AMDGPU: Make i64 loads/stores promote to v2i32

[arsenm]

Summary

Now that unaligned access expansion should not attempt
to produce i64 accesses, we can remove the hack in
PreprocessISelDAG where this is done.

This allows splitting i64 private accesses while
allowing the new add nodes indexing the vector components
can be folded with the base pointer arithmetic.

[arsenm]

The problem with this is that the promoted node counts as legalized, so the custom lowering for v2i32 is never called for it, so we can't split it up when necessary for private accesses

[• tstellarAMD]

LGTM.

[arsenm]

r268293

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

class AMDGPUDAGToDAGISel : public SelectionDAGISel {
   const AMDGPUSubtarget *Subtarget;
 
 public:
   AMDGPUDAGToDAGISel(TargetMachine &TM);
   virtual ~AMDGPUDAGToDAGISel();
   bool runOnMachineFunction(MachineFunction &MF) override;
   SDNode *Select(SDNode *N) override;
   const char *getPassName() const override;
-  void PreprocessISelDAG() override;
   void PostprocessISelDAG() override;
 
 private:
   bool isInlineImmediate(SDNode *N) const;
   bool FoldOperand(SDValue &Src, SDValue &Sel, SDValue &Neg, SDValue &Abs,
                    const R600InstrInfo *TII);
   bool FoldOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &);
   bool FoldDotOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &);

 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src,
                                                    SDValue &SrcMods,
                                                    SDValue &Clamp,
                                                    SDValue &Omod) const {
   Clamp = Omod = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);
   return SelectVOP3Mods(In, Src, SrcMods);
 }
 
-void AMDGPUDAGToDAGISel::PreprocessISelDAG() {
-  bool Modified = false;
-
-  // XXX - Other targets seem to be able to do this without a worklist.
-  SmallVector<LoadSDNode *, 8> LoadsToReplace;
-  SmallVector<StoreSDNode *, 8> StoresToReplace;
-
-  for (SDNode &Node : CurDAG->allnodes()) {
-    if (LoadSDNode *LD = dyn_cast<LoadSDNode>(&Node)) {
-      EVT VT = LD->getValueType(0);
-      if (VT != MVT::i64 || LD->getExtensionType() != ISD::NON_EXTLOAD)
-        continue;
-
-      // To simplify the TableGen patters, we replace all i64 loads with v2i32
-      // loads.  Alternatively, we could promote i64 loads to v2i32 during DAG
-      // legalization, however, so places (ExpandUnalignedLoad) in the DAG
-      // legalizer assume that if i64 is legal, so doing this promotion early
-      // can cause problems.
-      LoadsToReplace.push_back(LD);
-    } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(&Node)) {
-      // Handle i64 stores here for the same reason mentioned above for loads.
-      SDValue Value = ST->getValue();
-      if (Value.getValueType() != MVT::i64 || ST->isTruncatingStore())
-        continue;
-      StoresToReplace.push_back(ST);
-    }
-  }
-
-  for (LoadSDNode *LD : LoadsToReplace) {
-    SDLoc SL(LD);
-
-    SDValue NewLoad = CurDAG->getLoad(MVT::v2i32, SL, LD->getChain(),
-                                      LD->getBasePtr(), LD->getMemOperand());
-    SDValue BitCast = CurDAG->getNode(ISD::BITCAST, SL,
-                                      MVT::i64, NewLoad);
-    CurDAG->ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLoad.getValue(1));
-    CurDAG->ReplaceAllUsesOfValueWith(SDValue(LD, 0), BitCast);
-    Modified = true;
-  }
-
-  for (StoreSDNode *ST : StoresToReplace) {
-    SDValue NewValue = CurDAG->getNode(ISD::BITCAST, SDLoc(ST),
-                                       MVT::v2i32, ST->getValue());
-    const SDValue StoreOps[] = {
-      ST->getChain(),
-      NewValue,
-      ST->getBasePtr(),
-      ST->getOffset()
-    };
-
-    CurDAG->UpdateNodeOperands(ST, StoreOps);
-    Modified = true;
-  }
-
-  // XXX - Is this necessary?
-  if (Modified)
-    CurDAG->RemoveDeadNodes();
-}
-
 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
   const AMDGPUTargetLowering& Lowering =
     *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
   bool IsModified = false;
   do {
     IsModified = false;
     // Go over all selected nodes and try to fold them a bit more
     for (SDNode &Node : CurDAG->allnodes()) {

lib/Target/AMDGPU/SIISelLowering.cpp

SITargetLowering::SITargetLowering(TargetMachine &TM,
   setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
   setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
 
   // We need to custom lower vector stores from local memory
   setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
 
+  setOperationAction(ISD::LOAD, MVT::f64, Promote);
+  AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32);
+
+  setOperationAction(ISD::LOAD, MVT::i64, Promote);
+  AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
+
   setOperationAction(ISD::STORE, MVT::v8i32, Custom);
   setOperationAction(ISD::STORE, MVT::v16i32, Custom);
 
   setOperationAction(ISD::STORE, MVT::i1, Custom);
   setOperationAction(ISD::STORE, MVT::v4i32, Custom);
 
+  setOperationAction(ISD::STORE, MVT::f64, Promote);
+  AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32);
+
+  setOperationAction(ISD::STORE, MVT::i64, Promote);
+  AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
+
   setOperationAction(ISD::SELECT, MVT::i64, Custom);
   setOperationAction(ISD::SELECT, MVT::f64, Promote);
   AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
 
   setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
   setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
   setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
   setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);