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 (const auto *Relocate = dyn_cast<GCRelocateInst>(Val)) {
132 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
133 Builder.FuncInfo.StatepointRelocatedValues[Relocate->getStatepoint()];
135 auto It = SpillMap.find(Relocate->getDerivedPtr());
136 if (It == SpillMap.end())
137 return Optional<int>();
142 // Look through bitcast instructions.
143 if (const BitCastInst *Cast = dyn_cast<BitCastInst>(Val)) {
144 return findPreviousSpillSlot(Cast->getOperand(0), Builder, LookUpDepth - 1);
147 // Look through phi nodes
148 // All incoming values should have same known stack slot, otherwise result
150 if (const PHINode *Phi = dyn_cast<PHINode>(Val)) {
151 Optional<int> MergedResult = None;
153 for (auto &IncomingValue : Phi->incoming_values()) {
154 Optional<int> SpillSlot =
155 findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth - 1);
156 if (!SpillSlot.hasValue())
157 return Optional<int>();
159 if (MergedResult.hasValue() && *MergedResult != *SpillSlot)
160 return Optional<int>();
162 MergedResult = SpillSlot;
167 // TODO: We can do better for PHI nodes. In cases like this:
168 // ptr = phi(relocated_pointer, not_relocated_pointer)
170 // We will return that stack slot for ptr is unknown. And later we might
171 // assign different stack slots for ptr and relocated_pointer. This limits
172 // llvm's ability to remove redundant stores.
173 // Unfortunately it's hard to accomplish in current infrastructure.
174 // We use this function to eliminate spill store completely, while
175 // in example we still need to emit store, but instead of any location
176 // we need to use special "preferred" location.
178 // TODO: handle simple updates. If a value is modified and the original
179 // value is no longer live, it would be nice to put the modified value in the
180 // same slot. This allows folding of the memory accesses for some
181 // instructions types (like an increment).
185 // However we need to be careful for cases like this:
189 // Here we want to reserve spill slot for 'i', but not for 'i+1'. If we just
190 // put handling of simple modifications in this function like it's done
191 // for bitcasts we might end up reserving i's slot for 'i+1' because order in
192 // which we visit values is unspecified.
194 // Don't know any information about this instruction
195 return Optional<int>();
198 /// Try to find existing copies of the incoming values in stack slots used for
199 /// statepoint spilling. If we can find a spill slot for the incoming value,
200 /// mark that slot as allocated, and reuse the same slot for this safepoint.
201 /// This helps to avoid series of loads and stores that only serve to reshuffle
202 /// values on the stack between calls.
203 static void reservePreviousStackSlotForValue(const Value *IncomingValue,
204 SelectionDAGBuilder &Builder) {
206 SDValue Incoming = Builder.getValue(IncomingValue);
208 if (isa<ConstantSDNode>(Incoming) || isa<FrameIndexSDNode>(Incoming)) {
209 // We won't need to spill this, so no need to check for previously
210 // allocated stack slots
214 SDValue OldLocation = Builder.StatepointLowering.getLocation(Incoming);
215 if (OldLocation.getNode())
216 // duplicates in input
219 const int LookUpDepth = 6;
220 Optional<int> Index =
221 findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth);
222 if (!Index.hasValue())
225 auto Itr = std::find(Builder.FuncInfo.StatepointStackSlots.begin(),
226 Builder.FuncInfo.StatepointStackSlots.end(), *Index);
227 assert(Itr != Builder.FuncInfo.StatepointStackSlots.end() &&
228 "value spilled to the unknown stack slot");
230 // This is one of our dedicated lowering slots
232 std::distance(Builder.FuncInfo.StatepointStackSlots.begin(), Itr);
233 if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) {
234 // stack slot already assigned to someone else, can't use it!
235 // TODO: currently we reserve space for gc arguments after doing
236 // normal allocation for deopt arguments. We should reserve for
237 // _all_ deopt and gc arguments, then start allocating. This
238 // will prevent some moves being inserted when vm state changes,
239 // but gc state doesn't between two calls.
