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/GCMetadata.h"
21 #include "llvm/CodeGen/GCStrategy.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/StackMaps.h"
24 #include "llvm/IR/CallingConv.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/Statepoint.h"
29 #include "llvm/Target/TargetLowering.h"
33 #define DEBUG_TYPE "statepoint-lowering"
35 STATISTIC(NumSlotsAllocatedForStatepoints,
36 "Number of stack slots allocated for statepoints");
37 STATISTIC(NumOfStatepoints, "Number of statepoint nodes encountered");
38 STATISTIC(StatepointMaxSlotsRequired,
39 "Maximum number of stack slots required for a singe statepoint");
41 static void pushStackMapConstant(SmallVectorImpl<SDValue>& Ops,
42 SelectionDAGBuilder &Builder, uint64_t Value) {
43 SDLoc L = Builder.getCurSDLoc();
44 Ops.push_back(Builder.DAG.getTargetConstant(StackMaps::ConstantOp, L,
46 Ops.push_back(Builder.DAG.getTargetConstant(Value, L, MVT::i64));
49 void StatepointLoweringState::startNewStatepoint(SelectionDAGBuilder &Builder) {
51 assert(PendingGCRelocateCalls.empty() &&
52 "Trying to visit statepoint before finished processing previous one");
54 NextSlotToAllocate = 0;
55 // Need to resize this on each safepoint - we need the two to stay in
56 // sync and the clear patterns of a SelectionDAGBuilder have no relation
57 // to FunctionLoweringInfo.
58 AllocatedStackSlots.resize(Builder.FuncInfo.StatepointStackSlots.size());
59 for (size_t i = 0; i < AllocatedStackSlots.size(); i++) {
60 AllocatedStackSlots[i] = false;
64 void StatepointLoweringState::clear() {
66 AllocatedStackSlots.clear();
67 assert(PendingGCRelocateCalls.empty() &&
68 "cleared before statepoint sequence completed");
72 StatepointLoweringState::allocateStackSlot(EVT ValueType,
73 SelectionDAGBuilder &Builder) {
75 NumSlotsAllocatedForStatepoints++;
77 // The basic scheme here is to first look for a previously created stack slot
78 // which is not in use (accounting for the fact arbitrary slots may already
79 // be reserved), or to create a new stack slot and use it.
81 // If this doesn't succeed in 40000 iterations, something is seriously wrong
82 for (int i = 0; i < 40000; i++) {
83 assert(Builder.FuncInfo.StatepointStackSlots.size() ==
84 AllocatedStackSlots.size() &&
86 const size_t NumSlots = AllocatedStackSlots.size();
87 assert(NextSlotToAllocate <= NumSlots && "broken invariant");
89 if (NextSlotToAllocate >= NumSlots) {
90 assert(NextSlotToAllocate == NumSlots);
92 if (NumSlots + 1 > StatepointMaxSlotsRequired) {
93 StatepointMaxSlotsRequired = NumSlots + 1;
96 SDValue SpillSlot = Builder.DAG.CreateStackTemporary(ValueType);
97 const unsigned FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
98 Builder.FuncInfo.StatepointStackSlots.push_back(FI);
99 AllocatedStackSlots.push_back(true);
102 if (!AllocatedStackSlots[NextSlotToAllocate]) {
103 const int FI = Builder.FuncInfo.StatepointStackSlots[NextSlotToAllocate];
104 AllocatedStackSlots[NextSlotToAllocate] = true;
105 return Builder.DAG.getFrameIndex(FI, ValueType);
107 // Note: We deliberately choose to advance this only on the failing path.
108 // Doing so on the succeeding path involves a bit of complexity that caused
109 // a minor bug previously. Unless performance shows this matters, please
110 // keep this code as simple as possible.
111 NextSlotToAllocate++;
113 llvm_unreachable("infinite loop?");
116 /// Utility function for reservePreviousStackSlotForValue. Tries to find
117 /// stack slot index to which we have spilled value for previous statepoints.
118 /// LookUpDepth specifies maximum DFS depth this function is allowed to look.
