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. Also need to ensure used bits get cleared.
59 AllocatedStackSlots.clear();
60 AllocatedStackSlots.resize(Builder.FuncInfo.StatepointStackSlots.size());
63 void StatepointLoweringState::clear() {
65 AllocatedStackSlots.clear();
66 assert(PendingGCRelocateCalls.empty() &&
67 "cleared before statepoint sequence completed");
71 StatepointLoweringState::allocateStackSlot(EVT ValueType,
72 SelectionDAGBuilder &Builder) {
73 NumSlotsAllocatedForStatepoints++;
74 MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
76 unsigned SpillSize = ValueType.getSizeInBits() / 8;
77 assert((SpillSize * 8) == ValueType.getSizeInBits() && "Size not in bytes?");
79 // First look for a previously created stack slot which is not in
80 // use (accounting for the fact arbitrary slots may already be
81 // reserved), or to create a new stack slot and use it.
83 const size_t NumSlots = AllocatedStackSlots.size();
84 assert(NextSlotToAllocate <= NumSlots && "Broken invariant");
86 assert(AllocatedStackSlots.size() ==
87 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 // TODO: Is ValueType the right thing to use here?
96 return Builder.DAG.getFrameIndex(FI, ValueType);
101 // Couldn't find a free slot, so create a new one:
103 SDValue SpillSlot = Builder.DAG.CreateStackTemporary(ValueType);
104 const unsigned FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
105 MFI.markAsStatepointSpillSlotObjectIndex(FI);
107 Builder.FuncInfo.StatepointStackSlots.push_back(FI);
108 AllocatedStackSlots.resize(AllocatedStackSlots.size()+1, true);
109 assert(AllocatedStackSlots.size() ==
110 Builder.FuncInfo.StatepointStackSlots.size() &&
113 StatepointMaxSlotsRequired = std::max<unsigned long>(
114 StatepointMaxSlotsRequired, Builder.FuncInfo.StatepointStackSlots.size());
119 /// Utility function for reservePreviousStackSlotForValue. Tries to find
120 /// stack slot index to which we have spilled value for previous statepoints.
121 /// LookUpDepth specifies maximum DFS depth this function is allowed to look.
122 static Optional<int> findPreviousSpillSlot(const Value *Val,
123 SelectionDAGBuilder &Builder,
125 // Can not look any further - give up now
126 if (LookUpDepth <= 0)
129 // Spill location is known for gc relocates
130 if (const auto *Relocate = dyn_cast<GCRelocateInst>(Val)) {
131 const auto &SpillMap =
132 Builder.FuncInfo.StatepointSpillMaps[Relocate->getStatepoint()];
134 auto It = SpillMap.find(Relocate->getDerivedPtr());
135 if (It == SpillMap.end())
141 // Look through bitcast instructions.
142 if (const BitCastInst *Cast = dyn_cast<BitCastInst>(Val))
143 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())
157 if (MergedResult.hasValue() && *MergedResult != *SpillSlot)
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
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 const auto &StatepointSlots = Builder.FuncInfo.StatepointStackSlots;
225 auto SlotIt = find(StatepointSlots, *Index);
226 assert(SlotIt != StatepointSlots.end() &&
227 "Value spilled to the unknown stack slot");
229 // This is one of our dedicated lowering slots
230 const int Offset = std::distance(StatepointSlots.begin(), SlotIt);
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.
