1 //===-- StatepointLowering.cpp - SDAGBuilder's statepoint code -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file includes support code use by SelectionDAGBuilder when lowering a
11 // statepoint sequence in SelectionDAG IR.
13 //===----------------------------------------------------------------------===//
15 #include "StatepointLowering.h"
16 #include "SelectionDAGBuilder.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/CodeGen/FunctionLoweringInfo.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/GCMetadata.h"
22 #include "llvm/CodeGen/GCStrategy.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/CodeGen/StackMaps.h"
25 #include "llvm/IR/CallingConv.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/Statepoint.h"
30 #include "llvm/Target/TargetLowering.h"
34 #define DEBUG_TYPE "statepoint-lowering"
36 STATISTIC(NumSlotsAllocatedForStatepoints,
37 "Number of stack slots allocated for statepoints");
38 STATISTIC(NumOfStatepoints, "Number of statepoint nodes encountered");
39 STATISTIC(StatepointMaxSlotsRequired,
40 "Maximum number of stack slots required for a singe statepoint");
42 static void pushStackMapConstant(SmallVectorImpl<SDValue>& Ops,
43 SelectionDAGBuilder &Builder, uint64_t Value) {
44 SDLoc L = Builder.getCurSDLoc();
45 Ops.push_back(Builder.DAG.getTargetConstant(StackMaps::ConstantOp, L,
47 Ops.push_back(Builder.DAG.getTargetConstant(Value, L, MVT::i64));
50 void StatepointLoweringState::startNewStatepoint(SelectionDAGBuilder &Builder) {
52 assert(PendingGCRelocateCalls.empty() &&
53 "Trying to visit statepoint before finished processing previous one");
55 NextSlotToAllocate = 0;
56 // Need to resize this on each safepoint - we need the two to stay in
57 // sync and the clear patterns of a SelectionDAGBuilder have no relation
58 // to FunctionLoweringInfo.
59 AllocatedStackSlots.resize(Builder.FuncInfo.StatepointStackSlots.size());
60 for (size_t i = 0; i < AllocatedStackSlots.size(); i++) {
61 AllocatedStackSlots[i] = false;
65 void StatepointLoweringState::clear() {
67 AllocatedStackSlots.clear();
68 assert(PendingGCRelocateCalls.empty() &&
69 "cleared before statepoint sequence completed");
73 StatepointLoweringState::allocateStackSlot(EVT ValueType,
74 SelectionDAGBuilder &Builder) {
76 NumSlotsAllocatedForStatepoints++;
78 // The basic scheme here is to first look for a previously created stack slot
79 // which is not in use (accounting for the fact arbitrary slots may already
80 // be reserved), or to create a new stack slot and use it.
82 // If this doesn't succeed in 40000 iterations, something is seriously wrong
83 for (int i = 0; i < 40000; i++) {
84 assert(Builder.FuncInfo.StatepointStackSlots.size() ==
85 AllocatedStackSlots.size() &&
87 const size_t NumSlots = AllocatedStackSlots.size();
88 assert(NextSlotToAllocate <= NumSlots && "broken invariant");
90 if (NextSlotToAllocate >= NumSlots) {
91 assert(NextSlotToAllocate == NumSlots);
93 if (NumSlots + 1 > StatepointMaxSlotsRequired) {
94 StatepointMaxSlotsRequired = NumSlots + 1;
97 SDValue SpillSlot = Builder.DAG.CreateStackTemporary(ValueType);
98 const unsigned FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
99 auto *MFI = Builder.DAG.getMachineFunction().getFrameInfo();
100 MFI->markAsStatepointSpillSlotObjectIndex(FI);
102 Builder.FuncInfo.StatepointStackSlots.push_back(FI);
103 AllocatedStackSlots.push_back(true);
106 if (!AllocatedStackSlots[NextSlotToAllocate]) {
107 const int FI = Builder.FuncInfo.StatepointStackSlots[NextSlotToAllocate];
108 AllocatedStackSlots[NextSlotToAllocate] = true;
109 return Builder.DAG.getFrameIndex(FI, ValueType);
111 // Note: We deliberately choose to advance this only on the failing path.
112 // Doing so on the succeeding path involves a bit of complexity that caused
113 // a minor bug previously. Unless performance shows this matters, please
114 // keep this code as simple as possible.