119 static Optional<int> findPreviousSpillSlot(const Value *Val,
120 SelectionDAGBuilder &Builder,
122 // Can not look any further - give up now
123 if (LookUpDepth <= 0)
124 return Optional<int>();
126 // Spill location is known for gc relocates
127 if (isGCRelocate(Val)) {
128 GCRelocateOperands RelocOps(cast<Instruction>(Val));
130 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
131 Builder.FuncInfo.StatepointRelocatedValues[RelocOps.getStatepoint()];
133 auto It = SpillMap.find(RelocOps.getDerivedPtr());
134 if (It == SpillMap.end())
135 return Optional<int>();
140 // Look through bitcast instructions.
141 if (const BitCastInst *Cast = dyn_cast<BitCastInst>(Val)) {
142 return findPreviousSpillSlot(Cast->getOperand(0), Builder, LookUpDepth - 1);
145 // Look through phi nodes
146 // All incoming values should have same known stack slot, otherwise result
148 if (const PHINode *Phi = dyn_cast<PHINode>(Val)) {
149 Optional<int> MergedResult = None;
151 for (auto &IncomingValue : Phi->incoming_values()) {
152 Optional<int> SpillSlot =
153 findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth - 1);
154 if (!SpillSlot.hasValue())
155 return Optional<int>();
157 if (MergedResult.hasValue() && *MergedResult != *SpillSlot)
158 return Optional<int>();
160 MergedResult = SpillSlot;
165 // TODO: We can do better for PHI nodes. In cases like this:
166 // ptr = phi(relocated_pointer, not_relocated_pointer)
168 // We will return that stack slot for ptr is unknown. And later we might
169 // assign different stack slots for ptr and relocated_pointer. This limits
170 // llvm's ability to remove redundant stores.
171 // Unfortunately it's hard to accomplish in current infrastructure.
172 // We use this function to eliminate spill store completely, while
173 // in example we still need to emit store, but instead of any location
174 // we need to use special "preferred" location.
176 // TODO: handle simple updates. If a value is modified and the original
177 // value is no longer live, it would be nice to put the modified value in the
178 // same slot. This allows folding of the memory accesses for some
179 // instructions types (like an increment).
183 // However we need to be careful for cases like this:
187 // Here we want to reserve spill slot for 'i', but not for 'i+1'. If we just
188 // put handling of simple modifications in this function like it's done
189 // for bitcasts we might end up reserving i's slot for 'i+1' because order in
190 // which we visit values is unspecified.
192 // Don't know any information about this instruction
193 return Optional<int>();
196 /// Try to find existing copies of the incoming values in stack slots used for
197 /// statepoint spilling. If we can find a spill slot for the incoming value,
198 /// mark that slot as allocated, and reuse the same slot for this safepoint.
199 /// This helps to avoid series of loads and stores that only serve to reshuffle
200 /// values on the stack between calls.
201 static void reservePreviousStackSlotForValue(const Value *IncomingValue,
202 SelectionDAGBuilder &Builder) {
204 SDValue Incoming = Builder.getValue(IncomingValue);
206 if (isa<ConstantSDNode>(Incoming) || isa<FrameIndexSDNode>(Incoming)) {
207 // We won't need to spill this, so no need to check for previously
208 // allocated stack slots
212 SDValue OldLocation = Builder.StatepointLowering.getLocation(Incoming);
213 if (OldLocation.getNode())
214 // duplicates in input
217 const int LookUpDepth = 6;
218 Optional<int> Index =
219 findPreviousSpillSlot(IncomingValue, Builder, LookUpDepth);
220 if (!Index.hasValue())
223 auto Itr = std::find(Builder.FuncInfo.StatepointStackSlots.begin(),
224 Builder.FuncInfo.StatepointStackSlots.end(), *Index);
225 assert(Itr != Builder.FuncInfo.StatepointStackSlots.end() &&
226 "value spilled to the unknown stack slot");
228 // This is one of our dedicated lowering slots
230 std::distance(Builder.FuncInfo.StatepointStackSlots.begin(), Itr);
231 if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) {
232 // stack slot already assigned to someone else, can't use it!
233 // TODO: currently we reserve space for gc arguments after doing
234 // normal allocation for deopt arguments. We should reserve for
235 // _all_ deopt and gc arguments, then start allocating. This
236 // will prevent some moves being inserted when vm state changes,
237 // but gc state doesn't between two calls.