254 removeDuplicateGCPtrs(SmallVectorImpl<const Value *> &Bases,
255 SmallVectorImpl<const Value *> &Ptrs,
256 SmallVectorImpl<const GCRelocateInst *> &Relocs,
257 SelectionDAGBuilder &Builder,
258 FunctionLoweringInfo::StatepointSpillMap &SSM) {
259 DenseMap<SDValue, const Value *> Seen;
261 SmallVector<const Value *, 64> NewBases, NewPtrs;
262 SmallVector<const GCRelocateInst *, 64> NewRelocs;
263 for (size_t i = 0, e = Ptrs.size(); i < e; i++) {
264 SDValue SD = Builder.getValue(Ptrs[i]);
265 auto SeenIt = Seen.find(SD);
267 if (SeenIt == Seen.end()) {
268 // Only add non-duplicates
269 NewBases.push_back(Bases[i]);
270 NewPtrs.push_back(Ptrs[i]);
271 NewRelocs.push_back(Relocs[i]);
274 // Duplicate pointer found, note in SSM and move on:
275 SSM.DuplicateMap[Ptrs[i]] = SeenIt->second;
278 assert(Bases.size() >= NewBases.size());
279 assert(Ptrs.size() >= NewPtrs.size());
280 assert(Relocs.size() >= NewRelocs.size());
284 assert(Ptrs.size() == Bases.size());
285 assert(Ptrs.size() == Relocs.size());
288 /// Extract call from statepoint, lower it and return pointer to the
289 /// call node. Also update NodeMap so that getValue(statepoint) will
290 /// reference lowered call result
291 static std::pair<SDValue, SDNode *> lowerCallFromStatepointLoweringInfo(
292 SelectionDAGBuilder::StatepointLoweringInfo &SI,
293 SelectionDAGBuilder &Builder, SmallVectorImpl<SDValue> &PendingExports) {
295 SDValue ReturnValue, CallEndVal;
296 std::tie(ReturnValue, CallEndVal) =
297 Builder.lowerInvokable(SI.CLI, SI.EHPadBB);
298 SDNode *CallEnd = CallEndVal.getNode();
300 // Get a call instruction from the call sequence chain. Tail calls are not
301 // allowed. The following code is essentially reverse engineering X86's
304 // We are expecting DAG to have the following form:
306 // ch = eh_label (only in case of invoke statepoint)
307 // ch, glue = callseq_start ch
308 // ch, glue = X86::Call ch, glue
309 // ch, glue = callseq_end ch, glue
310 // get_return_value ch, glue
312 // get_return_value can either be a sequence of CopyFromReg instructions
313 // to grab the return value from the return register(s), or it can be a LOAD
314 // to load a value returned by reference via a stack slot.
316 bool HasDef = !SI.CLI.RetTy->isVoidTy();
318 if (CallEnd->getOpcode() == ISD::LOAD)
319 CallEnd = CallEnd->getOperand(0).getNode();
321 while (CallEnd->getOpcode() == ISD::CopyFromReg)
322 CallEnd = CallEnd->getOperand(0).getNode();
325 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
326 return std::make_pair(ReturnValue, CallEnd->getOperand(0).getNode());
329 /// Spill a value incoming to the statepoint. It might be either part of
331 /// or gcstate. In both cases unconditionally spill it on the stack unless it
332 /// is a null constant. Return pair with first element being frame index
333 /// containing saved value and second element with outgoing chain from the
335 static std::pair<SDValue, SDValue>
336 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
337 SelectionDAGBuilder &Builder) {
338 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
340 // Emit new store if we didn't do it for this ptr before
341 if (!Loc.getNode()) {
342 Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
344 int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
345 // We use TargetFrameIndex so that isel will not select it into LEA
346 Loc = Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
348 // TODO: We can create TokenFactor node instead of
349 // chaining stores one after another, this may allow
350 // a bit more optimal scheduling for them
353 // Right now we always allocate spill slots that are of the same
354 // size as the value we're about to spill (the size of spillee can
355 // vary since we spill vectors of pointers too). At some point we
356 // can consider allowing spills of smaller values to larger slots
357 // (i.e. change the '==' in the assert below to a '>=').
358 MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
359 assert((MFI.getObjectSize(Index) * 8) == Incoming.getValueSizeInBits() &&
360 "Bad spill: stack slot does not match!");
363 Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
364 MachinePointerInfo::getFixedStack(
365 Builder.DAG.getMachineFunction(), Index));
367 Builder.StatepointLowering.setLocation(Incoming, Loc);
370 assert(Loc.getNode());
371 return std::make_pair(Loc, Chain);
374 /// Lower a single value incoming to a statepoint node. This value can be
375 /// either a deopt value or a gc value, the handling is the same. We special
376 /// case constants and allocas, then fall back to spilling if required.