115 NextSlotToAllocate++;
117 llvm_unreachable("infinite loop?");
120 /// Utility function for reservePreviousStackSlotForValue. Tries to find
121 /// stack slot index to which we have spilled value for previous statepoints.
122 /// LookUpDepth specifies maximum DFS depth this function is allowed to look.
123 static Optional<int> findPreviousSpillSlot(const Value *Val,
124 SelectionDAGBuilder &Builder,
126 // Can not look any further - give up now
127 if (LookUpDepth <= 0)
128 return Optional<int>();
130 // Spill location is known for gc relocates
131 if (isGCRelocate(Val)) {
132 GCRelocateOperands RelocOps(cast<Instruction>(Val));
134 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
135 Builder.FuncInfo.StatepointRelocatedValues[RelocOps.getStatepoint()];
137 auto It = SpillMap.find(RelocOps.getDerivedPtr());
138 if (It == SpillMap.end())
139 return Optional<int>();
144 // Look through bitcast instructions.
145 if (const BitCastInst *Cast = dyn_cast<BitCastInst>(Val)) {
146 return findPreviousSpillSlot(Cast->getOperand(0), Builder, LookUpDepth - 1);
149 // Look through phi nodes
150 // All incoming values should have same known stack slot, otherwise result
152 if (const PHINode *Phi = dyn_cast<PHINode>(Val)) {
153 Optional<int> MergedResult = None;
155 for (auto &IncomingValue : Phi->incoming_values()) {
156 Optional<int> SpillSlot =
157 findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth - 1);
158 if (!SpillSlot.hasValue())
159 return Optional<int>();
161 if (MergedResult.hasValue() && *MergedResult != *SpillSlot)
162 return Optional<int>();
164 MergedResult = SpillSlot;
169 // TODO: We can do better for PHI nodes. In cases like this:
170 // ptr = phi(relocated_pointer, not_relocated_pointer)
172 // We will return that stack slot for ptr is unknown. And later we might
173 // assign different stack slots for ptr and relocated_pointer. This limits
174 // llvm's ability to remove redundant stores.
175 // Unfortunately it's hard to accomplish in current infrastructure.
176 // We use this function to eliminate spill store completely, while
177 // in example we still need to emit store, but instead of any location
178 // we need to use special "preferred" location.
180 // TODO: handle simple updates. If a value is modified and the original
181 // value is no longer live, it would be nice to put the modified value in the
182 // same slot. This allows folding of the memory accesses for some
183 // instructions types (like an increment).
187 // However we need to be careful for cases like this:
191 // Here we want to reserve spill slot for 'i', but not for 'i+1'. If we just
192 // put handling of simple modifications in this function like it's done
193 // for bitcasts we might end up reserving i's slot for 'i+1' because order in
194 // which we visit values is unspecified.
196 // Don't know any information about this instruction
197 return Optional<int>();
200 /// Try to find existing copies of the incoming values in stack slots used for
201 /// statepoint spilling. If we can find a spill slot for the incoming value,
202 /// mark that slot as allocated, and reuse the same slot for this safepoint.
203 /// This helps to avoid series of loads and stores that only serve to reshuffle
204 /// values on the stack between calls.
205 static void reservePreviousStackSlotForValue(const Value *IncomingValue,
206 SelectionDAGBuilder &Builder) {
208 SDValue Incoming = Builder.getValue(IncomingValue);
210 if (isa<ConstantSDNode>(Incoming) || isa<FrameIndexSDNode>(Incoming)) {
211 // We won't need to spill this, so no need to check for previously
212 // allocated stack slots
216 SDValue OldLocation = Builder.StatepointLowering.getLocation(Incoming);
217 if (OldLocation.getNode())
218 // duplicates in input
221 const int LookUpDepth = 6;
222 Optional<int> Index =
223 findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth);
224 if (!Index.hasValue())
227 auto Itr = std::find(Builder.FuncInfo.StatepointStackSlots.begin(),
228 Builder.FuncInfo.StatepointStackSlots.end(), *Index);
229 assert(Itr != Builder.FuncInfo.StatepointStackSlots.end() &&
230 "value spilled to the unknown stack slot");
232 // This is one of our dedicated lowering slots
234 std::distance(Builder.FuncInfo.StatepointStackSlots.begin(), Itr);
235 if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) {
236 // stack slot already assigned to someone else, can't use it!