240 // Reserve this stack slot
241 Builder.StatepointLowering.reserveStackSlot(Offset);
243 // Cache this slot so we find it when going through the normal
245 SDValue Loc = Builder.DAG.getTargetFrameIndex(*Index, Incoming.getValueType());
246 Builder.StatepointLowering.setLocation(Incoming, Loc);
249 /// Remove any duplicate (as SDValues) from the derived pointer pairs. This
250 /// is not required for correctness. It's purpose is to reduce the size of
251 /// StackMap section. It has no effect on the number of spill slots required
252 /// or the actual lowering.
253 static void removeDuplicatesGCPtrs(SmallVectorImpl<const Value *> &Bases,
254 SmallVectorImpl<const Value *> &Ptrs,
255 SmallVectorImpl<const Value *> &Relocs,
256 SelectionDAGBuilder &Builder) {
258 // This is horribly inefficient, but I don't care right now
259 SmallSet<SDValue, 64> Seen;
261 SmallVector<const Value *, 64> NewBases, NewPtrs, NewRelocs;
262 for (size_t i = 0; i < Ptrs.size(); i++) {
263 SDValue SD = Builder.getValue(Ptrs[i]);
264 // Only add non-duplicates
265 if (Seen.count(SD) == 0) {
266 NewBases.push_back(Bases[i]);
267 NewPtrs.push_back(Ptrs[i]);
268 NewRelocs.push_back(Relocs[i]);
272 assert(Bases.size() >= NewBases.size());
273 assert(Ptrs.size() >= NewPtrs.size());
274 assert(Relocs.size() >= NewRelocs.size());
278 assert(Ptrs.size() == Bases.size());
279 assert(Ptrs.size() == Relocs.size());
282 /// Extract call from statepoint, lower it and return pointer to the
283 /// call node. Also update NodeMap so that getValue(statepoint) will
284 /// reference lowered call result
286 lowerCallFromStatepoint(ImmutableStatepoint ISP, const BasicBlock *EHPadBB,
287 SelectionDAGBuilder &Builder,
288 SmallVectorImpl<SDValue> &PendingExports) {
290 ImmutableCallSite CS(ISP.getCallSite());
292 SDValue ActualCallee;
294 if (ISP.getNumPatchBytes() > 0) {
295 // If we've been asked to emit a nop sequence instead of a call instruction
296 // for this statepoint then don't lower the call target, but use a constant
297 // `null` instead. Not lowering the call target lets statepoint clients get
298 // away without providing a physical address for the symbolic call target at
301 const auto &TLI = Builder.DAG.getTargetLoweringInfo();
302 const auto &DL = Builder.DAG.getDataLayout();
304 unsigned AS = ISP.getCalledValue()->getType()->getPointerAddressSpace();
305 ActualCallee = Builder.DAG.getConstant(0, Builder.getCurSDLoc(),
306 TLI.getPointerTy(DL, AS));
308 ActualCallee = Builder.getValue(ISP.getCalledValue());
310 assert(CS.getCallingConv() != CallingConv::AnyReg &&
311 "anyregcc is not supported on statepoints!");
313 Type *DefTy = ISP.getActualReturnType();
314 bool HasDef = !DefTy->isVoidTy();
316 SDValue ReturnValue, CallEndVal;
317 std::tie(ReturnValue, CallEndVal) = Builder.lowerCallOperands(
318 ISP.getCallSite(), ImmutableStatepoint::CallArgsBeginPos,
319 ISP.getNumCallArgs(), ActualCallee, DefTy, EHPadBB,
320 false /* IsPatchPoint */);
322 SDNode *CallEnd = CallEndVal.getNode();
324 // Get a call instruction from the call sequence chain. Tail calls are not
325 // allowed. The following code is essentially reverse engineering X86's
328 // We are expecting DAG to have the following form:
330 // ch = eh_label (only in case of invoke statepoint)
331 // ch, glue = callseq_start ch
332 // ch, glue = X86::Call ch, glue
333 // ch, glue = callseq_end ch, glue
334 // get_return_value ch, glue
336 // get_return_value can either be a CopyFromReg to grab the return value from
337 // %RAX, or it can be a LOAD to load a value returned by reference via a stack
340 if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg ||
341 CallEnd->getOpcode() == ISD::LOAD))
342 CallEnd = CallEnd->getOperand(0).getNode();
344 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
346 // Export the result value if needed
347 const Instruction *GCResult = ISP.getGCResult();
348 if (HasDef && GCResult) {
349 if (GCResult->getParent() != CS.getParent()) {
350 // Result value will be used in a different basic block so we need to
352 // Default exporting mechanism will not work here because statepoint call
353 // has a different type than the actual call. It means that by default
354 // llvm will create export register of the wrong type (always i32 in our
355 // case). So instead we need to create export register with correct type
357 // TODO: To eliminate this problem we can remove gc.result intrinsics
358 // completely and make statepoint call to return a tuple.
