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 sync and
57 // the clear patterns of a SelectionDAGBuilder have no relation to
58 // FunctionLoweringInfo. SmallBitVector::reset initializes all bits to false.
59 AllocatedStackSlots.resize(Builder.FuncInfo.StatepointStackSlots.size());
62 void StatepointLoweringState::clear() {
64 AllocatedStackSlots.clear();
65 assert(PendingGCRelocateCalls.empty() &&
66 "cleared before statepoint sequence completed");
70 StatepointLoweringState::allocateStackSlot(EVT ValueType,
71 SelectionDAGBuilder &Builder) {
72 NumSlotsAllocatedForStatepoints++;
73 auto *MFI = Builder.DAG.getMachineFunction().getFrameInfo();
75 unsigned SpillSize = ValueType.getSizeInBits() / 8;
76 assert((SpillSize * 8) == ValueType.getSizeInBits() && "Size not in bytes?");
78 // First look for a previously created stack slot which is not in
79 // use (accounting for the fact arbitrary slots may already be
80 // reserved), or to create a new stack slot and use it.
82 const size_t NumSlots = AllocatedStackSlots.size();
83 assert(NextSlotToAllocate <= NumSlots && "Broken invariant");
85 // The stack slots in StatepointStackSlots beyond the first NumSlots were
86 // added in this instance of StatepointLoweringState, and cannot be re-used.
87 assert(NumSlots <= Builder.FuncInfo.StatepointStackSlots.size() &&
90 for (; NextSlotToAllocate < NumSlots; NextSlotToAllocate++) {
91 if (!AllocatedStackSlots.test(NextSlotToAllocate)) {
92 const int FI = Builder.FuncInfo.StatepointStackSlots[NextSlotToAllocate];
93 if (MFI->getObjectSize(FI) == SpillSize) {
94 AllocatedStackSlots.set(NextSlotToAllocate);
95 return Builder.DAG.getFrameIndex(FI, ValueType);
100 // Couldn't find a free slot, so create a new one:
102 SDValue SpillSlot = Builder.DAG.CreateStackTemporary(ValueType);
103 const unsigned FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
104 MFI->markAsStatepointSpillSlotObjectIndex(FI);
106 Builder.FuncInfo.StatepointStackSlots.push_back(FI);
108 StatepointMaxSlotsRequired = std::max<unsigned long>(
109 StatepointMaxSlotsRequired, Builder.FuncInfo.StatepointStackSlots.size());
114 /// Utility function for reservePreviousStackSlotForValue. Tries to find
115 /// stack slot index to which we have spilled value for previous statepoints.
116 /// LookUpDepth specifies maximum DFS depth this function is allowed to look.
117 static Optional<int> findPreviousSpillSlot(const Value *Val,
118 SelectionDAGBuilder &Builder,
120 // Can not look any further - give up now
121 if (LookUpDepth <= 0)
124 // Spill location is known for gc relocates
125 if (const auto *Relocate = dyn_cast<GCRelocateInst>(Val)) {
126 const auto &SpillMap =
127 Builder.FuncInfo.StatepointSpillMaps[Relocate->getStatepoint()];
129 auto It = SpillMap.find(Relocate->getDerivedPtr());
130 if (It == SpillMap.end())
136 // Look through bitcast instructions.
137 if (const BitCastInst *Cast = dyn_cast<BitCastInst>(Val))
138 return findPreviousSpillSlot(Cast->getOperand(0), Builder, LookUpDepth - 1);
140 // Look through phi nodes
141 // All incoming values should have same known stack slot, otherwise result
143 if (const PHINode *Phi = dyn_cast<PHINode>(Val)) {
144 Optional<int> MergedResult = None;
146 for (auto &IncomingValue : Phi->incoming_values()) {
147 Optional<int> SpillSlot =
148 findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth - 1);
149 if (!SpillSlot.hasValue())
152 if (MergedResult.hasValue() && *MergedResult != *SpillSlot)
155 MergedResult = SpillSlot;
160 // TODO: We can do better for PHI nodes. In cases like this:
161 // ptr = phi(relocated_pointer, not_relocated_pointer)
163 // We will return that stack slot for ptr is unknown. And later we might
164 // assign different stack slots for ptr and relocated_pointer. This limits
165 // llvm's ability to remove redundant stores.
166 // Unfortunately it's hard to accomplish in current infrastructure.
167 // We use this function to eliminate spill store completely, while
168 // in example we still need to emit store, but instead of any location
169 // we need to use special "preferred" location.