242 // Reserve this stack slot
243 Builder.StatepointLowering.reserveStackSlot(Offset);
245 // Cache this slot so we find it when going through the normal
247 SDValue Loc = Builder.DAG.getTargetFrameIndex(*Index, Incoming.getValueType());
248 Builder.StatepointLowering.setLocation(Incoming, Loc);
251 /// Remove any duplicate (as SDValues) from the derived pointer pairs. This
252 /// is not required for correctness. It's purpose is to reduce the size of
253 /// StackMap section. It has no effect on the number of spill slots required
254 /// or the actual lowering.
255 static void removeDuplicatesGCPtrs(SmallVectorImpl<const Value *> &Bases,
256 SmallVectorImpl<const Value *> &Ptrs,
257 SmallVectorImpl<const Value *> &Relocs,
258 SelectionDAGBuilder &Builder) {
260 // This is horribly inefficient, but I don't care right now
261 SmallSet<SDValue, 64> Seen;
263 SmallVector<const Value *, 64> NewBases, NewPtrs, NewRelocs;
264 for (size_t i = 0; i < Ptrs.size(); i++) {
265 SDValue SD = Builder.getValue(Ptrs[i]);
266 // Only add non-duplicates
267 if (Seen.count(SD) == 0) {
268 NewBases.push_back(Bases[i]);
269 NewPtrs.push_back(Ptrs[i]);
270 NewRelocs.push_back(Relocs[i]);
274 assert(Bases.size() >= NewBases.size());
275 assert(Ptrs.size() >= NewPtrs.size());
276 assert(Relocs.size() >= NewRelocs.size());
280 assert(Ptrs.size() == Bases.size());
281 assert(Ptrs.size() == Relocs.size());
284 /// Extract call from statepoint, lower it and return pointer to the
285 /// call node. Also update NodeMap so that getValue(statepoint) will
286 /// reference lowered call result
288 lowerCallFromStatepoint(ImmutableStatepoint ISP, const BasicBlock *EHPadBB,
289 SelectionDAGBuilder &Builder,
290 SmallVectorImpl<SDValue> &PendingExports) {
292 ImmutableCallSite CS(ISP.getCallSite());
294 SDValue ActualCallee;
296 if (ISP.getNumPatchBytes() > 0) {
297 // If we've been asked to emit a nop sequence instead of a call instruction
298 // for this statepoint then don't lower the call target, but use a constant
299 // `null` instead. Not lowering the call target lets statepoint clients get
300 // away without providing a physical address for the symbolic call target at
303 const auto &TLI = Builder.DAG.getTargetLoweringInfo();
304 const auto &DL = Builder.DAG.getDataLayout();
306 unsigned AS = ISP.getCalledValue()->getType()->getPointerAddressSpace();
307 ActualCallee = Builder.DAG.getConstant(0, Builder.getCurSDLoc(),
308 TLI.getPointerTy(DL, AS));
310 ActualCallee = Builder.getValue(ISP.getCalledValue());
312 assert(CS.getCallingConv() != CallingConv::AnyReg &&
313 "anyregcc is not supported on statepoints!");
315 Type *DefTy = ISP.getActualReturnType();
316 bool HasDef = !DefTy->isVoidTy();
318 SDValue ReturnValue, CallEndVal;
319 std::tie(ReturnValue, CallEndVal) = Builder.lowerCallOperands(
320 ISP.getCallSite(), ImmutableStatepoint::CallArgsBeginPos,
321 ISP.getNumCallArgs(), ActualCallee, DefTy, EHPadBB,
322 false /* IsPatchPoint */);
324 SDNode *CallEnd = CallEndVal.getNode();
326 // Get a call instruction from the call sequence chain. Tail calls are not
327 // allowed. The following code is essentially reverse engineering X86's
330 // We are expecting DAG to have the following form:
332 // ch = eh_label (only in case of invoke statepoint)
333 // ch, glue = callseq_start ch
334 // ch, glue = X86::Call ch, glue
335 // ch, glue = callseq_end ch, glue
336 // get_return_value ch, glue
338 // get_return_value can either be a sequence of CopyFromReg instructions
339 // to grab the return value from the return register(s), or it can be a LOAD
340 // to load a value returned by reference via a stack slot.
