X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FSelectionDAG%2FStatepointLowering.cpp;h=050ec2116c5d8b2e783fe0200ff271d33c00e9ec;hb=da801219ba8d6d2a8663d4dd3c14e8e3fca35ba5;hp=ed5ecd13667a1a2bdb834cfdc0617baaa1d90c64;hpb=716c5d8a308a2257298f1f227edf7f7ae102cf4f;p=oota-llvm.git diff --git a/lib/CodeGen/SelectionDAG/StatepointLowering.cpp b/lib/CodeGen/SelectionDAG/StatepointLowering.cpp index ed5ecd13667..050ec2116c5 100644 --- a/lib/CodeGen/SelectionDAG/StatepointLowering.cpp +++ b/lib/CodeGen/SelectionDAG/StatepointLowering.cpp @@ -17,6 +17,7 @@ #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" #include "llvm/CodeGen/FunctionLoweringInfo.h" +#include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/GCMetadata.h" #include "llvm/CodeGen/GCStrategy.h" #include "llvm/CodeGen/SelectionDAG.h" @@ -38,13 +39,19 @@ STATISTIC(NumOfStatepoints, "Number of statepoint nodes encountered"); STATISTIC(StatepointMaxSlotsRequired, "Maximum number of stack slots required for a singe statepoint"); -void -StatepointLoweringState::startNewStatepoint(SelectionDAGBuilder &Builder) { +static void pushStackMapConstant(SmallVectorImpl& Ops, + SelectionDAGBuilder &Builder, uint64_t Value) { + SDLoc L = Builder.getCurSDLoc(); + Ops.push_back(Builder.DAG.getTargetConstant(StackMaps::ConstantOp, L, + MVT::i64)); + Ops.push_back(Builder.DAG.getTargetConstant(Value, L, MVT::i64)); +} + +void StatepointLoweringState::startNewStatepoint(SelectionDAGBuilder &Builder) { // Consistency check assert(PendingGCRelocateCalls.empty() && "Trying to visit statepoint before finished processing previous one"); Locations.clear(); - RelocLocations.clear(); NextSlotToAllocate = 0; // Need to resize this on each safepoint - we need the two to stay in // sync and the clear patterns of a SelectionDAGBuilder have no relation @@ -54,9 +61,9 @@ StatepointLoweringState::startNewStatepoint(SelectionDAGBuilder &Builder) { AllocatedStackSlots[i] = false; } } + void StatepointLoweringState::clear() { Locations.clear(); - RelocLocations.clear(); AllocatedStackSlots.clear(); assert(PendingGCRelocateCalls.empty() && "cleared before statepoint sequence completed"); @@ -89,6 +96,9 @@ StatepointLoweringState::allocateStackSlot(EVT ValueType, SDValue SpillSlot = Builder.DAG.CreateStackTemporary(ValueType); const unsigned FI = cast(SpillSlot)->getIndex(); + auto *MFI = Builder.DAG.getMachineFunction().getFrameInfo(); + MFI->markAsStatepointSpillSlotObjectIndex(FI); + Builder.FuncInfo.StatepointStackSlots.push_back(FI); AllocatedStackSlots.push_back(true); return SpillSlot; @@ -99,92 +109,145 @@ StatepointLoweringState::allocateStackSlot(EVT ValueType, return Builder.DAG.getFrameIndex(FI, ValueType); } // Note: We deliberately choose to advance this only on the failing path. - // Doing so on the suceeding path involes a bit of complexity that caused a - // minor bug previously. Unless performance shows this matters, please + // Doing so on the succeeding path involves a bit of complexity that caused + // a minor bug previously. Unless performance shows this matters, please // keep this code as simple as possible. NextSlotToAllocate++; } llvm_unreachable("infinite loop?"); } +/// Utility function for reservePreviousStackSlotForValue. Tries to find +/// stack slot index to which we have spilled value for previous statepoints. +/// LookUpDepth specifies maximum DFS depth this function is allowed to look. +static Optional findPreviousSpillSlot(const Value *Val, + SelectionDAGBuilder &Builder, + int LookUpDepth) { + // Can not look any further - give up now + if (LookUpDepth <= 0) + return Optional(); + + // Spill location is known for gc relocates + if (isGCRelocate(Val)) { + GCRelocateOperands RelocOps(cast(Val)); + + FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap = + Builder.FuncInfo.StatepointRelocatedValues[RelocOps.getStatepoint()]; + + auto It = SpillMap.find(RelocOps.getDerivedPtr()); + if (It == SpillMap.end()) + return Optional(); + + return It->second; + } + + // Look through bitcast instructions. + if (const BitCastInst *Cast = dyn_cast(Val)) { + return findPreviousSpillSlot(Cast->getOperand(0), Builder, LookUpDepth - 1); + } + + // Look through phi nodes + // All incoming values should have same known stack slot, otherwise result + // is unknown. + if (const PHINode *Phi = dyn_cast(Val)) { + Optional MergedResult = None; + + for (auto &IncomingValue : Phi->incoming_values()) { + Optional SpillSlot = + findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth - 1); + if (!SpillSlot.hasValue()) + return Optional(); + + if (MergedResult.hasValue() && *MergedResult != *SpillSlot) + return Optional(); + + MergedResult = SpillSlot; + } + return MergedResult; + } + + // TODO: We can do better for PHI nodes. In cases like this: + // ptr = phi(relocated_pointer, not_relocated_pointer) + // statepoint(ptr) + // We will return that stack slot for ptr is