171 // TODO: handle simple updates. If a value is modified and the original
172 // value is no longer live, it would be nice to put the modified value in the
173 // same slot. This allows folding of the memory accesses for some
174 // instructions types (like an increment).
178 // However we need to be careful for cases like this:
182 // Here we want to reserve spill slot for 'i', but not for 'i+1'. If we just
183 // put handling of simple modifications in this function like it's done
184 // for bitcasts we might end up reserving i's slot for 'i+1' because order in
185 // which we visit values is unspecified.
187 // Don't know any information about this instruction
191 /// Try to find existing copies of the incoming values in stack slots used for
192 /// statepoint spilling. If we can find a spill slot for the incoming value,
193 /// mark that slot as allocated, and reuse the same slot for this safepoint.
194 /// This helps to avoid series of loads and stores that only serve to reshuffle
195 /// values on the stack between calls.
196 static void reservePreviousStackSlotForValue(const Value *IncomingValue,
197 SelectionDAGBuilder &Builder) {
199 SDValue Incoming = Builder.getValue(IncomingValue);
201 if (isa<ConstantSDNode>(Incoming) || isa<FrameIndexSDNode>(Incoming)) {
202 // We won't need to spill this, so no need to check for previously
203 // allocated stack slots
207 SDValue OldLocation = Builder.StatepointLowering.getLocation(Incoming);
208 if (OldLocation.getNode())
209 // Duplicates in input
212 const int LookUpDepth = 6;
213 Optional<int> Index =
214 findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth);
215 if (!Index.hasValue())
218 const auto &StatepointSlots = Builder.FuncInfo.StatepointStackSlots;
220 auto SlotIt = find(StatepointSlots, *Index);
221 assert(SlotIt != StatepointSlots.end() &&
222 "Value spilled to the unknown stack slot");
224 // This is one of our dedicated lowering slots
225 const int Offset = std::distance(StatepointSlots.begin(), SlotIt);
226 if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) {
227 // stack slot already assigned to someone else, can't use it!
228 // TODO: currently we reserve space for gc arguments after doing
229 // normal allocation for deopt arguments. We should reserve for
230 // _all_ deopt and gc arguments, then start allocating. This
231 // will prevent some moves being inserted when vm state changes,
232 // but gc state doesn't between two calls.
235 // Reserve this stack slot
236 Builder.StatepointLowering.reserveStackSlot(Offset);
238 // Cache this slot so we find it when going through the normal
240 SDValue Loc = Builder.DAG.getTargetFrameIndex(*Index, Incoming.getValueType());
241 Builder.StatepointLowering.setLocation(Incoming, Loc);
244 /// Remove any duplicate (as SDValues) from the derived pointer pairs. This
245 /// is not required for correctness. It's purpose is to reduce the size of
246 /// StackMap section. It has no effect on the number of spill slots required
247 /// or the actual lowering.
249 removeDuplicateGCPtrs(SmallVectorImpl<const Value *> &Bases,
250 SmallVectorImpl<const Value *> &Ptrs,
251 SmallVectorImpl<const GCRelocateInst *> &Relocs,
252 SelectionDAGBuilder &Builder,
253 FunctionLoweringInfo::StatepointSpillMap &SSM) {
254 DenseMap<SDValue, const Value *> Seen;
256 SmallVector<const Value *, 64> NewBases, NewPtrs;
257 SmallVector<const GCRelocateInst *, 64> NewRelocs;
258 for (size_t i = 0, e = Ptrs.size(); i < e; i++) {
259 SDValue SD = Builder.getValue(Ptrs[i]);
260 auto SeenIt = Seen.find(SD);
262 if (SeenIt == Seen.end()) {
263 // Only add non-duplicates
264 NewBases.push_back(Bases[i]);
265 NewPtrs.push_back(Ptrs[i]);
266 NewRelocs.push_back(Relocs[i]);
269 // Duplicate pointer found, note in SSM and move on:
270 SSM.DuplicateMap[Ptrs[i]] = SeenIt->second;
273 assert(Bases.size() >= NewBases.size());
274 assert(Ptrs.size() >= NewPtrs.size());
275 assert(Relocs.size() >= NewRelocs.size());
279 assert(Ptrs.size() == Bases.size());
280 assert(Ptrs.size() == Relocs.size());
283 /// Extract call from statepoint, lower it and return pointer to the
284 /// call node. Also update NodeMap so that getValue(statepoint) will
285 /// reference lowered call result
286 static std::pair<SDValue, SDNode *> lowerCallFromStatepointLoweringInfo(
287 SelectionDAGBuilder::StatepointLoweringInfo &SI,
288 SelectionDAGBuilder &Builder, SmallVectorImpl<SDValue> &PendingExports) {
290 SDValue ReturnValue, CallEndVal;
291 std::tie(ReturnValue, CallEndVal) =
292 Builder.lowerInvokable(SI.CLI, SI.EHPadBB);
293 SDNode *CallEnd = CallEndVal.getNode();
295 // Get a call instruction from the call sequence chain. Tail calls are not
296 // allowed. The following code is essentially reverse engineering X86's
299 // We are expecting DAG to have the following form:
301 // ch = eh_label (only in case of invoke statepoint)
302 // ch, glue = callseq_start ch
303 // ch, glue = X86::Call ch, glue
304 // ch, glue = callseq_end ch, glue
305 // get_return_value ch, glue
307 // get_return_value can either be a sequence of CopyFromReg instructions
308 // to grab the return value from the return register(s), or it can be a LOAD
309 // to load a value returned by reference via a stack slot.