343 if (CallEnd->getOpcode() == ISD::LOAD)
344 CallEnd = CallEnd->getOperand(0).getNode();
346 while (CallEnd->getOpcode() == ISD::CopyFromReg)
347 CallEnd = CallEnd->getOperand(0).getNode();
350 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
352 // Export the result value if needed
353 const Instruction *GCResult = ISP.getGCResult();
354 if (HasDef && GCResult) {
355 if (GCResult->getParent() != CS.getParent()) {
356 // Result value will be used in a different basic block so we need to
358 // Default exporting mechanism will not work here because statepoint call
359 // has a different type than the actual call. It means that by default
360 // llvm will create export register of the wrong type (always i32 in our
361 // case). So instead we need to create export register with correct type
363 // TODO: To eliminate this problem we can remove gc.result intrinsics
364 // completely and make statepoint call to return a tuple.
365 unsigned Reg = Builder.FuncInfo.CreateRegs(ISP.getActualReturnType());
367 *Builder.DAG.getContext(), Builder.DAG.getTargetLoweringInfo(),
368 Builder.DAG.getDataLayout(), Reg, ISP.getActualReturnType());
369 SDValue Chain = Builder.DAG.getEntryNode();
371 RFV.getCopyToRegs(ReturnValue, Builder.DAG, Builder.getCurSDLoc(), Chain,
373 PendingExports.push_back(Chain);
374 Builder.FuncInfo.ValueMap[CS.getInstruction()] = Reg;
376 // Result value will be used in a same basic block. Don't export it or
377 // perform any explicit register copies.
378 // We'll replace the actuall call node shortly. gc_result will grab
380 Builder.setValue(CS.getInstruction(), ReturnValue);
383 // The token value is never used from here on, just generate a poison value
384 Builder.setValue(CS.getInstruction(),
385 Builder.DAG.getIntPtrConstant(-1, Builder.getCurSDLoc()));
388 return CallEnd->getOperand(0).getNode();
391 /// Callect all gc pointers coming into statepoint intrinsic, clean them up,
392 /// and return two arrays:
393 /// Bases - base pointers incoming to this statepoint
394 /// Ptrs - derived pointers incoming to this statepoint
395 /// Relocs - the gc_relocate corresponding to each base/ptr pair
396 /// Elements of this arrays should be in one-to-one correspondence with each
397 /// other i.e Bases[i], Ptrs[i] are from the same gcrelocate call
398 static void getIncomingStatepointGCValues(
399 SmallVectorImpl<const Value *> &Bases, SmallVectorImpl<const Value *> &Ptrs,
400 SmallVectorImpl<const Value *> &Relocs, ImmutableStatepoint StatepointSite,
401 SelectionDAGBuilder &Builder) {
402 for (const GCRelocateInst *Relocate : StatepointSite.getRelocates()) {
403 Relocs.push_back(Relocate);
404 Bases.push_back(Relocate->getBasePtr());
405 Ptrs.push_back(Relocate->getDerivedPtr());
408 // Remove any redundant llvm::Values which map to the same SDValue as another
409 // input. Also has the effect of removing duplicates in the original
410 // llvm::Value input list as well. This is a useful optimization for
411 // reducing the size of the StackMap section. It has no other impact.
412 removeDuplicatesGCPtrs(Bases, Ptrs, Relocs, Builder);
414 assert(Bases.size() == Ptrs.size() && Ptrs.size() == Relocs.size());
417 /// Spill a value incoming to the statepoint. It might be either part of
419 /// or gcstate. In both cases unconditionally spill it on the stack unless it
420 /// is a null constant. Return pair with first element being frame index
421 /// containing saved value and second element with outgoing chain from the
423 static std::pair<SDValue, SDValue>
424 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
425 SelectionDAGBuilder &Builder) {
426 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
428 // Emit new store if we didn't do it for this ptr before
429 if (!Loc.getNode()) {
430 Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
432 assert(isa<FrameIndexSDNode>(Loc));
433 int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
434 // We use TargetFrameIndex so that isel will not select it into LEA
435 Loc = Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
437 // TODO: We can create TokenFactor node instead of
438 // chaining stores one after another, this may allow
439 // a bit more optimal scheduling for them
440 Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
441 MachinePointerInfo::getFixedStack(
442 Builder.DAG.getMachineFunction(), Index),
445 Builder.StatepointLowering.setLocation(Incoming, Loc);
448 assert(Loc.getNode());
449 return std::make_pair(Loc, Chain);
452 /// Lower a single value incoming to a statepoint node. This value can be
453 /// either a deopt value or a gc value, the handling is the same. We special
454 /// case constants and allocas, then fall back to spilling if required.
