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 suceeding path involes a bit of complexity that caused a
109 // 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 /// Try to find existing copies of the incoming values in stack slots used for
117 /// statepoint spilling. If we can find a spill slot for the incoming value,
118 /// mark that slot as allocated, and reuse the same slot for this safepoint.
119 /// This helps to avoid series of loads and stores that only serve to resuffle
120 /// values on the stack between calls.
121 static void reservePreviousStackSlotForValue(SDValue Incoming,
122 SelectionDAGBuilder &Builder) {
124 if (isa<ConstantSDNode>(Incoming) || isa<FrameIndexSDNode>(Incoming)) {
125 // We won't need to spill this, so no need to check for previously
126 // allocated stack slots
130 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
132 // duplicates in input
136 // Search back for the load from a stack slot pattern to find the original
137 // slot we allocated for this value. We could extend this to deal with
138 // simple modification patterns, but simple dealing with trivial load/store
139 // sequences helps a lot already.
140 if (LoadSDNode *Load = dyn_cast<LoadSDNode>(Incoming)) {
141 if (auto *FI = dyn_cast<FrameIndexSDNode>(Load->getBasePtr())) {
142 const int Index = FI->getIndex();
143 auto Itr = std::find(Builder.FuncInfo.StatepointStackSlots.begin(),
144 Builder.FuncInfo.StatepointStackSlots.end(), Index);
145 if (Itr == Builder.FuncInfo.StatepointStackSlots.end()) {
146 // not one of the lowering stack slots, can't reuse!
147 // TODO: Actually, we probably could reuse the stack slot if the value
148 // hasn't changed at all, but we'd need to look for intervening writes
151 // This is one of our dedicated lowering slots
153 std::distance(Builder.FuncInfo.StatepointStackSlots.begin(), Itr);
154 if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) {
155 // stack slot already assigned to someone else, can't use it!
156 // TODO: currently we reserve space for gc arguments after doing
157 // normal allocation for deopt arguments. We should reserve for
158 // _all_ deopt and gc arguments, then start allocating. This
159 // will prevent some moves being inserted when vm state changes,
160 // but gc state doesn't between two calls.
163 // Reserve this stack slot
164 Builder.StatepointLowering.reserveStackSlot(Offset);
167 // Cache this slot so we find it when going through the normal
170 Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
172 Builder.StatepointLowering.setLocation(Incoming, Loc);
176 // TODO: handle case where a reloaded value flows through a phi to
177 // another safepoint. e.g.
180 // bb2: % pred: bb1, bb3, bb4, etc.
181 // a_phi = phi(a', ...)
182 // statepoint ... a_phi
183 // NOTE: This will require reasoning about cross basic block values. This is
184 // decidedly non trivial and this might not be the right place to do it. We
185 // don't really have the information we need here...
187 // TODO: handle simple updates. If a value is modified and the original
188 // value is no longer live, it would be nice to put the modified value in the
189 // same slot. This allows folding of the memory accesses for some
190 // instructions types (like an increment).
196 /// Remove any duplicate (as SDValues) from the derived pointer pairs. This
197 /// is not required for correctness. It's purpose is to reduce the size of
198 /// StackMap section. It has no effect on the number of spill slots required
199 /// or the actual lowering.
