[AArch64] Expand vector ops when NEON and SVE are unavailable. (#90833)

Unlike `+noneon` we must assume that vector types are available, i.e.
it is valid to pass/return vector arguments to and from functions.
However, the compiler must make sure to scalarize any vector
operations.

Author: sdesmalen-arm

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -360,24 +360,24 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
   if (Subtarget->hasNEON()) {
     addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
     addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
-    // Someone set us up the NEON.
-    addDRTypeForNEON(MVT::v2f32);
-    addDRTypeForNEON(MVT::v8i8);
-    addDRTypeForNEON(MVT::v4i16);
-    addDRTypeForNEON(MVT::v2i32);
-    addDRTypeForNEON(MVT::v1i64);
-    addDRTypeForNEON(MVT::v1f64);
-    addDRTypeForNEON(MVT::v4f16);
-    addDRTypeForNEON(MVT::v4bf16);
-
-    addQRTypeForNEON(MVT::v4f32);
-    addQRTypeForNEON(MVT::v2f64);
-    addQRTypeForNEON(MVT::v16i8);
-    addQRTypeForNEON(MVT::v8i16);
-    addQRTypeForNEON(MVT::v4i32);
-    addQRTypeForNEON(MVT::v2i64);
-    addQRTypeForNEON(MVT::v8f16);
-    addQRTypeForNEON(MVT::v8bf16);
+
+    addDRType(MVT::v2f32);
+    addDRType(MVT::v8i8);
+    addDRType(MVT::v4i16);
+    addDRType(MVT::v2i32);
+    addDRType(MVT::v1i64);
+    addDRType(MVT::v1f64);
+    addDRType(MVT::v4f16);
+    addDRType(MVT::v4bf16);
+
+    addQRType(MVT::v4f32);
+    addQRType(MVT::v2f64);
+    addQRType(MVT::v16i8);
+    addQRType(MVT::v8i16);
+    addQRType(MVT::v4i32);
+    addQRType(MVT::v2i64);
+    addQRType(MVT::v8f16);
+    addQRType(MVT::v8bf16);
   }
 
   if (Subtarget->hasSVEorSME()) {
@@ -1125,7 +1125,7 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
 
   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
 
-  if (Subtarget->hasNEON()) {
+  if (Subtarget->isNeonAvailable()) {
     // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
     // silliness like this:
     for (auto Op :
@@ -1337,6 +1337,24 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
     // FADDP custom lowering
     for (MVT VT : { MVT::v16f16, MVT::v8f32, MVT::v4f64 })
       setOperationAction(ISD::FADD, VT, Custom);
+  } else /* !isNeonAvailable */ {
+    for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
+      for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
+        setOperationAction(Op, VT, Expand);
+
+      if (VT.is128BitVector() || VT.is64BitVector()) {
+        setOperationAction(ISD::LOAD, VT, Legal);
+        setOperationAction(ISD::STORE, VT, Legal);
+        setOperationAction(ISD::BITCAST, VT,
+                           Subtarget->isLittleEndian() ? Legal : Expand);
+      }
+      for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
+        setTruncStoreAction(VT, InnerVT, Expand);
+        setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
+        setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
+        setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
+      }
+    }
   }
 
   if (Subtarget->hasSME()) {
@@ -2020,14 +2038,16 @@ void AArch64TargetLowering::addTypeForFixedLengthSVE(MVT VT) {
   setOperationAction(ISD::ZERO_EXTEND, VT, Default);
 }
 
-void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
+void AArch64TargetLowering::addDRType(MVT VT) {
   addRegisterClass(VT, &AArch64::FPR64RegClass);
-  addTypeForNEON(VT);
+  if (Subtarget->isNeonAvailable())
+    addTypeForNEON(VT);
 }
 
-void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
+void AArch64TargetLowering::addQRType(MVT VT) {
   addRegisterClass(VT, &AArch64::FPR128RegClass);
-  addTypeForNEON(VT);
+  if (Subtarget->isNeonAvailable())
+    addTypeForNEON(VT);
 }
 
 EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &,
@@ -9445,7 +9465,8 @@ SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
 
 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
                                               SelectionDAG &DAG) const {
-  if (!Subtarget->hasNEON())
+  if (!Subtarget->isNeonAvailable() &&
+      !Subtarget->useSVEForFixedLengthVectors())
     return SDValue();
 
   EVT VT = Op.getValueType();
@@ -14141,6 +14162,13 @@ SDValue AArch64TargetLowering::LowerDIV(SDValue Op, SelectionDAG &DAG) const {
   return DAG.getNode(AArch64ISD::UZP1, dl, VT, ResultLo, ResultHi);
 }
 
+bool AArch64TargetLowering::shouldExpandBuildVectorWithShuffles(
+    EVT VT, unsigned DefinedValues) const {
+  if (!Subtarget->isNeonAvailable())
+    return false;
+  return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues);
+}
+
 bool AArch64TargetLowering::isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const {
   // Currently no fixed length shuffles that require SVE are legal.
   if (useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()))
@@ -19838,7 +19866,8 @@ performSVEMulAddSubCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
 // help, for example, to produce ssra from sshr+add.
 static SDValue performAddSubIntoVectorOp(SDNode *N, SelectionDAG &DAG) {
   EVT VT = N->getValueType(0);
-  if (VT != MVT::i64)
+  if (VT != MVT::i64 ||
+      DAG.getTargetLoweringInfo().isOperationExpand(N->getOpcode(), MVT::v1i64))
     return SDValue();
   SDValue Op0 = N->getOperand(0);
   SDValue Op1 = N->getOperand(1);

llvm/lib/Target/AArch64/AArch64ISelLowering.h

@@ -1017,8 +1017,10 @@ class AArch64TargetLowering : public TargetLowering {
 
   void addTypeForNEON(MVT VT);
   void addTypeForFixedLengthSVE(MVT VT);
-  void addDRTypeForNEON(MVT VT);
-  void addQRTypeForNEON(MVT VT);
+  void addDRType(MVT VT);
+  void addQRType(MVT VT);
+
+  bool shouldExpandBuildVectorWithShuffles(EVT, unsigned) const override;
 
   unsigned allocateLazySaveBuffer(SDValue &Chain, const SDLoc &DL,
                                   SelectionDAG &DAG) const;