455 static void lowerIncomingStatepointValue(SDValue Incoming,
456 SmallVectorImpl<SDValue> &Ops,
457 SelectionDAGBuilder &Builder) {
458 SDValue Chain = Builder.getRoot();
460 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
461 // If the original value was a constant, make sure it gets recorded as
462 // such in the stackmap. This is required so that the consumer can
463 // parse any internal format to the deopt state. It also handles null
464 // pointers and other constant pointers in GC states
465 pushStackMapConstant(Ops, Builder, C->getSExtValue());
466 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
467 // This handles allocas as arguments to the statepoint (this is only
468 // really meaningful for a deopt value. For GC, we'd be trying to
469 // relocate the address of the alloca itself?)
470 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
471 Incoming.getValueType()));
473 // Otherwise, locate a spill slot and explicitly spill it so it
474 // can be found by the runtime later. We currently do not support
475 // tracking values through callee saved registers to their eventual
476 // spill location. This would be a useful optimization, but would
477 // need to be optional since it requires a lot of complexity on the
478 // runtime side which not all would support.
479 std::pair<SDValue, SDValue> Res =
480 spillIncomingStatepointValue(Incoming, Chain, Builder);
481 Ops.push_back(Res.first);
485 Builder.DAG.setRoot(Chain);
488 /// Lower deopt state and gc pointer arguments of the statepoint. The actual
489 /// lowering is described in lowerIncomingStatepointValue. This function is
490 /// responsible for lowering everything in the right position and playing some
491 /// tricks to avoid redundant stack manipulation where possible. On
492 /// completion, 'Ops' will contain ready to use operands for machine code
493 /// statepoint. The chain nodes will have already been created and the DAG root
494 /// will be set to the last value spilled (if any were).
495 static void lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
496 ImmutableStatepoint StatepointSite,
497 SelectionDAGBuilder &Builder) {
499 // Lower the deopt and gc arguments for this statepoint. Layout will
500 // be: deopt argument length, deopt arguments.., gc arguments...
502 SmallVector<const Value *, 64> Bases, Ptrs, Relocations;
503 getIncomingStatepointGCValues(Bases, Ptrs, Relocations, StatepointSite,
507 // Check that each of the gc pointer and bases we've gotten out of the
508 // safepoint is something the strategy thinks might be a pointer (or vector
509 // of pointers) into the GC heap. This is basically just here to help catch
510 // errors during statepoint insertion. TODO: This should actually be in the
511 // Verifier, but we can't get to the GCStrategy from there (yet).
512 GCStrategy &S = Builder.GFI->getStrategy();
513 for (const Value *V : Bases) {
514 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
515 if (Opt.hasValue()) {
516 assert(Opt.getValue() &&
517 "non gc managed base pointer found in statepoint");
520 for (const Value *V : Ptrs) {
521 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
522 if (Opt.hasValue()) {
523 assert(Opt.getValue() &&
524 "non gc managed derived pointer found in statepoint");
527 for (const Value *V : Relocations) {
528 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
529 if (Opt.hasValue()) {
530 assert(Opt.getValue() && "non gc managed pointer relocated");
535 // Before we actually start lowering (and allocating spill slots for values),
536 // reserve any stack slots which we judge to be profitable to reuse for a
537 // particular value. This is purely an optimization over the code below and
538 // doesn't change semantics at all. It is important for performance that we
539 // reserve slots for both deopt and gc values before lowering either.
540 for (const Value *V : StatepointSite.vm_state_args()) {
541 reservePreviousStackSlotForValue(V, Builder);
543 for (unsigned i = 0; i < Bases.size(); ++i) {
544 reservePreviousStackSlotForValue(Bases[i], Builder);
545 reservePreviousStackSlotForValue(Ptrs[i], Builder);
548 // First, prefix the list with the number of unique values to be
549 // lowered. Note that this is the number of *Values* not the
550 // number of SDValues required to lower them.