200 static void removeDuplicatesGCPtrs(SmallVectorImpl<const Value *> &Bases,
201 SmallVectorImpl<const Value *> &Ptrs,
202 SmallVectorImpl<const Value *> &Relocs,
203 SelectionDAGBuilder &Builder) {
205 // This is horribly ineffecient, but I don't care right now
206 SmallSet<SDValue, 64> Seen;
208 SmallVector<const Value *, 64> NewBases, NewPtrs, NewRelocs;
209 for (size_t i = 0; i < Ptrs.size(); i++) {
210 SDValue SD = Builder.getValue(Ptrs[i]);
211 // Only add non-duplicates
212 if (Seen.count(SD) == 0) {
213 NewBases.push_back(Bases[i]);
214 NewPtrs.push_back(Ptrs[i]);
215 NewRelocs.push_back(Relocs[i]);
219 assert(Bases.size() >= NewBases.size());
220 assert(Ptrs.size() >= NewPtrs.size());
221 assert(Relocs.size() >= NewRelocs.size());
225 assert(Ptrs.size() == Bases.size());
226 assert(Ptrs.size() == Relocs.size());
229 /// Extract call from statepoint, lower it and return pointer to the
230 /// call node. Also update NodeMap so that getValue(statepoint) will
231 /// reference lowered call result
233 lowerCallFromStatepoint(ImmutableStatepoint ISP, MachineBasicBlock *LandingPad,
234 SelectionDAGBuilder &Builder,
235 SmallVectorImpl<SDValue> &PendingExports) {
237 ImmutableCallSite CS(ISP.getCallSite());
239 SDValue ActualCallee = Builder.getValue(ISP.getActualCallee());
241 assert(CS.getCallingConv() != CallingConv::AnyReg &&
242 "anyregcc is not supported on statepoints!");
244 Type *DefTy = ISP.getActualReturnType();
245 bool HasDef = !DefTy->isVoidTy();
247 SDValue ReturnValue, CallEndVal;
248 std::tie(ReturnValue, CallEndVal) = Builder.lowerCallOperands(
249 ISP.getCallSite(), ImmutableStatepoint::CallArgsBeginPos,
250 ISP.getNumCallArgs(), ActualCallee, DefTy, LandingPad,
251 false /* IsPatchPoint */);
253 SDNode *CallEnd = CallEndVal.getNode();
255 // Get a call instruction from the call sequence chain. Tail calls are not
256 // allowed. The following code is essentially reverse engineering X86's
259 // We are expecting DAG to have the following form:
261 // ch = eh_label (only in case of invoke statepoint)
262 // ch, glue = callseq_start ch
263 // ch, glue = X86::Call ch, glue
264 // ch, glue = callseq_end ch, glue
265 // get_return_value ch, glue
267 // get_return_value can either be a CopyFromReg to grab the return value from
268 // %RAX, or it can be a LOAD to load a value returned by reference via a stack
271 if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg ||
272 CallEnd->getOpcode() == ISD::LOAD))
273 CallEnd = CallEnd->getOperand(0).getNode();
275 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
279 // Result value will be used in different basic block for invokes
280 // so we need to export it now. But statepoint call has a different type
281 // than the actuall call. It means that standart exporting mechanism will
282 // create register of the wrong type. So instead we need to create
283 // register with correct type and save value into it manually.
284 // TODO: To eliminate this problem we can remove gc.result intrinsics
285 // completelly and make statepoint call to return a tuple.
286 unsigned Reg = Builder.FuncInfo.CreateRegs(ISP.getActualReturnType());
287 RegsForValue RFV(*Builder.DAG.getContext(),
288 Builder.DAG.getTargetLoweringInfo(), Reg,
289 ISP.getActualReturnType());
290 SDValue Chain = Builder.DAG.getEntryNode();
292 RFV.getCopyToRegs(ReturnValue, Builder.DAG, Builder.getCurSDLoc(), Chain,
294 PendingExports.push_back(Chain);
295 Builder.FuncInfo.ValueMap[CS.getInstruction()] = Reg;
297 // The value of the statepoint itself will be the value of call itself.
298 // We'll replace the actually call node shortly. gc_result will grab
300 Builder.setValue(CS.getInstruction(), ReturnValue);
303 // The token value is never used from here on, just generate a poison value
304 Builder.setValue(CS.getInstruction(),
305 Builder.DAG.getIntPtrConstant(-1, Builder.getCurSDLoc()));
308 return CallEnd->getOperand(0).getNode();
311 /// Callect all gc pointers coming into statepoint intrinsic, clean them up,
312 /// and return two arrays:
313 /// Bases - base pointers incoming to this statepoint
314 /// Ptrs - derived pointers incoming to this statepoint
315 /// Relocs - the gc_relocate corresponding to each base/ptr pair
316 /// Elements of this arrays should be in one-to-one correspondence with each
317 /// other i.e Bases[i], Ptrs[i] are from the same gcrelocate call
318 static void getIncomingStatepointGCValues(
319 SmallVectorImpl<const Value *> &Bases, SmallVectorImpl<const Value *> &Ptrs,
320 SmallVectorImpl<const Value *> &Relocs, ImmutableStatepoint StatepointSite,
321 SelectionDAGBuilder &Builder) {
322 for (GCRelocateOperands relocateOpers :
323 StatepointSite.getRelocates(StatepointSite)) {
324 Relocs.push_back(relocateOpers.getUnderlyingCallSite().getInstruction());
325 Bases.push_back(relocateOpers.getBasePtr());
326 Ptrs.push_back(relocateOpers.getDerivedPtr());
329 // Remove any redundant llvm::Values which map to the same SDValue as another
330 // input. Also has the effect of removing duplicates in the original
331 // llvm::Value input list as well. This is a useful optimization for
332 // reducing the size of the StackMap section. It has no other impact.