237 // TODO: currently we reserve space for gc arguments after doing
238 // normal allocation for deopt arguments. We should reserve for
239 // _all_ deopt and gc arguments, then start allocating. This
240 // will prevent some moves being inserted when vm state changes,
241 // but gc state doesn't between two calls.
244 // Reserve this stack slot
245 Builder.StatepointLowering.reserveStackSlot(Offset);
247 // Cache this slot so we find it when going through the normal
249 SDValue Loc = Builder.DAG.getTargetFrameIndex(*Index, Incoming.getValueType());
250 Builder.StatepointLowering.setLocation(Incoming, Loc);
253 /// Remove any duplicate (as SDValues) from the derived pointer pairs. This
254 /// is not required for correctness. It's purpose is to reduce the size of
255 /// StackMap section. It has no effect on the number of spill slots required
256 /// or the actual lowering.
257 static void removeDuplicatesGCPtrs(SmallVectorImpl<const Value *> &Bases,
258 SmallVectorImpl<const Value *> &Ptrs,
259 SmallVectorImpl<const Value *> &Relocs,
260 SelectionDAGBuilder &Builder) {
262 // This is horribly inefficient, but I don't care right now
263 SmallSet<SDValue, 64> Seen;
265 SmallVector<const Value *, 64> NewBases, NewPtrs, NewRelocs;
266 for (size_t i = 0; i < Ptrs.size(); i++) {
267 SDValue SD = Builder.getValue(Ptrs[i]);
268 // Only add non-duplicates
269 if (Seen.count(SD) == 0) {
270 NewBases.push_back(Bases[i]);
271 NewPtrs.push_back(Ptrs[i]);
272 NewRelocs.push_back(Relocs[i]);
276 assert(Bases.size() >= NewBases.size());
277 assert(Ptrs.size() >= NewPtrs.size());
278 assert(Relocs.size() >= NewRelocs.size());
282 assert(Ptrs.size() == Bases.size());
283 assert(Ptrs.size() == Relocs.size());
286 /// Extract call from statepoint, lower it and return pointer to the
287 /// call node. Also update NodeMap so that getValue(statepoint) will
288 /// reference lowered call result
290 lowerCallFromStatepoint(ImmutableStatepoint ISP, const BasicBlock *EHPadBB,
291 SelectionDAGBuilder &Builder,
292 SmallVectorImpl<SDValue> &PendingExports) {
294 ImmutableCallSite CS(ISP.getCallSite());
296 SDValue ActualCallee;
298 if (ISP.getNumPatchBytes() > 0) {
299 // If we've been asked to emit a nop sequence instead of a call instruction
300 // for this statepoint then don't lower the call target, but use a constant
301 // `null` instead. Not lowering the call target lets statepoint clients get
302 // away without providing a physical address for the symbolic call target at
305 const auto &TLI = Builder.DAG.getTargetLoweringInfo();
306 const auto &DL = Builder.DAG.getDataLayout();
308 unsigned AS = ISP.getCalledValue()->getType()->getPointerAddressSpace();
309 ActualCallee = Builder.DAG.getConstant(0, Builder.getCurSDLoc(),
310 TLI.getPointerTy(DL, AS));
312 ActualCallee = Builder.getValue(ISP.getCalledValue());
314 assert(CS.getCallingConv() != CallingConv::AnyReg &&
315 "anyregcc is not supported on statepoints!");
317 Type *DefTy = ISP.getActualReturnType();
318 bool HasDef = !DefTy->isVoidTy();
320 SDValue ReturnValue, CallEndVal;
321 std::tie(ReturnValue, CallEndVal) = Builder.lowerCallOperands(
322 ISP.getCallSite(), ImmutableStatepoint::CallArgsBeginPos,
323 ISP.getNumCallArgs(), ActualCallee, DefTy, EHPadBB,
324 false /* IsPatchPoint */);
326 SDNode *CallEnd = CallEndVal.getNode();
328 // Get a call instruction from the call sequence chain. Tail calls are not
329 // allowed. The following code is essentially reverse engineering X86's
332 // We are expecting DAG to have the following form:
334 // ch = eh_label (only in case of invoke statepoint)
335 // ch, glue = callseq_start ch
336 // ch, glue = X86::Call ch, glue
337 // ch, glue = callseq_end ch, glue
338 // get_return_value ch, glue
340 // get_return_value can either be a sequence of CopyFromReg instructions
341 // to grab the return value from the return register(s), or it can be a LOAD
342 // to load a value returned by reference via a stack slot.