377 static void lowerIncomingStatepointValue(SDValue Incoming, bool LiveInOnly,
378 SmallVectorImpl<SDValue> &Ops,
379 SelectionDAGBuilder &Builder) {
380 SDValue Chain = Builder.getRoot();
382 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
383 // If the original value was a constant, make sure it gets recorded as
384 // such in the stackmap. This is required so that the consumer can
385 // parse any internal format to the deopt state. It also handles null
386 // pointers and other constant pointers in GC states. Note the constant
387 // vectors do not appear to actually hit this path and that anything larger
388 // than an i64 value (not type!) will fail asserts here.
389 pushStackMapConstant(Ops, Builder, C->getSExtValue());
390 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
391 // This handles allocas as arguments to the statepoint (this is only
392 // really meaningful for a deopt value. For GC, we'd be trying to
393 // relocate the address of the alloca itself?)
394 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
395 Incoming.getValueType()));
396 } else if (LiveInOnly) {
397 // If this value is live in (not live-on-return, or live-through), we can
398 // treat it the same way patchpoint treats it's "live in" values. We'll
399 // end up folding some of these into stack references, but they'll be
400 // handled by the register allocator. Note that we do not have the notion
401 // of a late use so these values might be placed in registers which are
402 // clobbered by the call. This is fine for live-in.
403 Ops.push_back(Incoming);
405 // Otherwise, locate a spill slot and explicitly spill it so it
406 // can be found by the runtime later. We currently do not support
407 // tracking values through callee saved registers to their eventual
408 // spill location. This would be a useful optimization, but would
409 // need to be optional since it requires a lot of complexity on the
410 // runtime side which not all would support.
411 auto Res = spillIncomingStatepointValue(Incoming, Chain, Builder);
412 Ops.push_back(Res.first);
416 Builder.DAG.setRoot(Chain);
419 /// Lower deopt state and gc pointer arguments of the statepoint. The actual
420 /// lowering is described in lowerIncomingStatepointValue. This function is
421 /// responsible for lowering everything in the right position and playing some
422 /// tricks to avoid redundant stack manipulation where possible. On
423 /// completion, 'Ops' will contain ready to use operands for machine code
424 /// statepoint. The chain nodes will have already been created and the DAG root
425 /// will be set to the last value spilled (if any were).
427 lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
428 SelectionDAGBuilder::StatepointLoweringInfo &SI,
429 SelectionDAGBuilder &Builder) {
430 // Lower the deopt and gc arguments for this statepoint. Layout will be:
431 // deopt argument length, deopt arguments.., gc arguments...
433 if (auto *GFI = Builder.GFI) {
434 // Check that each of the gc pointer and bases we've gotten out of the
435 // safepoint is something the strategy thinks might be a pointer (or vector
436 // of pointers) into the GC heap. This is basically just here to help catch
437 // errors during statepoint insertion. TODO: This should actually be in the
438 // Verifier, but we can't get to the GCStrategy from there (yet).