unknown. And later we might + // assign different stack slots for ptr and relocated_pointer. This limits + // llvm's ability to remove redundant stores. + // Unfortunately it's hard to accomplish in current infrastructure. + // We use this function to eliminate spill store completely, while + // in example we still need to emit store, but instead of any location + // we need to use special "preferred" location. + + // TODO: handle simple updates. If a value is modified and the original + // value is no longer live, it would be nice to put the modified value in the + // same slot. This allows folding of the memory accesses for some + // instructions types (like an increment). + // statepoint (i) + // i1 = i+1 + // statepoint (i1) + // However we need to be careful for cases like this: + // statepoint(i) + // i1 = i+1 + // statepoint(i, i1) + // Here we want to reserve spill slot for 'i', but not for 'i+1'. If we just + // put handling of simple modifications in this function like it's done + // for bitcasts we might end up reserving i's slot for 'i+1' because order in + // which we visit values is unspecified. + + // Don't know any information about this instruction + return Optional(); +} + /// Try to find existing copies of the incoming values in stack slots used for /// statepoint spilling. If we can find a spill slot for the incoming value, /// mark that slot as allocated, and reuse the same slot for this safepoint. -/// This helps to avoid series of loads and stores that only serve to resuffle +/// This helps to avoid series of loads and stores that only serve to reshuffle /// values on the stack between calls. -static void reservePreviousStackSlotForValue(SDValue Incoming, +static void reservePreviousStackSlotForValue(const Value *IncomingValue, SelectionDAGBuilder &Builder) { + SDValue Incoming = Builder.getValue(IncomingValue); + if (isa(Incoming) || isa(Incoming)) { // We won't need to spill this, so no need to check for previously // allocated stack slots return; } - SDValue Loc = Builder.StatepointLowering.getLocation(Incoming); - if (Loc.getNode()) { + SDValue OldLocation = Builder.StatepointLowering.getLocation(Incoming); + if (OldLocation.getNode()) // duplicates in input return; - } - - // Search back for the load from a stack slot pattern to find the original - // slot we allocated for this value. We could extend this to deal with - // simple modification patterns, but simple dealing with trivial load/store - // sequences helps a lot already. - if (LoadSDNode *Load = dyn_cast(Incoming)) { - if (auto *FI = dyn_cast(Load->getBasePtr())) { - const int Index = FI->getIndex(); - auto Itr = std::find(Builder.FuncInfo.StatepointStackSlots.begin(), - Builder.FuncInfo.StatepointStackSlots.end(), Index); - if (Itr == Builder.FuncInfo.StatepointStackSlots.end()) { - // not one of the lowering stack slots, can't reuse! - // TODO: Actually, we probably could reuse the stack slot if the value - // hasn't changed at all, but we'd need to look for intervening writes - return; - } else { - // This is one of our dedicated lowering slots - const int Offset = - std::distance(Builder.FuncInfo.StatepointStackSlots.begin(), Itr); - if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) { - // stack slot already assigned to someone else, can't use it! - // TODO: currently we reserve space for gc arguments after doing - // normal allocation for deopt arguments. We should reserve for - // _all_ deopt and gc arguments, then start allocating. This - // will prevent some moves being inserted when vm state changes, - // but gc state doesn't between two calls. - return; - } - // Reserve this stack slot - Builder.StatepointLowering.reserveStackSlot(Offset); - } - // Cache this slot so we find it when going through the normal - // assignment loop. - SDValue Loc = - Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType()); + const int LookUpDepth = 6; + Optional Index = + findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth); + if (!Index.hasValue()) + return; - Builder.StatepointLowering.setLocation(Incoming, Loc); - } + auto Itr = std::find(Builder.FuncInfo.StatepointStackSlots.begin(), + Builder.FuncInfo.StatepointStackSlots.end(), *Index); + assert(Itr != Builder.FuncInfo.StatepointStackSlots.end() && + "value spilled to the unknown stack slot"); + + // This is one of our dedicated lowering slots + const int Offset = + std::distance(Builder.FuncInfo.StatepointStackSlots.begin(), Itr); + if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) { + // stack slot already assigned to someone else, can't use it! + // TODO: currently we reserve space for gc arguments after doing + // normal allocation for deopt arguments. We should reserve for + // _all_ deopt and gc arguments, then start allocating. This + // will prevent some moves being inserted when vm state changes, + // but gc state doesn't between two calls. + return; } + // Reserve