345 if (CallEnd->getOpcode() == ISD::LOAD)
346 CallEnd = CallEnd->getOperand(0).getNode();
348 while (CallEnd->getOpcode() == ISD::CopyFromReg)
349 CallEnd = CallEnd->getOperand(0).getNode();
352 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
354 // Export the result value if needed
355 const Instruction *GCResult = ISP.getGCResult();
356 if (HasDef && GCResult) {
357 if (GCResult->getParent() != CS.getParent()) {
358 // Result value will be used in a different basic block so we need to
360 // Default exporting mechanism will not work here because statepoint call
361 // has a different type than the actual call. It means that by default
362 // llvm will create export register of the wrong type (always i32 in our
363 // case). So instead we need to create export register with correct type
365 // TODO: To eliminate this problem we can remove gc.result intrinsics
366 // completely and make statepoint call to return a tuple.
367 unsigned Reg = Builder.FuncInfo.CreateRegs(ISP.getActualReturnType());
369 *Builder.DAG.getContext(), Builder.DAG.getTargetLoweringInfo(),
370 Builder.DAG.getDataLayout(), Reg, ISP.getActualReturnType());
371 SDValue Chain = Builder.DAG.getEntryNode();
373 RFV.getCopyToRegs(ReturnValue, Builder.DAG, Builder.getCurSDLoc(), Chain,
375 PendingExports.push_back(Chain);
376 Builder.FuncInfo.ValueMap[CS.getInstruction()] = Reg;
378 // Result value will be used in a same basic block. Don't export it or
379 // perform any explicit register copies.
380 // We'll replace the actuall call node shortly. gc_result will grab
382 Builder.setValue(CS.getInstruction(), ReturnValue);
385 // The token value is never used from here on, just generate a poison value
386 Builder.setValue(CS.getInstruction(),
387 Builder.DAG.getIntPtrConstant(-1, Builder.getCurSDLoc()));
390 return CallEnd->getOperand(0).getNode();
393 /// Callect all gc pointers coming into statepoint intrinsic, clean them up,
394 /// and return two arrays:
395 /// Bases - base pointers incoming to this statepoint
396 /// Ptrs - derived pointers incoming to this statepoint
397 /// Relocs - the gc_relocate corresponding to each base/ptr pair
398 /// Elements of this arrays should be in one-to-one correspondence with each
399 /// other i.e Bases[i], Ptrs[i] are from the same gcrelocate call
400 static void getIncomingStatepointGCValues(
401 SmallVectorImpl<const Value *> &Bases, SmallVectorImpl<const Value *> &Ptrs,
402 SmallVectorImpl<const Value *> &Relocs, ImmutableStatepoint StatepointSite,
403 SelectionDAGBuilder &Builder) {
404 for (GCRelocateOperands relocateOpers : StatepointSite.getRelocates()) {
405 Relocs.push_back(relocateOpers.getUnderlyingCallSite().getInstruction());
406 Bases.push_back(relocateOpers.getBasePtr());
407 Ptrs.push_back(relocateOpers.getDerivedPtr());
410 // Remove any redundant llvm::Values which map to the same SDValue as another
411 // input. Also has the effect of removing duplicates in the original
412 // llvm::Value input list as well. This is a useful optimization for
413 // reducing the size of the StackMap section. It has no other impact.
414 removeDuplicatesGCPtrs(Bases, Ptrs, Relocs, Builder);
416 assert(Bases.size() == Ptrs.size() && Ptrs.size() == Relocs.size());
419 /// Spill a value incoming to the statepoint. It might be either part of
421 /// or gcstate. In both cases unconditionally spill it on the stack unless it
422 /// is a null constant. Return pair with first element being frame index
423 /// containing saved value and second element with outgoing chain from the
425 static std::pair<SDValue, SDValue>
426 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
427 SelectionDAGBuilder &Builder) {
428 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
430 // Emit new store if we didn't do it for this ptr before
431 if (!Loc.getNode()) {
432 Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
434 assert(isa<FrameIndexSDNode>(Loc));
435 int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
436 // We use TargetFrameIndex so that isel will not select it into LEA
437 Loc = Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
439 // TODO: We can create TokenFactor node instead of
440 // chaining stores one after another, this may allow
441 // a bit more optimal scheduling for them
442 Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
443 MachinePointerInfo::getFixedStack(
444 Builder.DAG.getMachineFunction(), Index),
447 Builder.StatepointLowering.setLocation(Incoming, Loc);
450 assert(Loc.getNode());
451 return std::make_pair(Loc, Chain);
454 /// Lower a single value incoming to a statepoint node. This value can be
455 /// either a deopt value or a gc value, the handling is the same. We special
456 /// case constants and allocas, then fall back to spilling if required.
