MIR Statepoint refactoring: pass GC pointers in VRegs. Part 4/5.

[dantrushin]

Summary

Change MIR representation of STATEPOINT instruction from explicit
spill/reload of GC pointer arguments around statepoint to simply
assigning them to virtual registers. Then they can be spilled as
necessary. Relocated pointers are represented as STATEPOINT's Def
operands (tied to corresponding derived pointers from GC args list):

rel1,rel2,... = STATEPOINT ..., derived1<tied-def0>, derived2<tied-def1>, ...

This patch implements memory operand folding for statepoints.
InlineSpiller untie operands it need to fold, because it knows how to
reload them later. But if we asked to fold tied operand we do nothing,
because unaware caller does not know how to reload result.

Depends On: D80193

[dantrushin]

Abandoned in favor of D81646

llvm/lib/CodeGen/InlineSpiller.cpp

foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>> Ops,
   // Don't attempt folding in bundles.
   MachineInstr *MI = Ops.front().first;
   if (Ops.back().first != MI || MI->isBundled())
     return false;
 
   bool WasCopy = MI->isCopy();
   unsigned ImpReg = 0;
 
+  bool UntieRegs = MI->getOpcode() == TargetOpcode::STATEPOINT;
+
   // Spill subregs if the target allows it.
   // We always want to spill subregs for stackmap/patchpoint pseudos.
   bool SpillSubRegs = TII.isSubregFoldable() ||
                       MI->getOpcode() == TargetOpcode::STATEPOINT ||
                       MI->getOpcode() == TargetOpcode::PATCHPOINT ||
                       MI->getOpcode() == TargetOpcode::STACKMAP;
 
   // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied
   // operands.
   SmallVector<unsigned, 8> FoldOps;
   for (const auto &OpPair : Ops) {
     unsigned Idx = OpPair.second;
     assert(MI == OpPair.first && "Instruction conflict during operand folding");
     MachineOperand &MO = MI->getOperand(Idx);
     if (MO.isImplicit()) {
       ImpReg = MO.getReg();
       continue;
     }
 
+    if (UntieRegs && MO.isTied())
+      MI->untieRegOperand(Idx);
+
     if (!SpillSubRegs && MO.getSubReg())
       return false;
     // We cannot fold a load instruction into a def.
     if (LoadMI && MO.isDef())
       return false;
     // Tied use operands should not be passed to foldMemoryOperand.
     if (!MI->isRegTiedToDefOperand(Idx))
       FoldOps.push_back(Idx);

llvm/lib/CodeGen/TargetInstrInfo.cpp

 void TargetInstrInfo::getNoop(MCInst &NopInst) const {
   llvm_unreachable("Not implemented");
 }
 
 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr &MI,
                                     ArrayRef<unsigned> Ops, int FrameIndex,
                                     const TargetInstrInfo &TII) {
   unsigned StartIdx = 0;
+  unsigned NumDefs = 0;
   switch (MI.getOpcode()) {
   case TargetOpcode::STACKMAP: {
     // StackMapLiveValues are foldable
     StartIdx = StackMapOpers(&MI).getVarIdx();
     break;
   }
   case TargetOpcode::PATCHPOINT: {
     // For PatchPoint, the call args are not foldable (even if reported in the
     // stackmap e.g. via anyregcc).
     StartIdx = PatchPointOpers(&MI).getVarIdx();
     break;
   }
   case TargetOpcode::STATEPOINT: {
     // For statepoints, fold deopt and gc arguments, but not call arguments.
     StartIdx = StatepointOpers(&MI).getVarIdx();
+    NumDefs = MI.getNumDefs();
     break;
   }
   default:
     llvm_unreachable("unexpected stackmap opcode");
   }
 
+  unsigned DefToFoldIdx = MI.getNumOperands();
+
   // Return false if any operands requested for folding are not foldable (not
   // part of the stackmap's live values).
   for (unsigned Op : Ops) {
-    if (Op < StartIdx)
+    if (Op < NumDefs)
+      DefToFoldIdx = Op;
+    else if (Op < StartIdx)
+      return nullptr;
+    // When called from regalloc (InlineSpiller), operands must be untied,
+    // and regalloc will take care of (re)loading operand from memory.
+    // But when called from other places (e.g. peephole pass),
+    // we cannot fold operand which are tied - callers are unaware they
+    // need to reload destination register.
+    if (MI.getOperand(Op).isTied())
       return nullptr;
   }
 
   MachineInstr *NewMI =
       MF.CreateMachineInstr(TII.get(MI.getOpcode()), MI.getDebugLoc(), true);
   MachineInstrBuilder MIB(MF, NewMI);
 
   // No need to fold return, the meta data, and function arguments
   for (unsigned i = 0; i < StartIdx; ++i)
+    if (i != DefToFoldIdx)
       MIB.add(MI.getOperand(i));
 
-  for (unsigned i = StartIdx; i < MI.getNumOperands(); ++i) {
+  for (unsigned i = StartIdx, e = MI.getNumOperands(); i < e; ++i) {
     MachineOperand &MO = MI.getOperand(i);
+    unsigned TiedTo = e;
+    (void)MI.isRegTiedToDefOperand(i, &TiedTo);
+
     if (is_contained(Ops, i)) {
+      assert(TiedTo == e && "Cannot fold tied operands");
       unsigned SpillSize;
       unsigned SpillOffset;
       // Compute the spill slot size and offset.
       const TargetRegisterClass *RC =
         MF.getRegInfo().getRegClass(MO.getReg());
       bool Valid =
           TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
       if (!Valid)
         report_fatal_error("cannot spill patchpoint subregister operand");
       MIB.addImm(StackMaps::IndirectMemRefOp);
       MIB.addImm(SpillSize);
       MIB.addFrameIndex(FrameIndex);
       MIB.addImm(SpillOffset);
-    }
-    else
+    } else {
       MIB.add(MO);
+      if (TiedTo < e) {
+        assert(TiedTo < NumDefs && "Bad tied operand");
+        if (TiedTo > DefToFoldIdx)
+          --TiedTo;
+        NewMI->tieOperands(TiedTo, NewMI->getNumOperands() - 1);
+      }
+    }
   }
   return NewMI;
 }
 
 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
                                                  ArrayRef<unsigned> Ops, int FI,
                                                  LiveIntervals *LIS,
                                                  VirtRegMap *VRM) const {