333 removeDuplicatesGCPtrs(Bases, Ptrs, Relocs, Builder);
335 assert(Bases.size() == Ptrs.size() && Ptrs.size() == Relocs.size());
338 /// Spill a value incoming to the statepoint. It might be either part of
340 /// or gcstate. In both cases unconditionally spill it on the stack unless it
341 /// is a null constant. Return pair with first element being frame index
342 /// containing saved value and second element with outgoing chain from the
344 static std::pair<SDValue, SDValue>
345 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
346 SelectionDAGBuilder &Builder) {
347 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
349 // Emit new store if we didn't do it for this ptr before
350 if (!Loc.getNode()) {
351 Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
353 assert(isa<FrameIndexSDNode>(Loc));
354 int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
355 // We use TargetFrameIndex so that isel will not select it into LEA
356 Loc = Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
358 // TODO: We can create TokenFactor node instead of
359 // chaining stores one after another, this may allow
360 // a bit more optimal scheduling for them
361 Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
362 MachinePointerInfo::getFixedStack(Index),
365 Builder.StatepointLowering.setLocation(Incoming, Loc);
368 assert(Loc.getNode());
369 return std::make_pair(Loc, Chain);
372 /// Lower a single value incoming to a statepoint node. This value can be
373 /// either a deopt value or a gc value, the handling is the same. We special
374 /// case constants and allocas, then fall back to spilling if required.
375 static void lowerIncomingStatepointValue(SDValue Incoming,
376 SmallVectorImpl<SDValue> &Ops,
377 SelectionDAGBuilder &Builder) {
378 SDValue Chain = Builder.getRoot();
380 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
381 // If the original value was a constant, make sure it gets recorded as
382 // such in the stackmap. This is required so that the consumer can
383 // parse any internal format to the deopt state. It also handles null
384 // pointers and other constant pointers in GC states
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 std::pair<SDValue, SDValue> Res =
400 spillIncomingStatepointValue(Incoming, Chain, Builder);
401 Ops.push_back(Res.first);
405 Builder.DAG.setRoot(Chain);
408 /// Lower deopt state and gc pointer arguments of the statepoint. The actual
409 /// lowering is described in lowerIncomingStatepointValue. This function is
410 /// responsible for lowering everything in the right position and playing some
411 /// tricks to avoid redundant stack manipulation where possible. On
412 /// completion, 'Ops' will contain ready to use operands for machine code
413 /// statepoint. The chain nodes will have already been created and the DAG root
414 /// will be set to the last value spilled (if any were).
415 static void lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
416 ImmutableStatepoint StatepointSite,
417 SelectionDAGBuilder &Builder) {
419 // Lower the deopt and gc arguments for this statepoint. Layout will
420 // be: deopt argument length, deopt arguments.., gc arguments...
422 SmallVector<const Value *, 64> Bases, Ptrs, Relocations;
423 getIncomingStatepointGCValues(Bases, Ptrs, Relocations, StatepointSite,
427 // Check that each of the gc pointer and bases we've gotten out of the
428 // safepoint is something the strategy thinks might be a pointer into the GC
429 // heap. This is basically just here to help catch errors during statepoint
430 // insertion. TODO: This should actually be in the Verifier, but we can't get
431 // to the GCStrategy from there (yet).
432 GCStrategy &S = Builder.GFI->getStrategy();
433 for (const Value *V : Bases) {
434 auto Opt = S.isGCManagedPointer(V);
435 if (Opt.hasValue()) {
436 assert(Opt.getValue() &&
437 "non gc managed base pointer found in statepoint");
440 for (const Value *V : Ptrs) {
441 auto Opt = S.isGCManagedPointer(V);
442 if (Opt.hasValue()) {
443 assert(Opt.getValue() &&
444 "non gc managed derived pointer found in statepoint");
447 for (const Value *V : Relocations) {
448 auto Opt = S.isGCManagedPointer(V);
449 if (Opt.hasValue()) {
450 assert(Opt.getValue() && "non gc managed pointer relocated");
455 // Before we actually start lowering (and allocating spill slots for values),
456 // reserve any stack slots which we judge to be profitable to reuse for a
457 // particular value. This is purely an optimization over the code below and
458 // doesn't change semantics at all. It is important for performance that we
459 // reserve slots for both deopt and gc values before lowering either.
