[CodeGen] Add MMOs to statepoint nodes during SelectionDAG

[reames]

Summary

The existing statepoint lowering code does something odd; it adds machine memory operands post instruction selection. This was copied from the stackmap/patchpoint implementation, but appears to be non-idiomatic.

This change is largely NFC. It moves the MMO creation logic into SelectionDAG building. It ends up not quite being NFC because the size of the stack slot is reflected in the MMO. The old code blindly used pointer size for the MMO size, which appears to have always been incorrect for larger values. It just happened nothing actually relied on the MMOs, so it worked out okay.

For context, I'm planning on removing the MOVolatile flag from these in a future commit, and then removing the MOStore flag from deopt spill slots in a separate one. Doing so is motivated by a small test case where we should be able to better schedule spill slots, but don't do so due to a memory use/def implied by the statepoint.

[skatkov]

lgtm.

lib/CodeGen/SelectionDAG/StatepointLowering.cpp

473

new line missed.

lib/CodeGen/SelectionDAG/StatepointLowering.cpp

else
       while (CallEnd->getOpcode() == ISD::CopyFromReg)
         CallEnd = CallEnd->getOperand(0).getNode();
   }
 
   assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
   return std::make_pair(ReturnValue, CallEnd->getOperand(0).getNode());
 }
 
+static MachineMemOperand* getMachineMemOperand(MachineFunction &MF,
+                                               FrameIndexSDNode &FI) {
+  auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI.getIndex());
+  auto MMOFlags = MachineMemOperand::MOStore |
+    MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile;
+  auto &MFI = MF.getFrameInfo();
+  return MF.getMachineMemOperand(PtrInfo, MMOFlags, 
+                                 MFI.getObjectSize(FI.getIndex()),
+                                 MFI.getObjectAlignment(FI.getIndex()));
+}
+
 /// Spill a value incoming to the statepoint. It might be either part of
 /// vmstate
 /// or gcstate. In both cases unconditionally spill it on the stack unless it
 /// is a null constant. Return pair with first element being frame index
 /// containing saved value and second element with outgoing chain from the
 /// emitted store
-static std::pair<SDValue, SDValue>
+static std::tuple<SDValue, SDValue, MachineMemOperand*>
 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
                              SelectionDAGBuilder &Builder) {
   SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
+  MachineMemOperand* MMO = nullptr;
 
   // Emit new store if we didn't do it for this ptr before
   if (!Loc.getNode()) {
     Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
                                                        Builder);
     int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
     // We use TargetFrameIndex so that isel will not select it into LEA
     Loc = Builder.DAG.getTargetFrameIndex(Index, Builder.getFrameIndexTy());
 
 #ifndef NDEBUG
     // Right now we always allocate spill slots that are of the same
     // size as the value we're about to spill (the size of spillee can
     // vary since we spill vectors of pointers too).  At some point we
     // can consider allowing spills of smaller values to larger slots
     // (i.e. change the '==' in the assert below to a '>=').
     MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
     assert((MFI.getObjectSize(Index) * 8) == Incoming.getValueSizeInBits() &&
            "Bad spill:  stack slot does not match!");
 #endif
 
+    auto &MF = Builder.DAG.getMachineFunction();
+    auto PtrInfo = MachinePointerInfo::getFixedStack(MF, Index);
     Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
-                                 MachinePointerInfo::getFixedStack(
-                                     Builder.DAG.getMachineFunction(), Index));
+                                 PtrInfo);
+
+    MMO = getMachineMemOperand(MF, *cast<FrameIndexSDNode>(Loc));
     
     Builder.StatepointLowering.setLocation(Incoming, Loc);
   }
 
   assert(Loc.getNode());
-  return std::make_pair(Loc, Chain);
+  return std::make_tuple(Loc, Chain, MMO);
 }
 
 /// Lower a single value incoming to a statepoint node.  This value can be
 /// either a deopt value or a gc value, the handling is the same.  We special
 /// case constants and allocas, then fall back to spilling if required.
 static void lowerIncomingStatepointValue(SDValue Incoming, bool LiveInOnly,
                                          SmallVectorImpl<SDValue> &Ops,
+                                         SmallVectorImpl<MachineMemOperand*> &MemRefs,
                                          SelectionDAGBuilder &Builder) {
   // Note: We know all of these spills are independent, but don't bother to
   // exploit that chain wise.  DAGCombine will happily do so as needed, so
   // doing it here would be a small compile time win at most.
   SDValue Chain = Builder.getRoot();
 