311 bool HasDef = !SI.CLI.RetTy->isVoidTy();
313 if (CallEnd->getOpcode() == ISD::LOAD)
314 CallEnd = CallEnd->getOperand(0).getNode();
316 while (CallEnd->getOpcode() == ISD::CopyFromReg)
317 CallEnd = CallEnd->getOperand(0).getNode();
320 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
321 return std::make_pair(ReturnValue, CallEnd->getOperand(0).getNode());
324 /// Spill a value incoming to the statepoint. It might be either part of
326 /// or gcstate. In both cases unconditionally spill it on the stack unless it
327 /// is a null constant. Return pair with first element being frame index
328 /// containing saved value and second element with outgoing chain from the
330 static std::pair<SDValue, SDValue>
331 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
332 SelectionDAGBuilder &Builder) {
333 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
335 // Emit new store if we didn't do it for this ptr before
336 if (!Loc.getNode()) {
337 Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
339 int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
340 // We use TargetFrameIndex so that isel will not select it into LEA
341 Loc = Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
343 // TODO: We can create TokenFactor node instead of
344 // chaining stores one after another, this may allow
345 // a bit more optimal scheduling for them
348 // Right now we always allocate spill slots that are of the same
349 // size as the value we're about to spill (the size of spillee can
350 // vary since we spill vectors of pointers too). At some point we
351 // can consider allowing spills of smaller values to larger slots
352 // (i.e. change the '==' in the assert below to a '>=').
353 auto *MFI = Builder.DAG.getMachineFunction().getFrameInfo();
354 assert((MFI->getObjectSize(Index) * 8) ==
355 Incoming.getValueType().getSizeInBits() &&
356 "Bad spill: stack slot does not match!");
359 Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
360 MachinePointerInfo::getFixedStack(
361 Builder.DAG.getMachineFunction(), Index));
363 Builder.StatepointLowering.setLocation(Incoming, Loc);
366 assert(Loc.getNode());
367 return std::make_pair(Loc, Chain);
370 /// Lower a single value incoming to a statepoint node. This value can be
371 /// either a deopt value or a gc value, the handling is the same. We special
372 /// case constants and allocas, then fall back to spilling if required.
373 static void lowerIncomingStatepointValue(SDValue Incoming,
374 SmallVectorImpl<SDValue> &Ops,
375 SelectionDAGBuilder &Builder) {
376 SDValue Chain = Builder.getRoot();
378 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
379 // If the original value was a constant, make sure it gets recorded as
380 // such in the stackmap. This is required so that the consumer can
381 // parse any internal format to the deopt state. It also handles null
382 // pointers and other constant pointers in GC states. Note the constant
383 // vectors do not appear to actually hit this path and that anything larger
384 // than an i64 value (not type!) will fail asserts here.
385 pushStackMapConstant(Ops, Builder, C->getSExtValue());
386 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
387 // This handles allocas as arguments to the statepoint (this is only
388 // really meaningful for a deopt value. For GC, we'd be trying to
389 // relocate the address of the alloca itself?)
390 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
391 Incoming.getValueType()));
393 // Otherwise, locate a spill slot and explicitly spill it so it
394 // can be found by the runtime later. We currently do not support
395 // tracking values through callee saved registers to their eventual
396 // spill location. This would be a useful optimization, but would
397 // need to be optional since it requires a lot of complexity on the
398 // runtime side which not all would support.