this stack slot + Builder.StatepointLowering.reserveStackSlot(Offset); - // TODO: handle case where a reloaded value flows through a phi to - // another safepoint. e.g. - // bb1: - // a' = relocated... - // bb2: % pred: bb1, bb3, bb4, etc. - // a_phi = phi(a', ...) - // statepoint ... a_phi - // NOTE: This will require reasoning about cross basic block values. This is - // decidedly non trivial and this might not be the right place to do it. We - // don't really have the information we need here... - - // TODO: handle simple updates. If a value is modified and the original - // value is no longer live, it would be nice to put the modified value in the - // same slot. This allows folding of the memory accesses for some - // instructions types (like an increment). - // statepoint (i) - // i1 = i+1 - // statepoint (i1) + // Cache this slot so we find it when going through the normal + // assignment loop. + SDValue Loc = Builder.DAG.getTargetFrameIndex(*Index, Incoming.getValueType()); + Builder.StatepointLowering.setLocation(Incoming, Loc); } /// Remove any duplicate (as SDValues) from the derived pointer pairs. This @@ -196,7 +259,7 @@ static void removeDuplicatesGCPtrs(SmallVectorImpl &Bases, SmallVectorImpl &Relocs, SelectionDAGBuilder &Builder) { - // This is horribly ineffecient, but I don't care right now + // This is horribly inefficient, but I don't care right now SmallSet Seen; SmallVector NewBases, NewPtrs, NewRelocs; @@ -223,99 +286,108 @@ static void removeDuplicatesGCPtrs(SmallVectorImpl &Bases, /// Extract call from statepoint, lower it and return pointer to the /// call node. Also update NodeMap so that getValue(statepoint) will /// reference lowered call result -static SDNode *lowerCallFromStatepoint(ImmutableStatepoint StatepointSite, - MachineBasicBlock *LandingPad, - SelectionDAGBuilder &Builder) { - - ImmutableCallSite CS(StatepointSite.getCallSite()); - - // Lower the actual call itself - This is a bit of a hack, but we want to - // avoid modifying the actual lowering code. This is similiar in intent to - // the LowerCallOperands mechanism used by PATCHPOINT, but is structured - // differently. Hopefully, this is slightly more robust w.r.t. calling - // convention, return values, and other function attributes. - Value *ActualCallee = const_cast(StatepointSite.actualCallee()); - - std::vector Args; - CallInst::const_op_iterator arg_begin = StatepointSite.call_args_begin(); - CallInst::const_op_iterator arg_end = StatepointSite.call_args_end(); - Args.insert(Args.end(), arg_begin, arg_end); - // TODO: remove the creation of a new instruction! We should not be - // modifying the IR (even temporarily) at this point. - CallInst *Tmp = CallInst::Create(ActualCallee, Args); - Tmp->setTailCall(CS.isTailCall()); - Tmp->setCallingConv(CS.getCallingConv()); - Tmp->setAttributes(CS.getAttributes()); - Builder.LowerCallTo(Tmp, Builder.getValue(ActualCallee), false, LandingPad); - - // Handle the return value of the call iff any. - const bool HasDef = !Tmp->getType()->isVoidTy(); - if (HasDef) { - if (CS.isInvoke()) { - // Result value will be used in different basic block for invokes - // so we need to export it now. But statepoint call has a different type - // than the actuall call. It means that standart exporting mechanism will - // create register of the wrong type. So instead we need to create - // register with correct type and save value into it manually. - // TODO: To eliminate this problem we can remove gc.result intrinsics - // completelly and make statepoint call to return a tuple. - unsigned reg = Builder.FuncInfo.CreateRegs(Tmp->getType()); - Builder.CopyValueToVirtualRegister(Tmp, reg); - Builder.FuncInfo.ValueMap[CS.getInstruction()] = reg; - } - else { - // The value of the statepoint itself will be the value of call itself. - // We'll replace the actually call node shortly. gc_result will grab - // this value. - Builder.setValue(CS.getInstruction(), Builder.getValue(Tmp)); - } - } else { - // The token value is never used from here on, just generate a poison value - Builder.setValue(CS.getInstruction(), - Builder.DAG.getIntPtrConstant(-1, Builder.getCurSDLoc())); - } - // Remove the fake entry we created so we don't have a hanging reference - // after we delete this node. - Builder.removeValue(Tmp); - delete Tmp; - Tmp = nullptr; - - // Search for the call node - // The following code is essentially reverse engineering X86's +static SDNode * +lowerCallFromStatepoint(ImmutableStatepoint ISP, const BasicBlock *EHPadBB, + SelectionDAGBuilder &Builder, + SmallVectorImpl &PendingExports) { + + ImmutableCallSite CS(ISP.getCallSite()); + + SDValue ActualCallee; + + if (ISP.getNumPatchBytes() > 0) { + // If we've been asked to emit a nop sequence instead of a call instruction + // for this statepoint then don't lower the call target, but use a constant + // `null` instead. Not lowering