457 static void lowerIncomingStatepointValue(SDValue Incoming,
458 SmallVectorImpl<SDValue> &Ops,
459 SelectionDAGBuilder &Builder) {
460 SDValue Chain = Builder.getRoot();
462 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
463 // If the original value was a constant, make sure it gets recorded as
464 // such in the stackmap. This is required so that the consumer can
465 // parse any internal format to the deopt state. It also handles null
466 // pointers and other constant pointers in GC states
467 pushStackMapConstant(Ops, Builder, C->getSExtValue());
468 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
469 // This handles allocas as arguments to the statepoint (this is only
470 // really meaningful for a deopt value. For GC, we'd be trying to
471 // relocate the address of the alloca itself?)
472 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
473 Incoming.getValueType()));
475 // Otherwise, locate a spill slot and explicitly spill it so it
476 // can be found by the runtime later. We currently do not support
477 // tracking values through callee saved registers to their eventual
478 // spill location. This would be a useful optimization, but would
479 // need to be optional since it requires a lot of complexity on the
480 // runtime side which not all would support.
481 std::pair<SDValue, SDValue> Res =
482 spillIncomingStatepointValue(Incoming, Chain, Builder);
483 Ops.push_back(Res.first);
487 Builder.DAG.setRoot(Chain);
490 /// Lower deopt state and gc pointer arguments of the statepoint. The actual
491 /// lowering is described in lowerIncomingStatepointValue. This function is
492 /// responsible for lowering everything in the right position and playing some
493 /// tricks to avoid redundant stack manipulation where possible. On
494 /// completion, 'Ops' will contain ready to use operands for machine code
495 /// statepoint. The chain nodes will have already been created and the DAG root
496 /// will be set to the last value spilled (if any were).
497 static void lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
498 ImmutableStatepoint StatepointSite,
499 SelectionDAGBuilder &Builder) {
501 // Lower the deopt and gc arguments for this statepoint. Layout will
502 // be: deopt argument length, deopt arguments.., gc arguments...
504 SmallVector<const Value *, 64> Bases, Ptrs, Relocations;
505 getIncomingStatepointGCValues(Bases, Ptrs, Relocations, StatepointSite,
509 // Check that each of the gc pointer and bases we've gotten out of the
510 // safepoint is something the strategy thinks might be a pointer into the GC
511 // heap. This is basically just here to help catch errors during statepoint
512 // insertion. TODO: This should actually be in the Verifier, but we can't get
513 // to the GCStrategy from there (yet).
514 GCStrategy &S = Builder.GFI->getStrategy();
515 for (const Value *V : Bases) {
516 auto Opt = S.isGCManagedPointer(V->getType());
517 if (Opt.hasValue()) {
518 assert(Opt.getValue() &&
519 "non gc managed base pointer found in statepoint");
522 for (const Value *V : Ptrs) {
523 auto Opt = S.isGCManagedPointer(V->getType());
524 if (Opt.hasValue()) {
525 assert(Opt.getValue() &&
526 "non gc managed derived pointer found in statepoint");
529 for (const Value *V : Relocations) {
530 auto Opt = S.isGCManagedPointer(V->getType());
531 if (Opt.hasValue()) {
532 assert(Opt.getValue() && "non gc managed pointer relocated");
537 // Before we actually start lowering (and allocating spill slots for values),
538 // reserve any stack slots which we judge to be profitable to reuse for a
539 // particular value. This is purely an optimization over the code below and
540 // doesn't change semantics at all. It is important for performance that we
541 // reserve slots for both deopt and gc values before lowering either.
