[llvm] [WebAssembly] Combine i128 to v16i8 for setcc & expand memcmp for 16 byte loads with simd128 (PR #149461)

[Luke Lau via llvm-commits]

================
@@ -3383,15 +3383,65 @@ static SDValue TryMatchTrue(SDNode *N, EVT VecVT, SelectionDAG &DAG) {
   return DAG.getZExtOrTrunc(Ret, DL, N->getValueType(0));
 }
 
+/// Try to convert a i128 comparison to a v16i8 comparison before type
+/// legalization splits it up into chunks
+static SDValue
+combineVectorSizedSetCCEquality(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
+                                const WebAssemblySubtarget *Subtarget) {
+
+  SDLoc DL(N);
+  SDValue X = N->getOperand(0);
+  SDValue Y = N->getOperand(1);
+  EVT VT = N->getValueType(0);
+  EVT OpVT = X.getValueType();
+
+  SelectionDAG &DAG = DCI.DAG;
+  if (DCI.DAG.getMachineFunction().getFunction().hasFnAttribute(
+          Attribute::NoImplicitFloat))
+    return SDValue();
+
+  ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
+  // We're looking for an oversized integer equality comparison with SIMD
+  if (!OpVT.isScalarInteger() || !OpVT.isByteSized() || OpVT != MVT::i128 ||
+      !Subtarget->hasSIMD128() || !isIntEqualitySetCC(CC))
+    return SDValue();
+
+  // Don't perform this combine if constructing the vector will be expensive.
+  auto IsVectorBitCastCheap = [](SDValue X) {
+    X = peekThroughBitcasts(X);
+    return isa<ConstantSDNode>(X) || X.getOpcode() == ISD::LOAD;
+  };
+
+  if (!IsVectorBitCastCheap(X) || !IsVectorBitCastCheap(Y))
+    return SDValue();
+
+  SDValue VecX = DAG.getBitcast(MVT::v16i8, X);
+  SDValue VecY = DAG.getBitcast(MVT::v16i8, Y);
+  SDValue Cmp = DAG.getSetCC(DL, MVT::v16i8, VecX, VecY, CC);
+
+  SDValue Intr =
+      DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
+                  {DAG.getConstant(CC == ISD::SETEQ ? Intrinsic::wasm_alltrue
+                                                    : Intrinsic::wasm_anytrue,
+                                   DL, MVT::i32),
+                   Cmp});
+
+  return DAG.getSetCC(DL, VT, Intr, DAG.getConstant(0, DL, MVT::i32), CC);
----------------
lukel97 wrote:

We're accidentally negating the result, this should be

```suggestion
  return DAG.getSetCC(DL, VT, Intr, DAG.getConstant(0, DL, MVT::i32), ISD::SETNE);
```

I should have caught this in review earlier, sorry! You should open up another PR that reverts the revert and include this fix in it

https://github.com/llvm/llvm-project/pull/149461