359 unsigned Reg = Builder.FuncInfo.CreateRegs(ISP.getActualReturnType());
361 *Builder.DAG.getContext(), Builder.DAG.getTargetLoweringInfo(),
362 Builder.DAG.getDataLayout(), Reg, ISP.getActualReturnType());
363 SDValue Chain = Builder.DAG.getEntryNode();
365 RFV.getCopyToRegs(ReturnValue, Builder.DAG, Builder.getCurSDLoc(), Chain,
367 PendingExports.push_back(Chain);
368 Builder.FuncInfo.ValueMap[CS.getInstruction()] = Reg;
370 // Result value will be used in a same basic block. Don't export it or
371 // perform any explicit register copies.
372 // We'll replace the actuall call node shortly. gc_result will grab
374 Builder.setValue(CS.getInstruction(), ReturnValue);
377 // The token value is never used from here on, just generate a poison value
378 Builder.setValue(CS.getInstruction(),
379 Builder.DAG.getIntPtrConstant(-1, Builder.getCurSDLoc()));
382 return CallEnd->getOperand(0).getNode();
385 /// Callect all gc pointers coming into statepoint intrinsic, clean them up,
386 /// and return two arrays:
387 /// Bases - base pointers incoming to this statepoint
388 /// Ptrs - derived pointers incoming to this statepoint
389 /// Relocs - the gc_relocate corresponding to each base/ptr pair
390 /// Elements of this arrays should be in one-to-one correspondence with each
391 /// other i.e Bases[i], Ptrs[i] are from the same gcrelocate call
392 static void getIncomingStatepointGCValues(
393 SmallVectorImpl<const Value *> &Bases, SmallVectorImpl<const Value *> &Ptrs,
394 SmallVectorImpl<const Value *> &Relocs, ImmutableStatepoint StatepointSite,
395 SelectionDAGBuilder &Builder) {
396 for (GCRelocateOperands relocateOpers : StatepointSite.getRelocates()) {
397 Relocs.push_back(relocateOpers.getUnderlyingCallSite().getInstruction());
398 Bases.push_back(relocateOpers.getBasePtr());
399 Ptrs.push_back(relocateOpers.getDerivedPtr());
402 // Remove any redundant llvm::Values which map to the same SDValue as another
403 // input. Also has the effect of removing duplicates in the original
404 // llvm::Value input list as well. This is a useful optimization for
405 // reducing the size of the StackMap section. It has no other impact.
406 removeDuplicatesGCPtrs(Bases, Ptrs, Relocs, Builder);
408 assert(Bases.size() == Ptrs.size() && Ptrs.size() == Relocs.size());
411 /// Spill a value incoming to the statepoint. It might be either part of
413 /// or gcstate. In both cases unconditionally spill it on the stack unless it
414 /// is a null constant. Return pair with first element being frame index
415 /// containing saved value and second element with outgoing chain from the
417 static std::pair<SDValue, SDValue>
418 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
419 SelectionDAGBuilder &Builder) {
420 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
422 // Emit new store if we didn't do it for this ptr before
423 if (!Loc.getNode()) {
424 Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
426 assert(isa<FrameIndexSDNode>(Loc));
427 int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
428 // We use TargetFrameIndex so that isel will not select it into LEA
429 Loc = Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
431 // TODO: We can create TokenFactor node instead of
432 // chaining stores one after another, this may allow
433 // a bit more optimal scheduling for them
434 Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
435 MachinePointerInfo::getFixedStack(
436 Builder.DAG.getMachineFunction(), Index),
439 Builder.StatepointLowering.setLocation(Incoming, Loc);
442 assert(Loc.getNode());
443 return std::make_pair(Loc, Chain);
446 /// Lower a single value incoming to a statepoint node. This value can be
447 /// either a deopt value or a gc value, the handling is the same. We special
448 /// case constants and allocas, then fall back to spilling if required.
