != SrcVT.getSizeInBits()) {
-      assert((NumSizeInBits % SrcVT.getSizeInBits()) == 0 &&
-             "Illegal zero-extension type");
-      SrcVT = MVT::getVectorVT(SrcVT.getSimpleVT().getScalarType(),
-                               NumSizeInBits / NumSrcBitsPerElt);
-      Src = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), SrcVT,
-                        DAG.getUNDEF(SrcVT), Src,
-                        DAG.getIntPtrConstant(0, SDLoc(N)));
-    }
-
+    DecodeZeroExtendMask(SrcVT.getScalarSizeInBits(), NumBitsPerElt, NumElts,
+                         IsAnyExtend, Mask);
     Ops.push_back(Src);
     return true;
   }
@@ -35054,7 +35030,8 @@ static SDValue combineX86ShufflesRecursively(
 
   assert(Root.getSimpleValueType().isVector() &&
          "Shuffles operate on vector types!");
-  assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
+  unsigned RootSizeInBits = Root.getSimpleValueType().getSizeInBits();
+  assert(VT.getSizeInBits() == RootSizeInBits &&
          "Can only combine shuffles of the same vector register size.");
 
   // Extract target shuffle mask and resolve sentinels and inputs.
@@ -35068,12 +35045,17 @@ static SDValue combineX86ShufflesRecursively(
                               OpZero, DAG, Depth, false))
     return SDValue();
 
-  // Shuffle inputs must be the same size as the result.
-  if (llvm::any_of(OpInputs, [VT](SDValue Op) {
-        return VT.getSizeInBits() != Op.getValueSizeInBits();
+  // Shuffle inputs must be the same size as the result, bail on any larger
+  // inputs and widen any smaller inputs.
+  if (llvm::any_of(OpInputs, [RootSizeInBits](SDValue Op) {
+        return Op.getValueSizeInBits() > RootSizeInBits;
       }))
     return SDValue();
 
+  for (SDValue &Op : OpInputs)
+    if (Op.getValueSizeInBits() < RootSizeInBits)
+      Op = widenSubVector(Op, false, Subtarget, DAG, SDLoc(Op), RootSizeInBits);
+
   SmallVector<int, 64> Mask;
   SmallVector<SDValue, 16> Ops;