460 for (const Value *V : StatepointSite.vm_state_args()) {
461 SDValue Incoming = Builder.getValue(V);
462 reservePreviousStackSlotForValue(Incoming, Builder);
464 for (unsigned i = 0; i < Bases.size(); ++i) {
465 const Value *Base = Bases[i];
466 reservePreviousStackSlotForValue(Builder.getValue(Base), Builder);
468 const Value *Ptr = Ptrs[i];
469 reservePreviousStackSlotForValue(Builder.getValue(Ptr), Builder);
472 // First, prefix the list with the number of unique values to be
473 // lowered. Note that this is the number of *Values* not the
474 // number of SDValues required to lower them.
475 const int NumVMSArgs = StatepointSite.getNumTotalVMSArgs();
476 pushStackMapConstant(Ops, Builder, NumVMSArgs);
478 assert(NumVMSArgs == std::distance(StatepointSite.vm_state_begin(),
479 StatepointSite.vm_state_end()));
481 // The vm state arguments are lowered in an opaque manner. We do
482 // not know what type of values are contained within. We skip the
483 // first one since that happens to be the total number we lowered
484 // explicitly just above. We could have left it in the loop and
485 // not done it explicitly, but it's far easier to understand this
487 for (const Value *V : StatepointSite.vm_state_args()) {
488 SDValue Incoming = Builder.getValue(V);
489 lowerIncomingStatepointValue(Incoming, Ops, Builder);
492 // Finally, go ahead and lower all the gc arguments. There's no prefixed
493 // length for this one. After lowering, we'll have the base and pointer
494 // arrays interwoven with each (lowered) base pointer immediately followed by
495 // it's (lowered) derived pointer. i.e
496 // (base[0], ptr[0], base[1], ptr[1], ...)
497 for (unsigned i = 0; i < Bases.size(); ++i) {
498 const Value *Base = Bases[i];
499 lowerIncomingStatepointValue(Builder.getValue(Base), Ops, Builder);
501 const Value *Ptr = Ptrs[i];
502 lowerIncomingStatepointValue(Builder.getValue(Ptr), Ops, Builder);
505 // If there are any explicit spill slots passed to the statepoint, record
506 // them, but otherwise do not do anything special. These are user provided
507 // allocas and give control over placement to the consumer. In this case,
508 // it is the contents of the slot which may get updated, not the pointer to
510 for (Value *V : StatepointSite.gc_args()) {
511 SDValue Incoming = Builder.getValue(V);
512 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
513 // This handles allocas as arguments to the statepoint
514 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
515 Incoming.getValueType()));
519 // Record computed locations for all lowered values.
520 // This can not be embedded in lowering loops as we need to record *all*
521 // values, while previous loops account only values with unique SDValues.
522 const Instruction *StatepointInstr =
523 StatepointSite.getCallSite().getInstruction();
524 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
525 Builder.FuncInfo.StatepointRelocatedValues[StatepointInstr];
527 for (GCRelocateOperands RelocateOpers :
528 StatepointSite.getRelocates(StatepointSite)) {
529 const Value *V = RelocateOpers.getDerivedPtr();
530 SDValue SDV = Builder.getValue(V);
531 SDValue Loc = Builder.StatepointLowering.getLocation(SDV);
534 SpillMap[V] = cast<FrameIndexSDNode>(Loc)->getIndex();
536 // Record value as visited, but not spilled. This is case for allocas
537 // and constants. For this values we can avoid emiting spill load while
538 // visiting corresponding gc_relocate.
539 // Actually we do not need to record them in this map at all.
540 // We do this only to check that we are not relocating any unvisited value.
543 // Default llvm mechanisms for exporting values which are used in
544 // different basic blocks does not work for gc relocates.
545 // Note that it would be incorrect to teach llvm that all relocates are
546 // uses of the corresponging values so that it would automatically
547 // export them. Relocates of the spilled values does not use original
549 if (StatepointSite.getCallSite().isInvoke())
550 Builder.ExportFromCurrentBlock(V);
555 void SelectionDAGBuilder::visitStatepoint(const CallInst &CI) {
556 // Check some preconditions for sanity
557 assert(isStatepoint(&CI) &&
558 "function called must be the statepoint function");
560 LowerStatepoint(ImmutableStatepoint(&CI));
563 void SelectionDAGBuilder::LowerStatepoint(
564 ImmutableStatepoint ISP, MachineBasicBlock *LandingPad /*=nullptr*/) {
565 // The basic scheme here is that information about both the original call and
566 // the safepoint is encoded in the CallInst. We create a temporary call and
567 // lower it, then reverse engineer the calling sequence.