439 GCStrategy &S = GFI->getStrategy();
440 for (const Value *V : SI.Bases) {
441 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
442 if (Opt.hasValue()) {
443 assert(Opt.getValue() &&
444 "non gc managed base pointer found in statepoint");
447 for (const Value *V : SI.Ptrs) {
448 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
449 if (Opt.hasValue()) {
450 assert(Opt.getValue() &&
451 "non gc managed derived pointer found in statepoint");
454 assert(SI.Bases.size() == SI.Ptrs.size() && "Pointer without base!");
456 assert(SI.Bases.empty() && "No gc specified, so cannot relocate pointers!");
457 assert(SI.Ptrs.empty() && "No gc specified, so cannot relocate pointers!");
461 // Figure out what lowering strategy we're going to use for each part
462 // Note: Is is conservatively correct to lower both "live-in" and "live-out"
463 // as "live-through". A "live-through" variable is one which is "live-in",
464 // "live-out", and live throughout the lifetime of the call (i.e. we can find
465 // it from any PC within the transitive callee of the statepoint). In
466 // particular, if the callee spills callee preserved registers we may not
467 // be able to find a value placed in that register during the call. This is
468 // fine for live-out, but not for live-through. If we were willing to make
469 // assumptions about the code generator producing the callee, we could
470 // potentially allow live-through values in callee saved registers.
471 const bool LiveInDeopt =
472 SI.StatepointFlags & (uint64_t)StatepointFlags::DeoptLiveIn;
474 auto isGCValue =[&](const Value *V) {
475 return is_contained(SI.Ptrs, V) || is_contained(SI.Bases, V);
478 // Before we actually start lowering (and allocating spill slots for values),
479 // reserve any stack slots which we judge to be profitable to reuse for a
480 // particular value. This is purely an optimization over the code below and
481 // doesn't change semantics at all. It is important for performance that we
482 // reserve slots for both deopt and gc values before lowering either.
483 for (const Value *V : SI.DeoptState) {
484 if (!LiveInDeopt || isGCValue(V))
485 reservePreviousStackSlotForValue(V, Builder);
487 for (unsigned i = 0; i < SI.Bases.size(); ++i) {
488 reservePreviousStackSlotForValue(SI.Bases[i], Builder);
489 reservePreviousStackSlotForValue(SI.Ptrs[i], Builder);
492 // First, prefix the list with the number of unique values to be
493 // lowered. Note that this is the number of *Values* not the
494 // number of SDValues required to lower them.
495 const int NumVMSArgs = SI.DeoptState.size();
496 pushStackMapConstant(Ops, Builder, NumVMSArgs);
498 // The vm state arguments are lowered in an opaque manner. We do not know
499 // what type of values are contained within.
500 for (const Value *V : SI.DeoptState) {
501 SDValue Incoming = Builder.getValue(V);
502 const bool LiveInValue = LiveInDeopt && !isGCValue(V);
503 lowerIncomingStatepointValue(Incoming, LiveInValue, Ops, Builder);
506 // Finally, go ahead and lower all the gc arguments. There's no prefixed
507 // length for this one. After lowering, we'll have the base and pointer
508 // arrays interwoven with each (lowered) base pointer immediately followed by
509 // it's (lowered) derived pointer. i.e
510 // (base[0], ptr[0], base[1], ptr[1], ...)
511 for (unsigned i = 0; i < SI.Bases.size(); ++i) {
512 const Value *Base = SI.Bases[i];
513 lowerIncomingStatepointValue(Builder.getValue(Base), /*LiveInOnly*/ false,
516 const Value *Ptr = SI.Ptrs[i];
517 lowerIncomingStatepointValue(Builder.getValue(Ptr), /*LiveInOnly*/ false,
521 // If there are any explicit spill slots passed to the statepoint, record
522 // them, but otherwise do not do anything special. These are user provided
523 // allocas and give control over placement to the consumer. In this case,
524 // it is the contents of the slot which may get updated, not the pointer to
526 for (Value *V : SI.GCArgs) {
527 SDValue Incoming = Builder.getValue(V);
528 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
529 // This handles allocas as arguments to the statepoint
530 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
531 Incoming.getValueType()));
535 // Record computed locations for all lowered values.
536 // This can not be embedded in lowering loops as we need to record *all*
537 // values, while previous loops account only values with unique SDValues.