449 static void lowerIncomingStatepointValue(SDValue Incoming,
450 SmallVectorImpl<SDValue> &Ops,
451 SelectionDAGBuilder &Builder) {
452 SDValue Chain = Builder.getRoot();
454 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
455 // If the original value was a constant, make sure it gets recorded as
456 // such in the stackmap. This is required so that the consumer can
457 // parse any internal format to the deopt state. It also handles null
458 // pointers and other constant pointers in GC states
459 pushStackMapConstant(Ops, Builder, C->getSExtValue());
460 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
461 // This handles allocas as arguments to the statepoint (this is only
462 // really meaningful for a deopt value. For GC, we'd be trying to
463 // relocate the address of the alloca itself?)
464 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
465 Incoming.getValueType()));
467 // Otherwise, locate a spill slot and explicitly spill it so it
468 // can be found by the runtime later. We currently do not support
469 // tracking values through callee saved registers to their eventual
470 // spill location. This would be a useful optimization, but would
471 // need to be optional since it requires a lot of complexity on the
472 // runtime side which not all would support.
473 std::pair<SDValue, SDValue> Res =
474 spillIncomingStatepointValue(Incoming, Chain, Builder);
475 Ops.push_back(Res.first);
479 Builder.DAG.setRoot(Chain);
482 /// Lower deopt state and gc pointer arguments of the statepoint. The actual
483 /// lowering is described in lowerIncomingStatepointValue. This function is
484 /// responsible for lowering everything in the right position and playing some
485 /// tricks to avoid redundant stack manipulation where possible. On
486 /// completion, 'Ops' will contain ready to use operands for machine code
487 /// statepoint. The chain nodes will have already been created and the DAG root
488 /// will be set to the last value spilled (if any were).
489 static void lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
490 ImmutableStatepoint StatepointSite,
491 SelectionDAGBuilder &Builder) {
493 // Lower the deopt and gc arguments for this statepoint. Layout will
494 // be: deopt argument length, deopt arguments.., gc arguments...
496 SmallVector<const Value *, 64> Bases, Ptrs, Relocations;
497 getIncomingStatepointGCValues(Bases, Ptrs, Relocations, StatepointSite,
501 // Check that each of the gc pointer and bases we've gotten out of the
502 // safepoint is something the strategy thinks might be a pointer into the GC
503 // heap. This is basically just here to help catch errors during statepoint
504 // insertion. TODO: This should actually be in the Verifier, but we can't get
505 // to the GCStrategy from there (yet).
506 GCStrategy &S = Builder.GFI->getStrategy();
507 for (const Value *V : Bases) {
508 auto Opt = S.isGCManagedPointer(V);
509 if (Opt.hasValue()) {
510 assert(Opt.getValue() &&
511 "non gc managed base pointer found in statepoint");
514 for (const Value *V : Ptrs) {
515 auto Opt = S.isGCManagedPointer(V);
516 if (Opt.hasValue()) {
517 assert(Opt.getValue() &&
518 "non gc managed derived pointer found in statepoint");
521 for (const Value *V : Relocations) {
522 auto Opt = S.isGCManagedPointer(V);
523 if (Opt.hasValue()) {
524 assert(Opt.getValue() && "non gc managed pointer relocated");
529 // Before we actually start lowering (and allocating spill slots for values),
530 // reserve any stack slots which we judge to be profitable to reuse for a
531 // particular value. This is purely an optimization over the code below and
532 // doesn't change semantics at all. It is important for performance that we
533 // reserve slots for both deopt and gc values before lowering either.
534 for (const Value *V : StatepointSite.vm_state_args()) {
535 reservePreviousStackSlotForValue(V, Builder);
537 for (unsigned i = 0; i < Bases.size(); ++i) {
538 reservePreviousStackSlotForValue(Bases[i], Builder);
539 reservePreviousStackSlotForValue(Ptrs[i], Builder);
542 // First, prefix the list with the number of unique values to be
543 // lowered. Note that this is the number of *Values* not the
544 // number of SDValues required to lower them.