571 StatepointLowering.startNewStatepoint(*this);
573 ImmutableCallSite CS(ISP.getCallSite());
576 // Consistency check. Don't do this for invokes. It would be too
577 // expensive to preserve this information across different basic blocks
578 if (!CS.isInvoke()) {
579 for (const User *U : CS->users()) {
580 const CallInst *Call = cast<CallInst>(U);
581 if (isGCRelocate(Call))
582 StatepointLowering.scheduleRelocCall(*Call);
588 // If this is a malformed statepoint, report it early to simplify debugging.
589 // This should catch any IR level mistake that's made when constructing or
590 // transforming statepoints.
593 // Check that the associated GCStrategy expects to encounter statepoints.
594 assert(GFI->getStrategy().useStatepoints() &&
595 "GCStrategy does not expect to encounter statepoints");
598 // Lower statepoint vmstate and gcstate arguments
599 SmallVector<SDValue, 10> LoweredMetaArgs;
600 lowerStatepointMetaArgs(LoweredMetaArgs, ISP, *this);
602 // Get call node, we will replace it later with statepoint
604 lowerCallFromStatepoint(ISP, LandingPad, *this, PendingExports);
606 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
607 // nodes with all the appropriate arguments and return values.
609 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
610 SDValue Chain = CallNode->getOperand(0);
613 bool CallHasIncomingGlue = CallNode->getGluedNode();
614 if (CallHasIncomingGlue) {
615 // Glue is always last operand
616 Glue = CallNode->getOperand(CallNode->getNumOperands() - 1);
619 // Build the GC_TRANSITION_START node if necessary.
621 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
622 // order in which they appear in the call to the statepoint intrinsic. If
623 // any of the operands is a pointer-typed, that operand is immediately
624 // followed by a SRCVALUE for the pointer that may be used during lowering
625 // (e.g. to form MachinePointerInfo values for loads/stores).
626 const bool IsGCTransition =
627 (ISP.getFlags() & (uint64_t)StatepointFlags::GCTransition) ==
628 (uint64_t)StatepointFlags::GCTransition;
629 if (IsGCTransition) {
630 SmallVector<SDValue, 8> TSOps;
633 TSOps.push_back(Chain);
635 // Add GC transition arguments
636 for (const Value *V : ISP.gc_transition_args()) {
637 TSOps.push_back(getValue(V));
638 if (V->getType()->isPointerTy())
639 TSOps.push_back(DAG.getSrcValue(V));
642 // Add glue if necessary
643 if (CallHasIncomingGlue)
644 TSOps.push_back(Glue);
646 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
648 SDValue GCTransitionStart =
649 DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps);
651 Chain = GCTransitionStart.getValue(0);
652 Glue = GCTransitionStart.getValue(1);
655 // TODO: Currently, all of these operands are being marked as read/write in
656 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
657 // and flags to be read-only.
658 SmallVector<SDValue, 40> Ops;
660 // Add the <id> and <numBytes> constants.
661 Ops.push_back(DAG.getTargetConstant(ISP.getID(), getCurSDLoc(), MVT::i64));
663 DAG.getTargetConstant(ISP.getNumPatchBytes(), getCurSDLoc(), MVT::i32));
665 // Calculate and push starting position of vmstate arguments
666 // Get number of arguments incoming directly into call node
667 unsigned NumCallRegArgs =
668 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
669 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32));
672 SDValue CallTarget = SDValue(CallNode->getOperand(1).getNode(), 0);
673 Ops.push_back(CallTarget);
675 // Add call arguments
676 // Get position of register mask in the call
677 SDNode::op_iterator RegMaskIt;
678 if (CallHasIncomingGlue)
679 RegMaskIt = CallNode->op_end() - 2;
681 RegMaskIt = CallNode->op_end() - 1;
682 Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt);
684 // Add a constant argument for the calling convention
685 pushStackMapConstant(Ops, *this, CS.getCallingConv());
687 // Add a constant argument for the flags
688 uint64_t Flags = ISP.getFlags();
690 ((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0)
691 && "unknown flag used");
692 pushStackMapConstant(Ops, *this, Flags);
694 // Insert all vmstate and gcstate arguments
695 Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end());
697 // Add register mask from call node
698 Ops.push_back(*RegMaskIt);
701 Ops.push_back(Chain);
703 // Same for the glue, but we add it only if original call had it
707 // Compute return values. Provide a glue output since we consume one as
708 // input. This allows someone else to chain off us as needed.