399 auto Res = spillIncomingStatepointValue(Incoming, Chain, Builder);
400 Ops.push_back(Res.first);
404 Builder.DAG.setRoot(Chain);
407 /// Lower deopt state and gc pointer arguments of the statepoint. The actual
408 /// lowering is described in lowerIncomingStatepointValue. This function is
409 /// responsible for lowering everything in the right position and playing some
410 /// tricks to avoid redundant stack manipulation where possible. On
411 /// completion, 'Ops' will contain ready to use operands for machine code
412 /// statepoint. The chain nodes will have already been created and the DAG root
413 /// will be set to the last value spilled (if any were).
415 lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
416 SelectionDAGBuilder::StatepointLoweringInfo &SI,
417 SelectionDAGBuilder &Builder) {
418 // Lower the deopt and gc arguments for this statepoint. Layout will be:
419 // deopt argument length, deopt arguments.., gc arguments...
421 if (auto *GFI = Builder.GFI) {
422 // Check that each of the gc pointer and bases we've gotten out of the
423 // safepoint is something the strategy thinks might be a pointer (or vector
424 // of pointers) into the GC heap. This is basically just here to help catch
425 // errors during statepoint insertion. TODO: This should actually be in the
426 // Verifier, but we can't get to the GCStrategy from there (yet).
427 GCStrategy &S = GFI->getStrategy();
428 for (const Value *V : SI.Bases) {
429 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
430 if (Opt.hasValue()) {
431 assert(Opt.getValue() &&
432 "non gc managed base pointer found in statepoint");
435 for (const Value *V : SI.Ptrs) {
436 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
437 if (Opt.hasValue()) {
438 assert(Opt.getValue() &&
439 "non gc managed derived pointer found in statepoint");
443 assert(SI.Bases.empty() && "No gc specified, so cannot relocate pointers!");
444 assert(SI.Ptrs.empty() && "No gc specified, so cannot relocate pointers!");
448 // Before we actually start lowering (and allocating spill slots for values),
449 // reserve any stack slots which we judge to be profitable to reuse for a
450 // particular value. This is purely an optimization over the code below and
451 // doesn't change semantics at all. It is important for performance that we
452 // reserve slots for both deopt and gc values before lowering either.
453 for (const Value *V : SI.DeoptState) {
454 reservePreviousStackSlotForValue(V, Builder);
456 for (unsigned i = 0; i < SI.Bases.size(); ++i) {
457 reservePreviousStackSlotForValue(SI.Bases[i], Builder);
458 reservePreviousStackSlotForValue(SI.Ptrs[i], Builder);
461 // First, prefix the list with the number of unique values to be
462 // lowered. Note that this is the number of *Values* not the
463 // number of SDValues required to lower them.
464 const int NumVMSArgs = SI.DeoptState.size();
465 pushStackMapConstant(Ops, Builder, NumVMSArgs);
467 // The vm state arguments are lowered in an opaque manner. We do not know
468 // what type of values are contained within.
469 for (const Value *V : SI.DeoptState) {
470 SDValue Incoming = Builder.getValue(V);
471 lowerIncomingStatepointValue(Incoming, Ops, Builder);
474 // Finally, go ahead and lower all the gc arguments. There's no prefixed
475 // length for this one. After lowering, we'll have the base and pointer
476 // arrays interwoven with each (lowered) base pointer immediately followed by
477 // it's (lowered) derived pointer. i.e
478 // (base[0], ptr[0], base[1], ptr[1], ...)
479 for (unsigned i = 0; i < SI.Bases.size(); ++i) {
480 const Value *Base = SI.Bases[i];
481 lowerIncomingStatepointValue(Builder.getValue(Base), Ops, Builder);
483 const Value *Ptr = SI.Ptrs[i];
484 lowerIncomingStatepointValue(Builder.getValue(Ptr), Ops, Builder);
487 // If there are any explicit spill slots passed to the statepoint, record
488 // them, but otherwise do not do anything special. These are user provided
489 // allocas and give control over placement to the consumer. In this case,
490 // it is the contents of the slot which may get updated, not the pointer to
492 for (Value *V : SI.GCArgs) {
493 SDValue Incoming = Builder.getValue(V);
494 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
495 // This handles allocas as arguments to the statepoint
496 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
497 Incoming.getValueType()));
501 // Record computed locations for all lowered values.
502 // This can not be embedded in lowering loops as we need to record *all*
503 // values, while previous loops account only values with unique SDValues.