545 const int NumVMSArgs = StatepointSite.getNumTotalVMSArgs();
546 pushStackMapConstant(Ops, Builder, NumVMSArgs);
548 assert(NumVMSArgs == std::distance(StatepointSite.vm_state_begin(),
549 StatepointSite.vm_state_end()));
551 // The vm state arguments are lowered in an opaque manner. We do
552 // not know what type of values are contained within. We skip the
553 // first one since that happens to be the total number we lowered
554 // explicitly just above. We could have left it in the loop and
555 // not done it explicitly, but it's far easier to understand this
557 for (const Value *V : StatepointSite.vm_state_args()) {
558 SDValue Incoming = Builder.getValue(V);
559 lowerIncomingStatepointValue(Incoming, Ops, Builder);
562 // Finally, go ahead and lower all the gc arguments. There's no prefixed
563 // length for this one. After lowering, we'll have the base and pointer
564 // arrays interwoven with each (lowered) base pointer immediately followed by
565 // it's (lowered) derived pointer. i.e
566 // (base[0], ptr[0], base[1], ptr[1], ...)
567 for (unsigned i = 0; i < Bases.size(); ++i) {
568 const Value *Base = Bases[i];
569 lowerIncomingStatepointValue(Builder.getValue(Base), Ops, Builder);
571 const Value *Ptr = Ptrs[i];
572 lowerIncomingStatepointValue(Builder.getValue(Ptr), Ops, Builder);
575 // If there are any explicit spill slots passed to the statepoint, record
576 // them, but otherwise do not do anything special. These are user provided
577 // allocas and give control over placement to the consumer. In this case,
578 // it is the contents of the slot which may get updated, not the pointer to
580 for (Value *V : StatepointSite.gc_args()) {
581 SDValue Incoming = Builder.getValue(V);
582 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
583 // This handles allocas as arguments to the statepoint
584 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
585 Incoming.getValueType()));
589 // Record computed locations for all lowered values.
590 // This can not be embedded in lowering loops as we need to record *all*
591 // values, while previous loops account only values with unique SDValues.
592 const Instruction *StatepointInstr =
593 StatepointSite.getCallSite().getInstruction();
594 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
595 Builder.FuncInfo.StatepointRelocatedValues[StatepointInstr];
597 for (GCRelocateOperands RelocateOpers : StatepointSite.getRelocates()) {
598 const Value *V = RelocateOpers.getDerivedPtr();
599 SDValue SDV = Builder.getValue(V);
600 SDValue Loc = Builder.StatepointLowering.getLocation(SDV);
603 SpillMap[V] = cast<FrameIndexSDNode>(Loc)->getIndex();
605 // Record value as visited, but not spilled. This is case for allocas
606 // and constants. For this values we can avoid emitting spill load while
607 // visiting corresponding gc_relocate.
608 // Actually we do not need to record them in this map at all.
609 // We do this only to check that we are not relocating any unvisited
613 // Default llvm mechanisms for exporting values which are used in
614 // different basic blocks does not work for gc relocates.
615 // Note that it would be incorrect to teach llvm that all relocates are
616 // uses of the corresponding values so that it would automatically
617 // export them. Relocates of the spilled values does not use original
619 if (RelocateOpers.getUnderlyingCallSite().getParent() !=
620 StatepointInstr->getParent())
621 Builder.ExportFromCurrentBlock(V);
626 void SelectionDAGBuilder::visitStatepoint(const CallInst &CI) {
627 // Check some preconditions for sanity
628 assert(isStatepoint(&CI) &&
629 "function called must be the statepoint function");
631 LowerStatepoint(ImmutableStatepoint(&CI));
634 void SelectionDAGBuilder::LowerStatepoint(
635 ImmutableStatepoint ISP, const BasicBlock *EHPadBB /*= nullptr*/) {
636 // The basic scheme here is that information about both the original call and
637 // the safepoint is encoded in the CallInst. We create a temporary call and
638 // lower it, then reverse engineer the calling sequence.