542 for (const Value *V : StatepointSite.vm_state_args()) {
543 reservePreviousStackSlotForValue(V, Builder);
545 for (unsigned i = 0; i < Bases.size(); ++i) {
546 reservePreviousStackSlotForValue(Bases[i], Builder);
547 reservePreviousStackSlotForValue(Ptrs[i], Builder);
550 // First, prefix the list with the number of unique values to be
551 // lowered. Note that this is the number of *Values* not the
552 // number of SDValues required to lower them.
553 const int NumVMSArgs = StatepointSite.getNumTotalVMSArgs();
554 pushStackMapConstant(Ops, Builder, NumVMSArgs);
556 assert(NumVMSArgs == std::distance(StatepointSite.vm_state_begin(),
557 StatepointSite.vm_state_end()));
559 // The vm state arguments are lowered in an opaque manner. We do
560 // not know what type of values are contained within. We skip the
561 // first one since that happens to be the total number we lowered
562 // explicitly just above. We could have left it in the loop and
563 // not done it explicitly, but it's far easier to understand this
565 for (const Value *V : StatepointSite.vm_state_args()) {
566 SDValue Incoming = Builder.getValue(V);
567 lowerIncomingStatepointValue(Incoming, Ops, Builder);
570 // Finally, go ahead and lower all the gc arguments. There's no prefixed
571 // length for this one. After lowering, we'll have the base and pointer
572 // arrays interwoven with each (lowered) base pointer immediately followed by
573 // it's (lowered) derived pointer. i.e
574 // (base[0], ptr[0], base[1], ptr[1], ...)
575 for (unsigned i = 0; i < Bases.size(); ++i) {
576 const Value *Base = Bases[i];
577 lowerIncomingStatepointValue(Builder.getValue(Base), Ops, Builder);
579 const Value *Ptr = Ptrs[i];
580 lowerIncomingStatepointValue(Builder.getValue(Ptr), Ops, Builder);
583 // If there are any explicit spill slots passed to the statepoint, record
584 // them, but otherwise do not do anything special. These are user provided
585 // allocas and give control over placement to the consumer. In this case,
586 // it is the contents of the slot which may get updated, not the pointer to
588 for (Value *V : StatepointSite.gc_args()) {
589 SDValue Incoming = Builder.getValue(V);
590 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
591 // This handles allocas as arguments to the statepoint
592 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
593 Incoming.getValueType()));
597 // Record computed locations for all lowered values.
598 // This can not be embedded in lowering loops as we need to record *all*
599 // values, while previous loops account only values with unique SDValues.
600 const Instruction *StatepointInstr =
601 StatepointSite.getCallSite().getInstruction();
602 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
603 Builder.FuncInfo.StatepointRelocatedValues[StatepointInstr];
605 for (GCRelocateOperands RelocateOpers : StatepointSite.getRelocates()) {
606 const Value *V = RelocateOpers.getDerivedPtr();
607 SDValue SDV = Builder.getValue(V);
608 SDValue Loc = Builder.StatepointLowering.getLocation(SDV);
611 SpillMap[V] = cast<FrameIndexSDNode>(Loc)->getIndex();
613 // Record value as visited, but not spilled. This is case for allocas
614 // and constants. For this values we can avoid emitting spill load while
615 // visiting corresponding gc_relocate.
616 // Actually we do not need to record them in this map at all.
617 // We do this only to check that we are not relocating any unvisited
621 // Default llvm mechanisms for exporting values which are used in
622 // different basic blocks does not work for gc relocates.
623 // Note that it would be incorrect to teach llvm that all relocates are
624 // uses of the corresponding values so that it would automatically
625 // export them. Relocates of the spilled values does not use original
627 if (RelocateOpers.getUnderlyingCallSite().getParent() !=
628 StatepointInstr->getParent())
629 Builder.ExportFromCurrentBlock(V);
634 void SelectionDAGBuilder::visitStatepoint(const CallInst &CI) {
635 // Check some preconditions for sanity
636 assert(isStatepoint(&CI) &&
637 "function called must be the statepoint function");
639 LowerStatepoint(ImmutableStatepoint(&CI));
642 void SelectionDAGBuilder::LowerStatepoint(
643 ImmutableStatepoint ISP, const BasicBlock *EHPadBB /*= nullptr*/) {
644 // The basic scheme here is that information about both the original call and
645 // the safepoint is encoded in the CallInst. We create a temporary call and
646 // lower it, then reverse engineer the calling sequence.