551 const int NumVMSArgs = StatepointSite.getNumTotalVMSArgs();
552 pushStackMapConstant(Ops, Builder, NumVMSArgs);
554 assert(NumVMSArgs == std::distance(StatepointSite.vm_state_begin(),
555 StatepointSite.vm_state_end()));
557 // The vm state arguments are lowered in an opaque manner. We do
558 // not know what type of values are contained within. We skip the
559 // first one since that happens to be the total number we lowered
560 // explicitly just above. We could have left it in the loop and
561 // not done it explicitly, but it's far easier to understand this
563 for (const Value *V : StatepointSite.vm_state_args()) {
564 SDValue Incoming = Builder.getValue(V);
565 lowerIncomingStatepointValue(Incoming, Ops, Builder);
568 // Finally, go ahead and lower all the gc arguments. There's no prefixed
569 // length for this one. After lowering, we'll have the base and pointer
570 // arrays interwoven with each (lowered) base pointer immediately followed by
571 // it's (lowered) derived pointer. i.e
572 // (base[0], ptr[0], base[1], ptr[1], ...)
573 for (unsigned i = 0; i < Bases.size(); ++i) {
574 const Value *Base = Bases[i];
575 lowerIncomingStatepointValue(Builder.getValue(Base), Ops, Builder);
577 const Value *Ptr = Ptrs[i];
578 lowerIncomingStatepointValue(Builder.getValue(Ptr), Ops, Builder);
581 // If there are any explicit spill slots passed to the statepoint, record
582 // them, but otherwise do not do anything special. These are user provided
583 // allocas and give control over placement to the consumer. In this case,
584 // it is the contents of the slot which may get updated, not the pointer to
586 for (Value *V : StatepointSite.gc_args()) {
587 SDValue Incoming = Builder.getValue(V);
588 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
589 // This handles allocas as arguments to the statepoint
590 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
591 Incoming.getValueType()));
595 // Record computed locations for all lowered values.
596 // This can not be embedded in lowering loops as we need to record *all*
597 // values, while previous loops account only values with unique SDValues.
598 const Instruction *StatepointInstr =
599 StatepointSite.getCallSite().getInstruction();
600 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
601 Builder.FuncInfo.StatepointRelocatedValues[StatepointInstr];
603 for (const GCRelocateInst *Relocate : StatepointSite.getRelocates()) {
604 const Value *V = Relocate->getDerivedPtr();
605 SDValue SDV = Builder.getValue(V);
606 SDValue Loc = Builder.StatepointLowering.getLocation(SDV);
609 SpillMap[V] = cast<FrameIndexSDNode>(Loc)->getIndex();
611 // Record value as visited, but not spilled. This is case for allocas
612 // and constants. For this values we can avoid emitting spill load while
613 // visiting corresponding gc_relocate.
614 // Actually we do not need to record them in this map at all.
615 // We do this only to check that we are not relocating any unvisited
619 // Default llvm mechanisms for exporting values which are used in
620 // different basic blocks does not work for gc relocates.
621 // Note that it would be incorrect to teach llvm that all relocates are
622 // uses of the corresponding values so that it would automatically
623 // export them. Relocates of the spilled values does not use original
625 if (Relocate->getParent() != StatepointInstr->getParent())
626 Builder.ExportFromCurrentBlock(V);
631 void SelectionDAGBuilder::visitStatepoint(const CallInst &CI) {
632 // Check some preconditions for sanity
633 assert(isStatepoint(&CI) &&
634 "function called must be the statepoint function");
636 LowerStatepoint(ImmutableStatepoint(&CI));
639 void SelectionDAGBuilder::LowerStatepoint(
640 ImmutableStatepoint ISP, const BasicBlock *EHPadBB /*= nullptr*/) {
641 // The basic scheme here is that information about both the original call and
642 // the safepoint is encoded in the CallInst. We create a temporary call and
643 // lower it, then reverse engineer the calling sequence.