504 const Instruction *StatepointInstr = SI.StatepointInstr;
505 auto &SpillMap = Builder.FuncInfo.StatepointSpillMaps[StatepointInstr];
507 for (const GCRelocateInst *Relocate : SI.GCRelocates) {
508 const Value *V = Relocate->getDerivedPtr();
509 SDValue SDV = Builder.getValue(V);
510 SDValue Loc = Builder.StatepointLowering.getLocation(SDV);
513 SpillMap.SlotMap[V] = cast<FrameIndexSDNode>(Loc)->getIndex();
515 // Record value as visited, but not spilled. This is case for allocas
516 // and constants. For this values we can avoid emitting spill load while
517 // visiting corresponding gc_relocate.
518 // Actually we do not need to record them in this map at all.
519 // We do this only to check that we are not relocating any unvisited
521 SpillMap.SlotMap[V] = None;
523 // Default llvm mechanisms for exporting values which are used in
524 // different basic blocks does not work for gc relocates.
525 // Note that it would be incorrect to teach llvm that all relocates are
526 // uses of the corresponding values so that it would automatically
527 // export them. Relocates of the spilled values does not use original
529 if (Relocate->getParent() != StatepointInstr->getParent())
530 Builder.ExportFromCurrentBlock(V);
535 SDValue SelectionDAGBuilder::LowerAsSTATEPOINT(
536 SelectionDAGBuilder::StatepointLoweringInfo &SI) {
537 // The basic scheme here is that information about both the original call and
538 // the safepoint is encoded in the CallInst. We create a temporary call and
539 // lower it, then reverse engineer the calling sequence.
543 StatepointLowering.startNewStatepoint(*this);
546 // We schedule gc relocates before removeDuplicateGCPtrs since we _will_
547 // encounter the duplicate gc relocates we elide in removeDuplicateGCPtrs.
548 for (auto *Reloc : SI.GCRelocates)
549 if (Reloc->getParent() == SI.StatepointInstr->getParent())
550 StatepointLowering.scheduleRelocCall(*Reloc);
553 // Remove any redundant llvm::Values which map to the same SDValue as another
554 // input. Also has the effect of removing duplicates in the original
555 // llvm::Value input list as well. This is a useful optimization for
556 // reducing the size of the StackMap section. It has no other impact.
557 removeDuplicateGCPtrs(SI.Bases, SI.Ptrs, SI.GCRelocates, *this,
558 FuncInfo.StatepointSpillMaps[SI.StatepointInstr]);
559 assert(SI.Bases.size() == SI.Ptrs.size() &&
560 SI.Ptrs.size() == SI.GCRelocates.size());
562 // Lower statepoint vmstate and gcstate arguments
563 SmallVector<SDValue, 10> LoweredMetaArgs;
564 lowerStatepointMetaArgs(LoweredMetaArgs, SI, *this);
566 // Now that we've emitted the spills, we need to update the root so that the
567 // call sequence is ordered correctly.
568 SI.CLI.setChain(getRoot());
570 // Get call node, we will replace it later with statepoint
573 std::tie(ReturnVal, CallNode) =
574 lowerCallFromStatepointLoweringInfo(SI, *this, PendingExports);
576 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
577 // nodes with all the appropriate arguments and return values.
579 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
580 SDValue Chain = CallNode->getOperand(0);
583 bool CallHasIncomingGlue = CallNode->getGluedNode();
584 if (CallHasIncomingGlue) {
585 // Glue is always last operand
586 Glue = CallNode->getOperand(CallNode->getNumOperands() - 1);
589 // Build the GC_TRANSITION_START node if necessary.
591 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
592 // order in which they appear in the call to the statepoint intrinsic. If
593 // any of the operands is a pointer-typed, that operand is immediately
594 // followed by a SRCVALUE for the pointer that may be used during lowering
595 // (e.g. to form MachinePointerInfo values for loads/stores).
596 const bool IsGCTransition =
597 (SI.StatepointFlags & (uint64_t)StatepointFlags::GCTransition) ==
598 (uint64_t)StatepointFlags::GCTransition;
599 if (IsGCTransition) {
600 SmallVector<SDValue, 8> TSOps;
603 TSOps.push_back(Chain);
605 // Add GC transition arguments
606 for (const Value *V : SI.GCTransitionArgs) {
607 TSOps.push_back(getValue(V));
608 if (V->getType()->isPointerTy())
609 TSOps.push_back(DAG.getSrcValue(V));
612 // Add glue if necessary
613 if (CallHasIncomingGlue)
614 TSOps.push_back(Glue);
616 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
618 SDValue GCTransitionStart =
619 DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps);
621 Chain = GCTransitionStart.getValue(0);
622 Glue = GCTransitionStart.getValue(1);
625 // TODO: Currently, all of these operands are being marked as read/write in
626 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
627 // and flags to be read-only.