   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
     // If the original value was a constant, make sure it gets recorded as
     // such in the stackmap.  This is required so that the consumer can
     // parse any internal format to the deopt state.  It also handles null
     // pointers and other constant pointers in GC states.  Note the constant
     // vectors do not appear to actually hit this path and that anything larger
     // than an i64 value (not type!) will fail asserts here.
     pushStackMapConstant(Ops, Builder, C->getSExtValue());
   } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
     // This handles allocas as arguments to the statepoint (this is only
     // really meaningful for a deopt value.  For GC, we'd be trying to
     // relocate the address of the alloca itself?)
     assert(Incoming.getValueType() == Builder.getFrameIndexTy() &&
            "Incoming value is a frame index!");
     Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
                                                   Builder.getFrameIndexTy()));
+
+    auto &MF = Builder.DAG.getMachineFunction();
+    auto *MMO = getMachineMemOperand(MF, *FI);
+    MemRefs.push_back(MMO);
+    
   } else if (LiveInOnly) {
     // If this value is live in (not live-on-return, or live-through), we can
     // treat it the same way patchpoint treats it's "live in" values.  We'll
     // end up folding some of these into stack references, but they'll be
     // handled by the register allocator.  Note that we do not have the notion
     // of a late use so these values might be placed in registers which are
     // clobbered by the call.  This is fine for live-in.
     Ops.push_back(Incoming);
   } else {
     // Otherwise, locate a spill slot and explicitly spill it so it
     // can be found by the runtime later.  We currently do not support
     // tracking values through callee saved registers to their eventual
     // spill location.  This would be a useful optimization, but would
     // need to be optional since it requires a lot of complexity on the
     // runtime side which not all would support.
     auto Res = spillIncomingStatepointValue(Incoming, Chain, Builder);
-    Ops.push_back(Res.first);
-    Chain = Res.second;
+    Ops.push_back(std::get<0>(Res));
+    if (auto *MMO = std::get<2>(Res))
+      MemRefs.push_back(MMO);
+    Chain = std::get<1>(Res);;
   }
 
   Builder.DAG.setRoot(Chain);
 }
 
 /// Lower deopt state and gc pointer arguments of the statepoint.  The actual
 /// lowering is described in lowerIncomingStatepointValue.  This function is
 /// responsible for lowering everything in the right position and playing some
 /// tricks to avoid redundant stack manipulation where possible.  On
 /// completion, 'Ops' will contain ready to use operands for machine code
 /// statepoint. The chain nodes will have already been created and the DAG root
 /// will be set to the last value spilled (if any were).
 static void
 lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
-                        SelectionDAGBuilder::StatepointLoweringInfo &SI,
+                        SmallVectorImpl<MachineMemOperand*> &MemRefs,                                    SelectionDAGBuilder::StatepointLoweringInfo &SI,
                         SelectionDAGBuilder &Builder) {
   // Lower the deopt and gc arguments for this statepoint.  Layout will be:
   // deopt argument length, deopt arguments.., gc arguments...
 #ifndef NDEBUG
   if (auto *GFI = Builder.GFI) {
     // Check that each of the gc pointer and bases we've gotten out of the
     // safepoint is something the strategy thinks might be a pointer (or vector
     // of pointers) into the GC heap.  This is basically just here to help catch
for (const Value *V : SI.DeoptState) {
     if (const Argument *Arg = dyn_cast<Argument>(V)) {
       int FI = Builder.FuncInfo.getArgumentFrameIndex(Arg);
       if (FI != INT_MAX)
         Incoming = Builder.DAG.getFrameIndex(FI, Builder.getFrameIndexTy());
     }
     if (!Incoming.getNode())
       Incoming = Builder.getValue(V);
     const bool LiveInValue = LiveInDeopt && !isGCValue(V);
-    lowerIncomingStatepointValue(Incoming, LiveInValue, Ops, Builder);
+    lowerIncomingStatepointValue(Incoming, LiveInValue, Ops, MemRefs, Builder);
   }
 
   // Finally, go ahead and lower all the gc arguments.  There's no prefixed
   // length for this one.  After lowering, we'll have the base and pointer
   // arrays interwoven with each (lowered) base pointer immediately followed by
   // it's (lowered) derived pointer.  i.e
   // (base[0], ptr[0], base[1], ptr[1], ...)
   for (unsigned i = 0; i < SI.Bases.size(); ++i) {
     const Value *Base = SI.Bases[i];
     lowerIncomingStatepointValue(Builder.getValue(Base), /*LiveInOnly*/ false,
-                                 Ops, Builder);
+                                 Ops, MemRefs, Builder);
 
     const Value *Ptr = SI.Ptrs[i];
     lowerIncomingStatepointValue(Builder.getValue(Ptr), /*LiveInOnly*/ false,
-                                 Ops, Builder);
+                                 Ops, MemRefs, Builder);
   }
 