647 StatepointLowering.startNewStatepoint(*this);
649 ImmutableCallSite CS(ISP.getCallSite());
652 // Consistency check. Check only relocates in the same basic block as thier
654 for (const User *U : CS->users()) {
655 const CallInst *Call = cast<CallInst>(U);
656 if (isa<GCRelocateInst>(Call) && Call->getParent() == CS.getParent())
657 StatepointLowering.scheduleRelocCall(*Call);
662 // If this is a malformed statepoint, report it early to simplify debugging.
663 // This should catch any IR level mistake that's made when constructing or
664 // transforming statepoints.
667 // Check that the associated GCStrategy expects to encounter statepoints.
668 assert(GFI->getStrategy().useStatepoints() &&
669 "GCStrategy does not expect to encounter statepoints");
672 // Lower statepoint vmstate and gcstate arguments
673 SmallVector<SDValue, 10> LoweredMetaArgs;
674 lowerStatepointMetaArgs(LoweredMetaArgs, ISP, *this);
676 // Get call node, we will replace it later with statepoint
678 lowerCallFromStatepoint(ISP, EHPadBB, *this, PendingExports);
680 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
681 // nodes with all the appropriate arguments and return values.
683 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
684 SDValue Chain = CallNode->getOperand(0);
687 bool CallHasIncomingGlue = CallNode->getGluedNode();
688 if (CallHasIncomingGlue) {
689 // Glue is always last operand
690 Glue = CallNode->getOperand(CallNode->getNumOperands() - 1);
693 // Build the GC_TRANSITION_START node if necessary.
695 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
696 // order in which they appear in the call to the statepoint intrinsic. If
697 // any of the operands is a pointer-typed, that operand is immediately
698 // followed by a SRCVALUE for the pointer that may be used during lowering
699 // (e.g. to form MachinePointerInfo values for loads/stores).
700 const bool IsGCTransition =
701 (ISP.getFlags() & (uint64_t)StatepointFlags::GCTransition) ==
702 (uint64_t)StatepointFlags::GCTransition;
703 if (IsGCTransition) {
704 SmallVector<SDValue, 8> TSOps;
707 TSOps.push_back(Chain);
709 // Add GC transition arguments
710 for (const Value *V : ISP.gc_transition_args()) {
711 TSOps.push_back(getValue(V));
712 if (V->getType()->isPointerTy())
713 TSOps.push_back(DAG.getSrcValue(V));
716 // Add glue if necessary
717 if (CallHasIncomingGlue)
718 TSOps.push_back(Glue);
720 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
722 SDValue GCTransitionStart =
723 DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps);
725 Chain = GCTransitionStart.getValue(0);
726 Glue = GCTransitionStart.getValue(1);
729 // TODO: Currently, all of these operands are being marked as read/write in
730 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
731 // and flags to be read-only.
732 SmallVector<SDValue, 40> Ops;
734 // Add the <id> and <numBytes> constants.
735 Ops.push_back(DAG.getTargetConstant(ISP.getID(), getCurSDLoc(), MVT::i64));
737 DAG.getTargetConstant(ISP.getNumPatchBytes(), getCurSDLoc(), MVT::i32));
739 // Calculate and push starting position of vmstate arguments
740 // Get number of arguments incoming directly into call node
741 unsigned NumCallRegArgs =
742 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
743 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32));
746 SDValue CallTarget = SDValue(CallNode->getOperand(1).getNode(), 0);
747 Ops.push_back(CallTarget);
749 // Add call arguments
750 // Get position of register mask in the call
751 SDNode::op_iterator RegMaskIt;
752 if (CallHasIncomingGlue)
753 RegMaskIt = CallNode->op_end() - 2;
755 RegMaskIt = CallNode->op_end() - 1;
756 Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt);
758 // Add a constant argument for the calling convention
759 pushStackMapConstant(Ops, *this, CS.getCallingConv());
761 // Add a constant argument for the flags
762 uint64_t Flags = ISP.getFlags();
764 ((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0)
765 && "unknown flag used");
766 pushStackMapConstant(Ops, *this, Flags);
768 // Insert all vmstate and gcstate arguments
769 Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end());
771 // Add register mask from call node
772 Ops.push_back(*RegMaskIt);
775 Ops.push_back(Chain);
777 // Same for the glue, but we add it only if original call had it
781 // Compute return values. Provide a glue output since we consume one as
782 // input. This allows someone else to chain off us as needed.