the call target lets statepoint clients get + // away without providing a physical address for the symbolic call target at + // link time. + + const auto &TLI = Builder.DAG.getTargetLoweringInfo(); + const auto &DL = Builder.DAG.getDataLayout(); + + unsigned AS = ISP.getCalledValue()->getType()->getPointerAddressSpace(); + ActualCallee = Builder.DAG.getConstant(0, Builder.getCurSDLoc(), + TLI.getPointerTy(DL, AS)); + } else + ActualCallee = Builder.getValue(ISP.getCalledValue()); + + assert(CS.getCallingConv() != CallingConv::AnyReg && + "anyregcc is not supported on statepoints!"); + + Type *DefTy = ISP.getActualReturnType(); + bool HasDef = !DefTy->isVoidTy(); + + SDValue ReturnValue, CallEndVal; + std::tie(ReturnValue, CallEndVal) = Builder.lowerCallOperands( + ISP.getCallSite(), ImmutableStatepoint::CallArgsBeginPos, + ISP.getNumCallArgs(), ActualCallee, DefTy, EHPadBB, + false /* IsPatchPoint */); + + SDNode *CallEnd = CallEndVal.getNode(); + + // Get a call instruction from the call sequence chain. Tail calls are not + // allowed. The following code is essentially reverse engineering X86's // LowerCallTo. + // // We are expecting DAG to have the following form: + // // ch = eh_label (only in case of invoke statepoint) // ch, glue = callseq_start ch // ch, glue = X86::Call ch, glue // ch, glue = callseq_end ch, glue - // ch = eh_label ch (only in case of invoke statepoint) + // get_return_value ch, glue // - // DAG root will be either last eh_label or callseq_end. - - SDNode *CallNode = nullptr; - - // We just emitted a call, so it should be last thing generated - SDValue Chain = Builder.DAG.getRoot(); + // get_return_value can either be a sequence of CopyFromReg instructions + // to grab the return value from the return register(s), or it can be a LOAD + // to load a value returned by reference via a stack slot. - // Find closest CALLSEQ_END walking back through lowered nodes if needed - SDNode *CallEnd = Chain.getNode(); - int Sanity = 0; - while (CallEnd->getOpcode() != ISD::CALLSEQ_END) { - assert(CallEnd->getNumOperands() >= 1 && - CallEnd->getOperand(0).getValueType() == MVT::Other); - - CallEnd = CallEnd->getOperand(0).getNode(); + if (HasDef) { + if (CallEnd->getOpcode() == ISD::LOAD) + CallEnd = CallEnd->getOperand(0).getNode(); + else + while (CallEnd->getOpcode() == ISD::CopyFromReg) + CallEnd = CallEnd->getOperand(0).getNode(); + } - assert(Sanity < 20 && "should have found call end already"); - Sanity++; + assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!"); + + // Export the result value if needed + const Instruction *GCResult = ISP.getGCResult(); + if (HasDef && GCResult) { + if (GCResult->getParent() != CS.getParent()) { + // Result value will be used in a different basic block so we need to + // export it now. + // Default exporting mechanism will not work here because statepoint call + // has a different type than the actual call. It means that by default + // llvm will create export register of the wrong type (always i32 in our + // case). So instead we need to create export register with correct type + // manually. + // TODO: To eliminate this problem we can remove gc.result intrinsics + // completely and make statepoint call to return a tuple. + unsigned Reg = Builder.FuncInfo.CreateRegs(ISP.getActualReturnType()); + RegsForValue RFV( + *Builder.DAG.getContext(), Builder.DAG.getTargetLoweringInfo(), + Builder.DAG.getDataLayout(), Reg, ISP.getActualReturnType()); + SDValue Chain = Builder.DAG.getEntryNode(); + + RFV.getCopyToRegs(ReturnValue, Builder.DAG, Builder.getCurSDLoc(), Chain, + nullptr); + PendingExports.push_back(Chain); + Builder.FuncInfo.ValueMap[CS.getInstruction()] = Reg; + } else { + // Result value will be used in a same basic block. Don't export it or + // perform any explicit register copies. + // We'll replace the actuall call node shortly. gc_result will grab + // this value. + Builder.setValue(CS.getInstruction(), ReturnValue); + } + } else { + // The token value is never used from here on, just generate a poison value + Builder.setValue(CS.getInstruction(), + Builder.DAG.getIntPtrConstant(-1, Builder.getCurSDLoc())); } - assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && - "Expected a callseq node."); - assert(CallEnd->getGluedNode()); - // Step back inside the CALLSEQ - CallNode = CallEnd->getGluedNode(); - return CallNode; + return CallEnd->getOperand(0).getNode(); } /// Callect all gc pointers coming into statepoint intrinsic, clean them up, @@ -325,17 +397,14 @@ static SDNode *lowerCallFromStatepoint(ImmutableStatepoint StatepointSite, /// Relocs - the gc_relocate corresponding to each base/ptr pair /// Elements of this arrays should be in one-to-one correspondence with each /// other i.e Bases[i], Ptrs[i] are from the same gcrelocate call -static void -getIncomingStatepointGCValues(SmallVectorImpl &Bases, - SmallVectorImpl &Ptrs, - SmallVectorImpl &Relocs, - ImmutableStatepoint StatepointSite, - SelectionDAGBuilder &Builder) { - for (GCRelocateOperands relocateOpers : - StatepointSite.getRelocates(StatepointSite)) { +static void getIncomingStatepointGCValues( + SmallVectorImpl &Bases, SmallVectorImpl &Ptrs, + SmallVectorImpl &Relocs, ImmutableStatepoint StatepointSite, + SelectionDAGBuilder &Builder) { + for (GCRelocateOperands relocateOpers : StatepointSite.getRelocates()) { Relocs.push_back(relocateOpers.getUnderlyingCallSite().getInstruction()); - Bases.push_back(relocateOpers.basePtr()); - Ptrs.push_back(relocateOpers.derivedPtr()); + Bases.push_back(relocateOpers.getBasePtr()); + Ptrs.push_back(relocateOpers.getDerivedPtr()); } // Remove any redundant llvm::Values which map to the same SDValue as another @@ -371,7 +440,8 @@ spillIncomingStatepointValue(SDValue Incoming, SDValue Chain, // chaining stores one after another, this may allow // a bit more optimal scheduling for them Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc, - MachinePointerInfo::getFixedStack(Index), + MachinePointerInfo::getFixedStack( + Builder.DAG.getMachineFunction(), Index), false, false, 0); Builder.StatepointLowering.setLocation(Incoming, Loc); @@ -394,17 +464,12 @@ static void lowerIncomingStatepointValue(SDValue Incoming, // 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 - Ops.push_back(Builder.DAG.getTargetConstant(StackMaps::ConstantOp, - Builder.getCurSDLoc(), - MVT::i64)); - Ops.push_back(Builder.DAG.getTargetConstant(C->getSExtValue(), - Builder.getCurSDLoc(), - MVT::i64)); + pushStackMapConstant(Ops, Builder, C->getSExtValue()); } else if (FrameIndexSDNode *FI = dyn_cast(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?) - Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(), + Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(), Incoming.getValueType())); } else { // Otherwise, locate a spill slot and explicitly spill it so it @@ -437,8 +502,8 @@ static void lowerStatepointMetaArgs(SmallVectorImpl &Ops, // be: deopt argument length, deopt arguments.., gc arguments... SmallVector Bases, Ptrs, Relocations; - getIncomingStatepointGCValues(Bases, Ptrs, Relocations, - StatepointSite, Builder); + getIncomingStatepointGCValues(Bases, Ptrs, Relocations, StatepointSite, + Builder); #ifndef NDEBUG // Check that each of the gc pointer and bases we've gotten out of the @@ -448,60 +513,48 @@ static void lowerStatepointMetaArgs(SmallVectorImpl &Ops, // to the GCStrategy from there (yet). GCStrategy &S = Builder.GFI->getStrategy(); for (const Value *V : Bases) { - auto Opt = S.isGCManagedPointer(V); + auto Opt = S.isGCManagedPointer(V->getType()); if (Opt.hasValue()) { assert(Opt.getValue() && "non gc managed base pointer found in statepoint"); } } for (const Value *V : Ptrs) { - auto Opt = S.isGCManagedPointer(V); + auto Opt = S.isGCManagedPointer(V->getType()); if (Opt.hasValue()) { assert(Opt.getValue() && "non gc managed derived pointer found in statepoint"); } } for (const Value *V : Relocations) { - auto Opt = S.isGCManagedPointer(V); + auto Opt = S.isGCManagedPointer(V->getType()); if (Opt.hasValue()) { assert(Opt.getValue() && "non gc managed pointer relocated"); } } #endif - - // Before we actually start lowering (and allocating spill slots for values), // reserve any stack slots which we judge to be profitable to reuse for a // particular value. This is purely an optimization over the code below and // doesn't change semantics at all. It is important for performance that we // reserve slots for both deopt and gc values before lowering either. - for (auto I = StatepointSite.vm_state_begin() + 1, - E = StatepointSite.vm_state_end(); - I != E; ++I) { - Value *V = *I; - SDValue Incoming = Builder.getValue(V); - reservePreviousStackSlotForValue(Incoming, Builder); + for (const Value *V : StatepointSite.vm_state_args()) { + reservePreviousStackSlotForValue(V, Builder); } - for (unsigned i = 0; i < Bases.size() * 2; ++i) { - // Even elements will contain base, odd elements - derived ptr - const Value *V = i % 2 ? Bases[i / 2] : Ptrs[i / 2]; - SDValue Incoming = Builder.getValue(V); - reservePreviousStackSlotForValue(Incoming, Builder); + for (unsigned i = 0; i < Bases.size(); ++i) { + reservePreviousStackSlotForValue(Bases[i], Builder); + reservePreviousStackSlotForValue(Ptrs[i], Builder); } // First, prefix the list with the number of unique values to be // lowered. Note that this is the number of *Values* not the // number of SDValues required to lower them. - const int NumVMSArgs = StatepointSite.numTotalVMSArgs(); - Ops.push_back( Builder.DAG.getTargetConstant(StackMaps::ConstantOp, - Builder.getCurSDLoc(), - MVT::i64)); - Ops.push_back(Builder.DAG.getTargetConstant(NumVMSArgs, Builder.getCurSDLoc(), - MVT::i64)); + const int NumVMSArgs = StatepointSite.getNumTotalVMSArgs(); + pushStackMapConstant(Ops, Builder, NumVMSArgs); - assert(NumVMSArgs + 1 == std::distance(StatepointSite.vm_state_begin(), - StatepointSite.vm_state_end())); + assert(NumVMSArgs == std::distance(StatepointSite.vm_state_begin(), + StatepointSite.vm_state_end())); // The vm state arguments are lowered in an opaque manner. We do // not know what type of values are contained within. We skip the @@ -509,10 +562,7 @@ static void lowerStatepointMetaArgs(SmallVectorImpl &Ops, // explicitly just above. We could have left it in the loop and // not done it explicitly, but it's far easier to understand this // way. - for (auto I = StatepointSite.vm_state_begin() + 1, - E = StatepointSite.vm_state_end(); - I != E; ++I) { - const Value *V = *I; + for (const Value *V : StatepointSite.vm_state_args()) { SDValue Incoming = Builder.getValue(V); lowerIncomingStatepointValue(Incoming, Ops, Builder); } @@ -522,25 +572,61 @@ static void lowerStatepointMetaArgs(SmallVectorImpl &Ops, // 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 < Bases.size() * 2; ++i) { - // Even elements will contain base, odd elements - derived ptr - const Value *V = i % 2 ? Bases[i / 2] : Ptrs[i / 2]; - SDValue Incoming = Builder.getValue(V); - lowerIncomingStatepointValue(Incoming, Ops, Builder); + for (unsigned i = 0; i < Bases.size(); ++i) { + const Value *Base = Bases[i]; + lowerIncomingStatepointValue(Builder.getValue(Base), Ops, Builder); + + const Value *Ptr = Ptrs[i]; + lowerIncomingStatepointValue(Builder.getValue(Ptr), Ops, Builder); } - // If there are any explicit spill slots passed to the statepoint, record + // 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, + // 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 : StatepointSite.gc_args()) { SDValue Incoming = Builder.getValue(V); if (FrameIndexSDNode *FI = dyn_cast(Incoming)) { // This handles allocas as arguments to the statepoint - Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(), + Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(), Incoming.getValueType())); + } + } + // 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 = + StatepointSite.getCallSite().getInstruction(); + FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap = + Builder.FuncInfo.StatepointRelocatedValues[StatepointInstr]; + + for (GCRelocateOperands RelocateOpers : StatepointSite.getRelocates()) { + const Value *V = RelocateOpers.getDerivedPtr(); + SDValue SDV = Builder.getValue(V); + SDValue Loc = Builder.StatepointLowering.getLocation(SDV); + + if (Loc.getNode()) { + SpillMap[V] = cast(Loc)->getIndex(); + } else { + // Record value as visited, but not spilled. This is case for allocas + // and constants. For this values we can avoid emitting spill load while + // visiting corresponding gc_relocate. + // Actually we do not need to record them in this map at all. + // We do this only to check that we are not relocating any unvisited + // value. + SpillMap[V] = None; + + // Default llvm mechanisms for exporting values which are used in + // different basic blocks does not work for gc relocates. + // Note that it would be incorrect to teach llvm that all relocates are + // uses of the corresponding values so that it would automatically + // export them. Relocates of the spilled values does not use original + // value. + if (RelocateOpers.getUnderlyingCallSite().getParent() != + StatepointInstr->getParent()) + Builder.ExportFromCurrentBlock(V); } } } @@ -553,9 +639,8 @@ void SelectionDAGBuilder::visitStatepoint(const CallInst &CI) { LowerStatepoint(ImmutableStatepoint(&CI)); } -void -SelectionDAGBuilder::LowerStatepoint(ImmutableStatepoint ISP, - MachineBasicBlock *LandingPad/*=nullptr*/) { +void SelectionDAGBuilder::LowerStatepoint( + ImmutableStatepoint ISP, const BasicBlock *EHPadBB /*= nullptr*/) { // The basic scheme here is that information about both the original call and // the safepoint is encoded in the CallInst. We create a temporary call and // lower it, then reverse engineer the calling sequence. @@ -567,10 +652,11 @@ SelectionDAGBuilder::LowerStatepoint(ImmutableStatepoint ISP, ImmutableCallSite CS(ISP.getCallSite()); #ifndef NDEBUG - // Consistency check + // Consistency check. Check only relocates in the same basic block as thier + // statepoint. for (const User *U : CS->users()) { const CallInst *Call = cast(U); - if (isGCRelocate(Call)) + if (isGCRelocate(Call) && Call->getParent() == CS.getParent()) StatepointLowering.scheduleRelocCall(*Call); } #endif @@ -582,38 +668,81 @@ SelectionDAGBuilder::LowerStatepoint(ImmutableStatepoint ISP, ISP.verify(); // Check that the associated GCStrategy expects to encounter statepoints. - // TODO: This if should become an assert. For now, we allow the GCStrategy - // to be optional for backwards compatibility. This will only last a short - // period (i.e. a couple of weeks). assert(GFI->getStrategy().useStatepoints() && "GCStrategy does not expect to encounter statepoints"); #endif // Lower statepoint vmstate and gcstate