709 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
711 SDNode *StatepointMCNode =
712 DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
714 SDNode *SinkNode = StatepointMCNode;
716 // Build the GC_TRANSITION_END node if necessary.
718 // See the comment above regarding GC_TRANSITION_START for the layout of
719 // the operands to the GC_TRANSITION_END node.
720 if (IsGCTransition) {
721 SmallVector<SDValue, 8> TEOps;
724 TEOps.push_back(SDValue(StatepointMCNode, 0));
726 // Add GC transition arguments
727 for (const Value *V : ISP.gc_transition_args()) {
728 TEOps.push_back(getValue(V));
729 if (V->getType()->isPointerTy())
730 TEOps.push_back(DAG.getSrcValue(V));
734 TEOps.push_back(SDValue(StatepointMCNode, 1));
736 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
738 SDValue GCTransitionStart =
739 DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps);
741 SinkNode = GCTransitionStart.getNode();
744 // Replace original call
745 DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root
746 // Remove originall call node
747 DAG.DeleteNode(CallNode);
749 // DON'T set the root - under the assumption that it's already set past the
750 // inserted node we created.
752 // TODO: A better future implementation would be to emit a single variable
753 // argument, variable return value STATEPOINT node here and then hookup the
754 // return value of each gc.relocate to the respective output of the
755 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
756 // to actually be possible today.
759 void SelectionDAGBuilder::visitGCResult(const CallInst &CI) {
760 // The result value of the gc_result is simply the result of the actual
761 // call. We've already emitted this, so just grab the value.
762 Instruction *I = cast<Instruction>(CI.getArgOperand(0));
763 assert(isStatepoint(I) && "first argument must be a statepoint token");
765 if (isa<InvokeInst>(I)) {
766 // For invokes we should have stored call result in a virtual register.
767 // We can not use default getValue() functionality to copy value from this
768 // register because statepoint and actuall call return types can be
769 // different, and getValue() will use CopyFromReg of the wrong type,
770 // which is always i32 in our case.
771 PointerType *CalleeType =
772 cast<PointerType>(ImmutableStatepoint(I).getActualCallee()->getType());
774 cast<FunctionType>(CalleeType->getElementType())->getReturnType();
775 SDValue CopyFromReg = getCopyFromRegs(I, RetTy);
777 assert(CopyFromReg.getNode());
778 setValue(&CI, CopyFromReg);
780 setValue(&CI, getValue(I));
784 void SelectionDAGBuilder::visitGCRelocate(const CallInst &CI) {
785 GCRelocateOperands RelocateOpers(&CI);
789 // We skip this check for invoke statepoints. It would be too expensive to
790 // preserve validation info through different basic blocks.
791 if (!RelocateOpers.isTiedToInvoke()) {
792 StatepointLowering.relocCallVisited(CI);
796 const Value *DerivedPtr = RelocateOpers.getDerivedPtr();
797 SDValue SD = getValue(DerivedPtr);
799 FunctionLoweringInfo::StatepointSpilledValueMapTy &SpillMap =
800 FuncInfo.StatepointRelocatedValues[RelocateOpers.getStatepoint()];
802 // We should have recorded location for this pointer
803 assert(SpillMap.count(DerivedPtr) && "Relocating not lowered gc value");
804 Optional<int> DerivedPtrLocation = SpillMap[DerivedPtr];
806 // We didn't need to spill these special cases (constants and allocas).
807 // See the handling in spillIncomingValueForStatepoint for detail.
808 if (!DerivedPtrLocation) {
813 SDValue SpillSlot = DAG.getTargetFrameIndex(*DerivedPtrLocation,
816 // Be conservative: flush all pending loads
817 // TODO: Probably we can be less restrictive on this,
818 // it may allow more scheduling opprtunities
819 SDValue Chain = getRoot();
822 DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, SpillSlot,
823 MachinePointerInfo::getFixedStack(*DerivedPtrLocation),
824 false, false, false, 0);
826 // Again, be conservative, don't emit pending loads
827 DAG.setRoot(SpillLoad.getValue(1));
829 assert(SpillLoad.getNode());
830 setValue(&CI, SpillLoad);