783 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
785 SDNode *StatepointMCNode =
786 DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
788 SDNode *SinkNode = StatepointMCNode;
790 // Build the GC_TRANSITION_END node if necessary.
792 // See the comment above regarding GC_TRANSITION_START for the layout of
793 // the operands to the GC_TRANSITION_END node.
794 if (IsGCTransition) {
795 SmallVector<SDValue, 8> TEOps;
798 TEOps.push_back(SDValue(StatepointMCNode, 0));
800 // Add GC transition arguments
801 for (const Value *V : ISP.gc_transition_args()) {
802 TEOps.push_back(getValue(V));
803 if (V->getType()->isPointerTy())
804 TEOps.push_back(DAG.getSrcValue(V));
808 TEOps.push_back(SDValue(StatepointMCNode, 1));
810 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
812 SDValue GCTransitionStart =
813 DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps);
815 SinkNode = GCTransitionStart.getNode();
818 // Replace original call
819 DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root
820 // Remove original call node
821 DAG.DeleteNode(CallNode);
823 // DON'T set the root - under the assumption that it's already set past the
824 // inserted node we created.
826 // TODO: A better future implementation would be to emit a single variable
827 // argument, variable return value STATEPOINT node here and then hookup the
828 // return value of each gc.relocate to the respective output of the
829 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
830 // to actually be possible today.
833 void SelectionDAGBuilder::visitGCResult(const CallInst &CI) {
834 // The result value of the gc_result is simply the result of the actual
835 // call. We've already emitted this, so just grab the value.
836 Instruction *I = cast<Instruction>(CI.getArgOperand(0));
837 assert(isStatepoint(I) && "first argument must be a statepoint token");
839 if (I->getParent() != CI.getParent()) {
840 // Statepoint is in different basic block so we should have stored call
841 // result in a virtual register.
842 // We can not use default getValue() functionality to copy value from this
843 // register because statepoint and actuall call return types can be
844 // different, and getValue() will use CopyFromReg of the wrong type,
845 // which is always i32 in our case.
846 PointerType *CalleeType = cast<PointerType>(
847 ImmutableStatepoint(I).getCalledValue()->getType());
849 cast<FunctionType>(CalleeType->getElementType())->getReturnType();
850 SDValue CopyFromReg = getCopyFromRegs(I, RetTy);
852 assert(CopyFromReg.getNode());
853 setValue(&CI, CopyFromReg);
855 setValue(&CI, getValue(I));
859 void SelectionDAGBuilder::visitGCRelocate(const GCRelocateInst &Relocate) {
862 // We skip this check for relocates not in the same basic block as thier
863 // statepoint. It would be too expensive to preserve validation info through
864 // different basic blocks.
865 if (Relocate.getStatepoint()->getParent() == Relocate.getParent()) {
866 StatepointLowering.relocCallVisited(Relocate);
870 const Value *DerivedPtr = Relocate.getDerivedPtr();
871 SDValue SD = getValue(DerivedPtr);
873 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
874 FuncInfo.StatepointRelocatedValues[Relocate.getStatepoint()];
876 // We should have recorded location for this pointer
877 assert(SpillMap.count(DerivedPtr) && "Relocating not lowered gc value");
878 Optional<int> DerivedPtrLocation = SpillMap[DerivedPtr];
880 // We didn't need to spill these special cases (constants and allocas).
881 // See the handling in spillIncomingValueForStatepoint for detail.
882 if (!DerivedPtrLocation) {
883 setValue(&Relocate, SD);
887 SDValue SpillSlot = DAG.getTargetFrameIndex(*DerivedPtrLocation,
890 // Be conservative: flush all pending loads
891 // TODO: Probably we can be less restrictive on this,
892 // it may allow more scheduling opportunities.
893 SDValue Chain = getRoot();
896 DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, SpillSlot,
897 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
898 *DerivedPtrLocation),
899 false, false, false, 0);
901 // Again, be conservative, don't emit pending loads
902 DAG.setRoot(SpillLoad.getValue(1));
904 assert(SpillLoad.getNode());
905 setValue(&Relocate, SpillLoad);