628 SmallVector<SDValue, 40> Ops;
630 // Add the <id> and <numBytes> constants.
631 Ops.push_back(DAG.getTargetConstant(SI.ID, getCurSDLoc(), MVT::i64));
633 DAG.getTargetConstant(SI.NumPatchBytes, getCurSDLoc(), MVT::i32));
635 // Calculate and push starting position of vmstate arguments
636 // Get number of arguments incoming directly into call node
637 unsigned NumCallRegArgs =
638 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
639 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32));
642 SDValue CallTarget = SDValue(CallNode->getOperand(1).getNode(), 0);
643 Ops.push_back(CallTarget);
645 // Add call arguments
646 // Get position of register mask in the call
647 SDNode::op_iterator RegMaskIt;
648 if (CallHasIncomingGlue)
649 RegMaskIt = CallNode->op_end() - 2;
651 RegMaskIt = CallNode->op_end() - 1;
652 Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt);
654 // Add a constant argument for the calling convention
655 pushStackMapConstant(Ops, *this, SI.CLI.CallConv);
657 // Add a constant argument for the flags
658 uint64_t Flags = SI.StatepointFlags;
659 assert(((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0) &&
660 "Unknown flag used");
661 pushStackMapConstant(Ops, *this, Flags);
663 // Insert all vmstate and gcstate arguments
664 Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end());
666 // Add register mask from call node
667 Ops.push_back(*RegMaskIt);
670 Ops.push_back(Chain);
672 // Same for the glue, but we add it only if original call had it
676 // Compute return values. Provide a glue output since we consume one as
677 // input. This allows someone else to chain off us as needed.
678 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
680 SDNode *StatepointMCNode =
681 DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
683 SDNode *SinkNode = StatepointMCNode;
685 // Build the GC_TRANSITION_END node if necessary.
687 // See the comment above regarding GC_TRANSITION_START for the layout of
688 // the operands to the GC_TRANSITION_END node.
689 if (IsGCTransition) {
690 SmallVector<SDValue, 8> TEOps;
693 TEOps.push_back(SDValue(StatepointMCNode, 0));
695 // Add GC transition arguments
696 for (const Value *V : SI.GCTransitionArgs) {
697 TEOps.push_back(getValue(V));
698 if (V->getType()->isPointerTy())
699 TEOps.push_back(DAG.getSrcValue(V));
703 TEOps.push_back(SDValue(StatepointMCNode, 1));
705 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
707 SDValue GCTransitionStart =
708 DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps);
710 SinkNode = GCTransitionStart.getNode();
713 // Replace original call
714 DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root
715 // Remove original call node
716 DAG.DeleteNode(CallNode);
718 // DON'T set the root - under the assumption that it's already set past the
719 // inserted node we created.
721 // TODO: A better future implementation would be to emit a single variable
722 // argument, variable return value STATEPOINT node here and then hookup the
723 // return value of each gc.relocate to the respective output of the
724 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
725 // to actually be possible today.
731 SelectionDAGBuilder::LowerStatepoint(ImmutableStatepoint ISP,
732 const BasicBlock *EHPadBB /*= nullptr*/) {
733 assert(ISP.getCallSite().getCallingConv() != CallingConv::AnyReg &&
734 "anyregcc is not supported on statepoints!");
737 // If this is a malformed statepoint, report it early to simplify debugging.
738 // This should catch any IR level mistake that's made when constructing or
739 // transforming statepoints.
742 // Check that the associated GCStrategy expects to encounter statepoints.