642 StatepointLowering.startNewStatepoint(*this);
644 ImmutableCallSite CS(ISP.getCallSite());
647 // Consistency check. Check only relocates in the same basic block as thier
649 for (const User *U : CS->users()) {
650 const CallInst *Call = cast<CallInst>(U);
651 if (isGCRelocate(Call) && Call->getParent() == CS.getParent())
652 StatepointLowering.scheduleRelocCall(*Call);
657 // If this is a malformed statepoint, report it early to simplify debugging.
658 // This should catch any IR level mistake that's made when constructing or
659 // transforming statepoints.
662 // Check that the associated GCStrategy expects to encounter statepoints.
663 assert(GFI->getStrategy().useStatepoints() &&
664 "GCStrategy does not expect to encounter statepoints");
667 // Lower statepoint vmstate and gcstate arguments
668 SmallVector<SDValue, 10> LoweredMetaArgs;
669 lowerStatepointMetaArgs(LoweredMetaArgs, ISP, *this);
671 // Get call node, we will replace it later with statepoint
673 lowerCallFromStatepoint(ISP, EHPadBB, *this, PendingExports);
675 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
676 // nodes with all the appropriate arguments and return values.
678 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
679 SDValue Chain = CallNode->getOperand(0);
682 bool CallHasIncomingGlue = CallNode->getGluedNode();
683 if (CallHasIncomingGlue) {
684 // Glue is always last operand
685 Glue = CallNode->getOperand(CallNode->getNumOperands() - 1);
688 // Build the GC_TRANSITION_START node if necessary.
690 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
691 // order in which they appear in the call to the statepoint intrinsic. If
692 // any of the operands is a pointer-typed, that operand is immediately
693 // followed by a SRCVALUE for the pointer that may be used during lowering
694 // (e.g. to form MachinePointerInfo values for loads/stores).
695 const bool IsGCTransition =
696 (ISP.getFlags() & (uint64_t)StatepointFlags::GCTransition) ==
697 (uint64_t)StatepointFlags::GCTransition;
698 if (IsGCTransition) {
699 SmallVector<SDValue, 8> TSOps;
702 TSOps.push_back(Chain);
704 // Add GC transition arguments
705 for (const Value *V : ISP.gc_transition_args()) {
706 TSOps.push_back(getValue(V));
707 if (V->getType()->isPointerTy())
708 TSOps.push_back(DAG.getSrcValue(V));
711 // Add glue if necessary
712 if (CallHasIncomingGlue)
713 TSOps.push_back(Glue);
715 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
717 SDValue GCTransitionStart =
718 DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps);
720 Chain = GCTransitionStart.getValue(0);
721 Glue = GCTransitionStart.getValue(1);
724 // TODO: Currently, all of these operands are being marked as read/write in
725 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
726 // and flags to be read-only.
727 SmallVector<SDValue, 40> Ops;
729 // Add the <id> and <numBytes> constants.
730 Ops.push_back(DAG.getTargetConstant(ISP.getID(), getCurSDLoc(), MVT::i64));
732 DAG.getTargetConstant(ISP.getNumPatchBytes(), getCurSDLoc(), MVT::i32));
734 // Calculate and push starting position of vmstate arguments
735 // Get number of arguments incoming directly into call node
736 unsigned NumCallRegArgs =
737 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
738 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32));
741 SDValue CallTarget = SDValue(CallNode->getOperand(1).getNode(), 0);
742 Ops.push_back(CallTarget);
744 // Add call arguments
745 // Get position of register mask in the call
746 SDNode::op_iterator RegMaskIt;
747 if (CallHasIncomingGlue)
748 RegMaskIt = CallNode->op_end() - 2;
750 RegMaskIt = CallNode->op_end() - 1;
751 Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt);
753 // Add a constant argument for the calling convention
754 pushStackMapConstant(Ops, *this, CS.getCallingConv());
756 // Add a constant argument for the flags
757 uint64_t Flags = ISP.getFlags();
759 ((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0)
760 && "unknown flag used");
761 pushStackMapConstant(Ops, *this, Flags);
763 // Insert all vmstate and gcstate arguments
764 Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end());
766 // Add register mask from call node
767 Ops.push_back(*RegMaskIt);
770 Ops.push_back(Chain);
772 // Same for the glue, but we add it only if original call had it
776 // Compute return values. Provide a glue output since we consume one as
777 // input. This allows someone else to chain off us as needed.