650 StatepointLowering.startNewStatepoint(*this);
652 ImmutableCallSite CS(ISP.getCallSite());
655 // Consistency check. Check only relocates in the same basic block as thier
657 for (const User *U : CS->users()) {
658 const CallInst *Call = cast<CallInst>(U);
659 if (isGCRelocate(Call) && Call->getParent() == CS.getParent())
660 StatepointLowering.scheduleRelocCall(*Call);
665 // If this is a malformed statepoint, report it early to simplify debugging.
666 // This should catch any IR level mistake that's made when constructing or
667 // transforming statepoints.
670 // Check that the associated GCStrategy expects to encounter statepoints.
671 assert(GFI->getStrategy().useStatepoints() &&
672 "GCStrategy does not expect to encounter statepoints");
675 // Lower statepoint vmstate and gcstate arguments
676 SmallVector<SDValue, 10> LoweredMetaArgs;
677 lowerStatepointMetaArgs(LoweredMetaArgs, ISP, *this);
679 // Get call node, we will replace it later with statepoint
681 lowerCallFromStatepoint(ISP, EHPadBB, *this, PendingExports);
683 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
684 // nodes with all the appropriate arguments and return values.
686 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
687 SDValue Chain = CallNode->getOperand(0);
690 bool CallHasIncomingGlue = CallNode->getGluedNode();
691 if (CallHasIncomingGlue) {
692 // Glue is always last operand
693 Glue = CallNode->getOperand(CallNode->getNumOperands() - 1);
696 // Build the GC_TRANSITION_START node if necessary.
698 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
699 // order in which they appear in the call to the statepoint intrinsic. If
700 // any of the operands is a pointer-typed, that operand is immediately
701 // followed by a SRCVALUE for the pointer that may be used during lowering
702 // (e.g. to form MachinePointerInfo values for loads/stores).
703 const bool IsGCTransition =
704 (ISP.getFlags() & (uint64_t)StatepointFlags::GCTransition) ==
705 (uint64_t)StatepointFlags::GCTransition;
706 if (IsGCTransition) {
707 SmallVector<SDValue, 8> TSOps;
710 TSOps.push_back(Chain);
712 // Add GC transition arguments
713 for (const Value *V : ISP.gc_transition_args()) {
714 TSOps.push_back(getValue(V));
715 if (V->getType()->isPointerTy())
716 TSOps.push_back(DAG.getSrcValue(V));
719 // Add glue if necessary
720 if (CallHasIncomingGlue)
721 TSOps.push_back(Glue);
723 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
725 SDValue GCTransitionStart =
726 DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps);
728 Chain = GCTransitionStart.getValue(0);
729 Glue = GCTransitionStart.getValue(1);
732 // TODO: Currently, all of these operands are being marked as read/write in
733 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
734 // and flags to be read-only.
735 SmallVector<SDValue, 40> Ops;
737 // Add the <id> and <numBytes> constants.
738 Ops.push_back(DAG.getTargetConstant(ISP.getID(), getCurSDLoc(), MVT::i64));
740 DAG.getTargetConstant(ISP.getNumPatchBytes(), getCurSDLoc(), MVT::i32));
742 // Calculate and push starting position of vmstate arguments
743 // Get number of arguments incoming directly into call node
744 unsigned NumCallRegArgs =
745 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
746 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32));
749 SDValue CallTarget = SDValue(CallNode->getOperand(1).getNode(), 0);
750 Ops.push_back(CallTarget);
752 // Add call arguments
753 // Get position of register mask in the call
754 SDNode::op_iterator RegMaskIt;
755 if (CallHasIncomingGlue)
756 RegMaskIt = CallNode->op_end() - 2;
758 RegMaskIt = CallNode->op_end() - 1;
759 Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt);
761 // Add a constant argument for the calling convention
762 pushStackMapConstant(Ops, *this, CS.getCallingConv());
764 // Add a constant argument for the flags
765 uint64_t Flags = ISP.getFlags();
767 ((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0)
768 && "unknown flag used");
769 pushStackMapConstant(Ops, *this, Flags);
771 // Insert all vmstate and gcstate arguments
772 Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end());
774 // Add register mask from call node
775 Ops.push_back(*RegMaskIt);
778 Ops.push_back(Chain);
780 // Same for the glue, but we add it only if original call had it
784 // Compute return values. Provide a glue output since we consume one as
785 // input. This allows someone else to chain off us as needed.