743 assert(GFI->getStrategy().useStatepoints() &&
744 "GCStrategy does not expect to encounter statepoints");
747 SDValue ActualCallee;
749 if (ISP.getNumPatchBytes() > 0) {
750 // If we've been asked to emit a nop sequence instead of a call instruction
751 // for this statepoint then don't lower the call target, but use a constant
752 // `null` instead. Not lowering the call target lets statepoint clients get
753 // away without providing a physical address for the symbolic call target at
756 const auto &TLI = DAG.getTargetLoweringInfo();
757 const auto &DL = DAG.getDataLayout();
759 unsigned AS = ISP.getCalledValue()->getType()->getPointerAddressSpace();
760 ActualCallee = DAG.getConstant(0, getCurSDLoc(), TLI.getPointerTy(DL, AS));
762 ActualCallee = getValue(ISP.getCalledValue());
765 StatepointLoweringInfo SI(DAG);
766 populateCallLoweringInfo(SI.CLI, ISP.getCallSite(),
767 ImmutableStatepoint::CallArgsBeginPos,
768 ISP.getNumCallArgs(), ActualCallee,
769 ISP.getActualReturnType(), false /* IsPatchPoint */);
771 for (const GCRelocateInst *Relocate : ISP.getRelocates()) {
772 SI.GCRelocates.push_back(Relocate);
773 SI.Bases.push_back(Relocate->getBasePtr());
774 SI.Ptrs.push_back(Relocate->getDerivedPtr());
777 SI.GCArgs = ArrayRef<const Use>(ISP.gc_args_begin(), ISP.gc_args_end());
778 SI.StatepointInstr = ISP.getInstruction();
779 SI.GCTransitionArgs =
780 ArrayRef<const Use>(ISP.gc_args_begin(), ISP.gc_args_end());
782 SI.DeoptState = ArrayRef<const Use>(ISP.vm_state_begin(), ISP.vm_state_end());
783 SI.StatepointFlags = ISP.getFlags();
784 SI.NumPatchBytes = ISP.getNumPatchBytes();
785 SI.EHPadBB = EHPadBB;
787 SDValue ReturnValue = LowerAsSTATEPOINT(SI);
789 // Export the result value if needed
790 const GCResultInst *GCResult = ISP.getGCResult();
791 Type *RetTy = ISP.getActualReturnType();
792 if (!RetTy->isVoidTy() && GCResult) {
793 if (GCResult->getParent() != ISP.getCallSite().getParent()) {
794 // Result value will be used in a different basic block so we need to
795 // export it now. Default exporting mechanism will not work here because
796 // statepoint call has a different type than the actual call. It means
797 // that by default llvm will create export register of the wrong type
798 // (always i32 in our case). So instead we need to create export register
799 // with correct type manually.
800 // TODO: To eliminate this problem we can remove gc.result intrinsics
801 // completely and make statepoint call to return a tuple.
802 unsigned Reg = FuncInfo.CreateRegs(RetTy);
803 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
804 DAG.getDataLayout(), Reg, RetTy);
805 SDValue Chain = DAG.getEntryNode();
807 RFV.getCopyToRegs(ReturnValue, DAG, getCurSDLoc(), Chain, nullptr);
808 PendingExports.push_back(Chain);
809 FuncInfo.ValueMap[ISP.getInstruction()] = Reg;
811 // Result value will be used in a same basic block. Don't export it or
812 // perform any explicit register copies.
813 // We'll replace the actuall call node shortly. gc_result will grab
815 setValue(ISP.getInstruction(), ReturnValue);
818 // The token value is never used from here on, just generate a poison value
819 setValue(ISP.getInstruction(), DAG.getIntPtrConstant(-1, getCurSDLoc()));
823 void SelectionDAGBuilder::LowerCallSiteWithDeoptBundleImpl(
824 ImmutableCallSite CS, SDValue Callee, const BasicBlock *EHPadBB,
825 bool VarArgDisallowed, bool ForceVoidReturnTy) {
826 StatepointLoweringInfo SI(DAG);
827 unsigned ArgBeginIndex = CS.arg_begin() - CS.getInstruction()->op_begin();
828 populateCallLoweringInfo(
829 SI.CLI, CS, ArgBeginIndex, CS.getNumArgOperands(), Callee,
830 ForceVoidReturnTy ? Type::getVoidTy(*DAG.getContext()) : CS.getType(),
832 if (!VarArgDisallowed)
833 SI.CLI.IsVarArg = CS.getFunctionType()->isVarArg();
835 auto DeoptBundle = *CS.getOperandBundle(LLVMContext::OB_deopt);
837 unsigned DefaultID = StatepointDirectives::DeoptBundleStatepointID;
839 auto SD = parseStatepointDirectivesFromAttrs(CS.getAttributes());
840 SI.ID = SD.StatepointID.getValueOr(DefaultID);
841 SI.NumPatchBytes = SD.NumPatchBytes.getValueOr(0);
844 ArrayRef<const Use>(DeoptBundle.Inputs.begin(), DeoptBundle.Inputs.end());
845 SI.StatepointFlags = static_cast<uint64_t>(StatepointFlags::None);
846 SI.EHPadBB = EHPadBB;
848 // NB! The GC arguments are deliberately left empty.