778 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
780 SDNode *StatepointMCNode =
781 DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
783 SDNode *SinkNode = StatepointMCNode;
785 // Build the GC_TRANSITION_END node if necessary.
787 // See the comment above regarding GC_TRANSITION_START for the layout of
788 // the operands to the GC_TRANSITION_END node.
789 if (IsGCTransition) {
790 SmallVector<SDValue, 8> TEOps;
793 TEOps.push_back(SDValue(StatepointMCNode, 0));
795 // Add GC transition arguments
796 for (const Value *V : ISP.gc_transition_args()) {
797 TEOps.push_back(getValue(V));
798 if (V->getType()->isPointerTy())
799 TEOps.push_back(DAG.getSrcValue(V));
803 TEOps.push_back(SDValue(StatepointMCNode, 1));
805 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
807 SDValue GCTransitionStart =
808 DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps);
810 SinkNode = GCTransitionStart.getNode();
813 // Replace original call
814 DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root
815 // Remove original call node
816 DAG.DeleteNode(CallNode);
818 // DON'T set the root - under the assumption that it's already set past the
819 // inserted node we created.
821 // TODO: A better future implementation would be to emit a single variable
822 // argument, variable return value STATEPOINT node here and then hookup the
823 // return value of each gc.relocate to the respective output of the
824 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
825 // to actually be possible today.
828 void SelectionDAGBuilder::visitGCResult(const CallInst &CI) {
829 // The result value of the gc_result is simply the result of the actual
830 // call. We've already emitted this, so just grab the value.
831 Instruction *I = cast<Instruction>(CI.getArgOperand(0));
832 assert(isStatepoint(I) && "first argument must be a statepoint token");
834 if (I->getParent() != CI.getParent()) {
835 // Statepoint is in different basic block so we should have stored call
836 // result in a virtual register.
837 // We can not use default getValue() functionality to copy value from this
838 // register because statepoint and actuall call return types can be
839 // different, and getValue() will use CopyFromReg of the wrong type,
840 // which is always i32 in our case.
841 PointerType *CalleeType = cast<PointerType>(
842 ImmutableStatepoint(I).getCalledValue()->getType());
844 cast<FunctionType>(CalleeType->getElementType())->getReturnType();
845 SDValue CopyFromReg = getCopyFromRegs(I, RetTy);
847 assert(CopyFromReg.getNode());
848 setValue(&CI, CopyFromReg);
850 setValue(&CI, getValue(I));
854 void SelectionDAGBuilder::visitGCRelocate(const CallInst &CI) {
855 GCRelocateOperands RelocateOpers(&CI);
859 // We skip this check for relocates not in the same basic block as thier
860 // statepoint. It would be too expensive to preserve validation info through
861 // different basic blocks.
862 if (RelocateOpers.getStatepoint()->getParent() == CI.getParent()) {
863 StatepointLowering.relocCallVisited(CI);
867 const Value *DerivedPtr = RelocateOpers.getDerivedPtr();
868 SDValue SD = getValue(DerivedPtr);
870 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
871 FuncInfo.StatepointRelocatedValues[RelocateOpers.getStatepoint()];
873 // We should have recorded location for this pointer
874 assert(SpillMap.count(DerivedPtr) && "Relocating not lowered gc value");
875 Optional<int> DerivedPtrLocation = SpillMap[DerivedPtr];
877 // We didn't need to spill these special cases (constants and allocas).
878 // See the handling in spillIncomingValueForStatepoint for detail.
879 if (!DerivedPtrLocation) {
884 SDValue SpillSlot = DAG.getTargetFrameIndex(*DerivedPtrLocation,
887 // Be conservative: flush all pending loads
888 // TODO: Probably we can be less restrictive on this,
889 // it may allow more scheduling opportunities.
890 SDValue Chain = getRoot();
893 DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, SpillSlot,
894 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
895 *DerivedPtrLocation),
896 false, false, false, 0);
898 // Again, be conservative, don't emit pending loads
899 DAG.setRoot(SpillLoad.getValue(1));
901 assert(SpillLoad.getNode());
902 setValue(&CI, SpillLoad);