   // If there are any explicit spill slots passed to the statepoint, record
   // them, but otherwise do not do anything special.  These are user provided
   // allocas and give control over placement to the consumer.  In this case,
   // it is the contents of the slot which may get updated, not the pointer to
   // the alloca
   for (Value *V : SI.GCArgs) {
     SDValue Incoming = Builder.getValue(V);
     if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
       // This handles allocas as arguments to the statepoint
       assert(Incoming.getValueType() == Builder.getFrameIndexTy() &&
              "Incoming value is a frame index!");
       Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
                                                     Builder.getFrameIndexTy()));
+
+      auto &MF = Builder.DAG.getMachineFunction();
+      auto *MMO = getMachineMemOperand(MF, *FI);
+      MemRefs.push_back(MMO);
     }
   }
 
   // Record computed locations for all lowered values.
   // This can not be embedded in lowering loops as we need to record *all*
   // values, while previous loops account only values with unique SDValues.
   const Instruction *StatepointInstr = SI.StatepointInstr;
   auto &SpillMap = Builder.FuncInfo.StatepointSpillMaps[StatepointInstr];
#endif
   // reducing the size of the StackMap section.  It has no other impact.
   removeDuplicateGCPtrs(SI.Bases, SI.Ptrs, SI.GCRelocates, *this,
                         FuncInfo.StatepointSpillMaps[SI.StatepointInstr]);
   assert(SI.Bases.size() == SI.Ptrs.size() &&
          SI.Ptrs.size() == SI.GCRelocates.size());
 
   // Lower statepoint vmstate and gcstate arguments
   SmallVector<SDValue, 10> LoweredMetaArgs;
-  lowerStatepointMetaArgs(LoweredMetaArgs, SI, *this);
+  SmallVector<MachineMemOperand*, 16> MemRefs;
+  lowerStatepointMetaArgs(LoweredMetaArgs, MemRefs, SI, *this);
 
   // Now that we've emitted the spills, we need to update the root so that the
   // call sequence is ordered correctly.
   SI.CLI.setChain(getRoot());
 
   // Get call node, we will replace it later with statepoint
   SDValue ReturnVal;
   SDNode *CallNode;
#endif
   // Same for the glue, but we add it only if original call had it
   if (Glue.getNode())
     Ops.push_back(Glue);
 
   // Compute return values.  Provide a glue output since we consume one as
   // input.  This allows someone else to chain off us as needed.
   SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
 
-  SDNode *StatepointMCNode =
-      DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
+  MachineSDNode *StatepointMCNode =
+  DAG.setNodeMemRefs(StatepointMCNode, MemRefs);
 
   SDNode *SinkNode = StatepointMCNode;
 
   // Build the GC_TRANSITION_END node if necessary.
   //
   // See the comment above regarding GC_TRANSITION_START for the layout of
   // the operands to the GC_TRANSITION_END node.
   if (IsGCTransition) {

lib/CodeGen/TargetLoweringBase.cpp

for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) {
 
     // Inherit previous memory operands.
     MIB.cloneMemRefs(*MI);
     assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
 
     // Add a new memory operand for this FI.
     assert(MFI.getObjectOffset(FI) != -1);
 
+    // Note: STATEPOINT MMOs are added during SelectionDAG.  STACKMAP, and
+    // PATCHPOINT should be updated to do the same. (TODO)
+    if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
       auto Flags = MachineMemOperand::MOLoad;
-    if (MI->getOpcode() == TargetOpcode::STATEPOINT) {
-      Flags |= MachineMemOperand::MOStore;
-      Flags |= MachineMemOperand::MOVolatile;
-    }
       MachineMemOperand *MMO = MF.getMachineMemOperand(
           MachinePointerInfo::getFixedStack(MF, FI), Flags,
           MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI));
       MIB->addMemOperand(MF, MMO);
+    }
     
     // Replace the instruction and update the operand index.
     MBB->insert(MachineBasicBlock::iterator(MI), MIB);
     OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1;
     MI->eraseFromParent();
     MI = MIB;
   }
   return MBB;
 }