arguments - SmallVector LoweredArgs; - lowerStatepointMetaArgs(LoweredArgs, ISP, *this); + SmallVector LoweredMetaArgs; + lowerStatepointMetaArgs(LoweredMetaArgs, ISP, *this); // Get call node, we will replace it later with statepoint - SDNode *CallNode = lowerCallFromStatepoint(ISP, LandingPad, *this); + SDNode *CallNode = + lowerCallFromStatepoint(ISP, EHPadBB, *this, PendingExports); + + // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END + // nodes with all the appropriate arguments and return values. + + // Call Node: Chain, Target, {Args}, RegMask, [Glue] + SDValue Chain = CallNode->getOperand(0); + + SDValue Glue; + bool CallHasIncomingGlue = CallNode->getGluedNode(); + if (CallHasIncomingGlue) { + // Glue is always last operand + Glue = CallNode->getOperand(CallNode->getNumOperands() - 1); + } + + // Build the GC_TRANSITION_START node if necessary. + // + // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the + // order in which they appear in the call to the statepoint intrinsic. If + // any of the operands is a pointer-typed, that operand is immediately + // followed by a SRCVALUE for the pointer that may be used during lowering + // (e.g. to form MachinePointerInfo values for loads/stores). + const bool IsGCTransition = + (ISP.getFlags() & (uint64_t)StatepointFlags::GCTransition) == + (uint64_t)StatepointFlags::GCTransition; + if (IsGCTransition) { + SmallVector TSOps; + + // Add chain + TSOps.push_back(Chain); + + // Add GC transition arguments + for (const Value *V : ISP.gc_transition_args()) { + TSOps.push_back(getValue(V)); + if (V->getType()->isPointerTy()) + TSOps.push_back(DAG.getSrcValue(V)); + } - // Construct the actual STATEPOINT node with all the appropriate arguments - // and return values. + // Add glue if necessary + if (CallHasIncomingGlue) + TSOps.push_back(Glue); + + SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); + + SDValue GCTransitionStart = + DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps); + + Chain = GCTransitionStart.getValue(0); + Glue = GCTransitionStart.getValue(1); + } // TODO: Currently, all of these operands are being marked as read/write in // PrologEpilougeInserter.cpp, we should special case the VMState arguments // and flags to be read-only. SmallVector Ops; + // Add the and constants. + Ops.push_back(DAG.getTargetConstant(ISP.getID(), getCurSDLoc(), MVT::i64)); + Ops.push_back( + DAG.getTargetConstant(ISP.getNumPatchBytes(), getCurSDLoc(), MVT::i32)); + // Calculate and push starting position of vmstate arguments - // Call Node: Chain, Target, {Args}, RegMask, [Glue] - SDValue Glue; - if (CallNode->getGluedNode()) { - // Glue is always last operand - Glue = CallNode->getOperand(CallNode->getNumOperands() - 1); - } // Get number of arguments incoming directly into call node unsigned NumCallRegArgs = - CallNode->getNumOperands() - (Glue.getNode() ? 4 : 3); + CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3); Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32)); // Add call target @@ -623,30 +752,30 @@ SelectionDAGBuilder::LowerStatepoint(ImmutableStatepoint ISP, // Add call arguments // Get position of register mask in the call SDNode::op_iterator RegMaskIt; - if (Glue.getNode()) + if (CallHasIncomingGlue) RegMaskIt = CallNode->op_end() - 2; else RegMaskIt = CallNode->op_end() - 1; Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt); - // Add a leading constant argument with the Flags and the calling convention - // masked together - CallingConv::ID CallConv = CS.getCallingConv(); - int Flags = cast(CS.getArgument(2))->getZExtValue(); - assert(Flags == 0 && "not expected to be used"); - Ops.push_back(DAG.getTargetConstant(StackMaps::ConstantOp, getCurSDLoc(), - MVT::i64)); - Ops.push_back(DAG.getTargetConstant(Flags | ((unsigned)CallConv << 1), - getCurSDLoc(), MVT::i64)); + // Add a constant argument for the calling convention + pushStackMapConstant(Ops, *this, CS.getCallingConv()); + + // Add a constant argument for the flags + uint64_t Flags = ISP.getFlags(); + assert( + ((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0) + && "unknown flag used"); + pushStackMapConstant(Ops, *this, Flags); // Insert all vmstate and gcstate arguments - Ops.insert(Ops.end(), LoweredArgs.begin(), LoweredArgs.end()); + Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end()); // Add register mask from call node Ops.push_back(*RegMaskIt); // Add chain - Ops.push_back(CallNode->getOperand(0)); + Ops.push_back(Chain); // Same for the glue, but we add it only if original call had it if (Glue.getNode()) @@ -656,12 +785,42 @@ SelectionDAGBuilder::LowerStatepoint(ImmutableStatepoint ISP, // 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); + SDNode *StatepointMCNode = + DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops); + + 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) { + SmallVector TEOps; + + // Add chain + TEOps.push_back(SDValue(StatepointMCNode, 0)); + + // Add GC transition arguments + for (const Value *V : ISP.gc_transition_args()) { + TEOps.push_back(getValue(V)); + if (V->getType()->isPointerTy()) + TEOps.push_back(DAG.getSrcValue(V)); + } + + // Add glue + TEOps.push_back(SDValue(StatepointMCNode, 1)); + + SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); + + SDValue GCTransitionStart = + DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps); + + SinkNode = GCTransitionStart.getNode(); + } // Replace original call - DAG.ReplaceAllUsesWith(CallNode, StatepointMCNode); // This may update Root - // Remove originall call node + DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root + // Remove original call node DAG.DeleteNode(CallNode); // DON'T set the root - under the assumption that it's already set past the @@ -678,66 +837,75 @@ void SelectionDAGBuilder::visitGCResult(const CallInst &CI) { // The result value of the gc_result is simply the result of the actual // call. We've already emitted this, so just grab the value. Instruction *I = cast(CI.getArgOperand(0)); - assert(isStatepoint(I) && - "first argument must be a statepoint token"); + assert(isStatepoint(I) && "first argument must be a statepoint token"); - if (isa(I)) { - // For invokes we should have stored call result in a virtual register. + if (I->getParent() != CI.getParent()) { + // Statepoint is in different basic block so we should have stored call + // result in a virtual register. // We can not use default getValue() functionality to copy value from this // register because statepoint and actuall call return types can be // different, and getValue() will use CopyFromReg of the wrong type, // which is always i32 in our case. PointerType *CalleeType = cast( - ImmutableStatepoint(I).actualCallee()->getType()); - Type *RetTy = cast( - CalleeType->getElementType())->getReturnType(); + ImmutableStatepoint(I).getCalledValue()->getType()); + Type *RetTy = + cast(CalleeType->getElementType())->getReturnType(); SDValue CopyFromReg = getCopyFromRegs(I, RetTy); assert(CopyFromReg.getNode()); setValue(&CI, CopyFromReg); - } - else { + } else { setValue(&CI, getValue(I)); } } void SelectionDAGBuilder::visitGCRelocate(const CallInst &CI) { + GCRelocateOperands RelocateOpers(&CI); + #ifndef NDEBUG // Consistency check - StatepointLowering.relocCallVisited(CI); + // We skip this check for relocates not in the same basic block as thier + // statepoint. It would be too expensive to preserve validation info through + // different basic blocks. + if (RelocateOpers.getStatepoint()->getParent() == CI.getParent()) { + StatepointLowering.relocCallVisited(CI); + } #endif - GCRelocateOperands relocateOpers(&CI); - SDValue SD = getValue(relocateOpers.derivedPtr()); + const Value *DerivedPtr = RelocateOpers.getDerivedPtr(); + SDValue SD = getValue(DerivedPtr); + + FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap = + FuncInfo.StatepointRelocatedValues[RelocateOpers.getStatepoint()]; - if (isa(SD) || isa(SD)) { - // We didn't need to spill these special cases (constants and allocas). - // See the handling in spillIncomingValueForStatepoint for detail. + // We should have recorded location for this pointer + assert(SpillMap.count(DerivedPtr) && "Relocating not lowered gc value"); + Optional DerivedPtrLocation = SpillMap[DerivedPtr]; + + // We didn't need to spill these special cases (constants and allocas). + // See the handling in spillIncomingValueForStatepoint for detail. + if (!DerivedPtrLocation) { setValue(&CI, SD); return; } - SDValue Loc = StatepointLowering.getRelocLocation(SD); - // Emit new load if we did not emit it before - if (!Loc.getNode()) { - SDValue SpillSlot = StatepointLowering.getLocation(SD); - int FI = cast(SpillSlot)->getIndex(); + SDValue SpillSlot = DAG.getTargetFrameIndex(*DerivedPtrLocation, + SD.getValueType()); - // Be conservative: flush all pending loads - // TODO: Probably we can be less restrictive on this, - // it may allow more scheduling opprtunities - SDValue Chain = getRoot(); + // Be conservative: flush all pending loads + // TODO: Probably we can be less restrictive on this, + // it may allow more scheduling opportunities. + SDValue Chain = getRoot(); - Loc = DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, - SpillSlot, MachinePointerInfo::getFixedStack(FI), false, - false, false, 0); + SDValue SpillLoad = + DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, SpillSlot, + MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), + *DerivedPtrLocation), + false, false, false, 0); - StatepointLowering.setRelocLocation(SD, Loc); + // Again, be conservative, don't emit pending loads + DAG.setRoot(SpillLoad.getValue(1)); - // Again, be conservative, don't emit pending loads - DAG.setRoot(Loc.getValue(1)); - } - - assert(Loc.getNode()); - setValue(&CI, Loc); + assert(SpillLoad.getNode()); + setValue(&CI, SpillLoad); }