850 if (SDValue ReturnVal = LowerAsSTATEPOINT(SI)) {
851 const Instruction *Inst = CS.getInstruction();
852 ReturnVal = lowerRangeToAssertZExt(DAG, *Inst, ReturnVal);
853 setValue(Inst, ReturnVal);
857 void SelectionDAGBuilder::LowerCallSiteWithDeoptBundle(
858 ImmutableCallSite CS, SDValue Callee, const BasicBlock *EHPadBB) {
859 LowerCallSiteWithDeoptBundleImpl(CS, Callee, EHPadBB,
860 /* VarArgDisallowed = */ false,
861 /* ForceVoidReturnTy = */ false);
864 void SelectionDAGBuilder::visitGCResult(const GCResultInst &CI) {
865 // The result value of the gc_result is simply the result of the actual
866 // call. We've already emitted this, so just grab the value.
867 const Instruction *I = CI.getStatepoint();
869 if (I->getParent() != CI.getParent()) {
870 // Statepoint is in different basic block so we should have stored call
871 // result in a virtual register.
872 // We can not use default getValue() functionality to copy value from this
873 // register because statepoint and actual call return types can be
874 // different, and getValue() will use CopyFromReg of the wrong type,
875 // which is always i32 in our case.
876 PointerType *CalleeType = cast<PointerType>(
877 ImmutableStatepoint(I).getCalledValue()->getType());
879 cast<FunctionType>(CalleeType->getElementType())->getReturnType();
880 SDValue CopyFromReg = getCopyFromRegs(I, RetTy);
882 assert(CopyFromReg.getNode());
883 setValue(&CI, CopyFromReg);
885 setValue(&CI, getValue(I));
889 void SelectionDAGBuilder::visitGCRelocate(const GCRelocateInst &Relocate) {
892 // We skip this check for relocates not in the same basic block as thier
893 // statepoint. It would be too expensive to preserve validation info through
894 // different basic blocks.
895 if (Relocate.getStatepoint()->getParent() == Relocate.getParent())
896 StatepointLowering.relocCallVisited(Relocate);
898 auto *Ty = Relocate.getType()->getScalarType();
899 if (auto IsManaged = GFI->getStrategy().isGCManagedPointer(Ty))
900 assert(*IsManaged && "Non gc managed pointer relocated!");
903 const Value *DerivedPtr = Relocate.getDerivedPtr();
904 SDValue SD = getValue(DerivedPtr);
906 auto &SpillMap = FuncInfo.StatepointSpillMaps[Relocate.getStatepoint()];
907 auto SlotIt = SpillMap.find(DerivedPtr);
908 assert(SlotIt != SpillMap.end() && "Relocating not lowered gc value");
909 Optional<int> DerivedPtrLocation = SlotIt->second;
911 // We didn't need to spill these special cases (constants and allocas).
912 // See the handling in spillIncomingValueForStatepoint for detail.
913 if (!DerivedPtrLocation) {
914 setValue(&Relocate, SD);
918 SDValue SpillSlot = DAG.getTargetFrameIndex(*DerivedPtrLocation,
921 // Be conservative: flush all pending loads
922 // TODO: Probably we can be less restrictive on this,
923 // it may allow more scheduling opportunities.
924 SDValue Chain = getRoot();
927 DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, SpillSlot,
928 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
929 *DerivedPtrLocation));
931 // Again, be conservative, don't emit pending loads
932 DAG.setRoot(SpillLoad.getValue(1));
934 assert(SpillLoad.getNode());
935 setValue(&Relocate, SpillLoad);
938 void SelectionDAGBuilder::LowerDeoptimizeCall(const CallInst *CI) {
939 const auto &TLI = DAG.getTargetLoweringInfo();
940 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(RTLIB::DEOPTIMIZE),
941 TLI.getPointerTy(DAG.getDataLayout()));
943 // We don't lower calls to __llvm_deoptimize as varargs, but as a regular
944 // call. We also do not lower the return value to any virtual register, and
945 // change the immediately following return to a trap instruction.
946 LowerCallSiteWithDeoptBundleImpl(CI, Callee, /* EHPadBB = */ nullptr,
947 /* VarArgDisallowed = */ true,
948 /* ForceVoidReturnTy = */ true);
951 void SelectionDAGBuilder::LowerDeoptimizingReturn() {
952 // We do not lower the return value from llvm.deoptimize to any virtual
953 // register, and change the immediately following return to a trap
955 if (DAG.getTarget().Options.TrapUnreachable)
957 DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));