786 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
788 SDNode *StatepointMCNode =
789 DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
791 SDNode *SinkNode = StatepointMCNode;
793 // Build the GC_TRANSITION_END node if necessary.
795 // See the comment above regarding GC_TRANSITION_START for the layout of
796 // the operands to the GC_TRANSITION_END node.
797 if (IsGCTransition) {
798 SmallVector<SDValue, 8> TEOps;
801 TEOps.push_back(SDValue(StatepointMCNode, 0));
803 // Add GC transition arguments
804 for (const Value *V : ISP.gc_transition_args()) {
805 TEOps.push_back(getValue(V));
806 if (V->getType()->isPointerTy())
807 TEOps.push_back(DAG.getSrcValue(V));
811 TEOps.push_back(SDValue(StatepointMCNode, 1));
813 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
815 SDValue GCTransitionStart =
816 DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps);
818 SinkNode = GCTransitionStart.getNode();
821 // Replace original call
822 DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root
823 // Remove original call node
824 DAG.DeleteNode(CallNode);
826 // DON'T set the root - under the assumption that it's already set past the
827 // inserted node we created.
829 // TODO: A better future implementation would be to emit a single variable
830 // argument, variable return value STATEPOINT node here and then hookup the
831 // return value of each gc.relocate to the respective output of the
832 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
833 // to actually be possible today.
836 void SelectionDAGBuilder::visitGCResult(const CallInst &CI) {
837 // The result value of the gc_result is simply the result of the actual
838 // call. We've already emitted this, so just grab the value.
839 Instruction *I = cast<Instruction>(CI.getArgOperand(0));
840 assert(isStatepoint(I) && "first argument must be a statepoint token");
842 if (I->getParent() != CI.getParent()) {
843 // Statepoint is in different basic block so we should have stored call
844 // result in a virtual register.
845 // We can not use default getValue() functionality to copy value from this
846 // register because statepoint and actuall call return types can be
847 // different, and getValue() will use CopyFromReg of the wrong type,
848 // which is always i32 in our case.
849 PointerType *CalleeType = cast<PointerType>(
850 ImmutableStatepoint(I).getCalledValue()->getType());
852 cast<FunctionType>(CalleeType->getElementType())->getReturnType();
853 SDValue CopyFromReg = getCopyFromRegs(I, RetTy);
855 assert(CopyFromReg.getNode());
856 setValue(&CI, CopyFromReg);
858 setValue(&CI, getValue(I));
862 void SelectionDAGBuilder::visitGCRelocate(const CallInst &CI) {
863 GCRelocateOperands RelocateOpers(&CI);
867 // We skip this check for relocates not in the same basic block as thier
868 // statepoint. It would be too expensive to preserve validation info through
869 // different basic blocks.
870 if (RelocateOpers.getStatepoint()->getParent() == CI.getParent()) {
871 StatepointLowering.relocCallVisited(CI);
875 const Value *DerivedPtr = RelocateOpers.getDerivedPtr();
876 SDValue SD = getValue(DerivedPtr);
878 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
879 FuncInfo.StatepointRelocatedValues[RelocateOpers.getStatepoint()];
881 // We should have recorded location for this pointer
882 assert(SpillMap.count(DerivedPtr) && "Relocating not lowered gc value");
883 Optional<int> DerivedPtrLocation = SpillMap[DerivedPtr];
885 // We didn't need to spill these special cases (constants and allocas).
886 // See the handling in spillIncomingValueForStatepoint for detail.
887 if (!DerivedPtrLocation) {
892 SDValue SpillSlot = DAG.getTargetFrameIndex(*DerivedPtrLocation,
895 // Be conservative: flush all pending loads
896 // TODO: Probably we can be less restrictive on this,
897 // it may allow more scheduling opportunities.
898 SDValue Chain = getRoot();
901 DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, SpillSlot,
902 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
903 *DerivedPtrLocation),
904 false, false, false, 0);
906 // Again, be conservative, don't emit pending loads
907 DAG.setRoot(SpillLoad.getValue(1));
909 assert(SpillLoad.getNode());
910 setValue(&CI, SpillLoad);