538 const Instruction *StatepointInstr = SI.StatepointInstr;
539 auto &SpillMap = Builder.FuncInfo.StatepointSpillMaps[StatepointInstr];
541 for (const GCRelocateInst *Relocate : SI.GCRelocates) {
542 const Value *V = Relocate->getDerivedPtr();
543 SDValue SDV = Builder.getValue(V);
544 SDValue Loc = Builder.StatepointLowering.getLocation(SDV);
547 SpillMap.SlotMap[V] = cast<FrameIndexSDNode>(Loc)->getIndex();
549 // Record value as visited, but not spilled. This is case for allocas
550 // and constants. For this values we can avoid emitting spill load while
551 // visiting corresponding gc_relocate.
552 // Actually we do not need to record them in this map at all.
553 // We do this only to check that we are not relocating any unvisited
555 SpillMap.SlotMap[V] = None;
557 // Default llvm mechanisms for exporting values which are used in
558 // different basic blocks does not work for gc relocates.
559 // Note that it would be incorrect to teach llvm that all relocates are
560 // uses of the corresponding values so that it would automatically
561 // export them. Relocates of the spilled values does not use original
563 if (Relocate->getParent() != StatepointInstr->getParent())
564 Builder.ExportFromCurrentBlock(V);
569 SDValue SelectionDAGBuilder::LowerAsSTATEPOINT(
570 SelectionDAGBuilder::StatepointLoweringInfo &SI) {
571 // The basic scheme here is that information about both the original call and
572 // the safepoint is encoded in the CallInst. We create a temporary call and
573 // lower it, then reverse engineer the calling sequence.
577 StatepointLowering.startNewStatepoint(*this);
580 // We schedule gc relocates before removeDuplicateGCPtrs since we _will_
581 // encounter the duplicate gc relocates we elide in removeDuplicateGCPtrs.
582 for (auto *Reloc : SI.GCRelocates)
583 if (Reloc->getParent() == SI.StatepointInstr->getParent())
584 StatepointLowering.scheduleRelocCall(*Reloc);
587 // Remove any redundant llvm::Values which map to the same SDValue as another
588 // input. Also has the effect of removing duplicates in the original
589 // llvm::Value input list as well. This is a useful optimization for
590 // reducing the size of the StackMap section. It has no other impact.
591 removeDuplicateGCPtrs(SI.Bases, SI.Ptrs, SI.GCRelocates, *this,
592 FuncInfo.StatepointSpillMaps[SI.StatepointInstr]);
593 assert(SI.Bases.size() == SI.Ptrs.size() &&
594 SI.Ptrs.size() == SI.GCRelocates.size());
596 // Lower statepoint vmstate and gcstate arguments
597 SmallVector<SDValue, 10> LoweredMetaArgs;
598 lowerStatepointMetaArgs(LoweredMetaArgs, SI, *this);
600 // Now that we've emitted the spills, we need to update the root so that the
601 // call sequence is ordered correctly.
602 SI.CLI.setChain(getRoot());
604 // Get call node, we will replace it later with statepoint
607 std::tie(ReturnVal, CallNode) =
608 lowerCallFromStatepointLoweringInfo(SI, *this, PendingExports);
610 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
611 // nodes with all the appropriate arguments and return values.
613 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
614 SDValue Chain = CallNode->getOperand(0);
617 bool CallHasIncomingGlue = CallNode->getGluedNode();
618 if (CallHasIncomingGlue) {
619 // Glue is always last operand
620 Glue = CallNode->getOperand(CallNode->getNumOperands() - 1);
623 // Build the GC_TRANSITION_START node if necessary.
625 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
626 // order in which they appear in the call to the statepoint intrinsic. If
627 // any of the operands is a pointer-typed, that operand is immediately
628 // followed by a SRCVALUE for the pointer that may be used during lowering
629 // (e.g. to form MachinePointerInfo values for loads/stores).
630 const bool IsGCTransition =
631 (SI.StatepointFlags & (uint64_t)StatepointFlags::GCTransition) ==
632 (uint64_t)StatepointFlags::GCTransition;
633 if (IsGCTransition) {
634 SmallVector<SDValue, 8> TSOps;
637 TSOps.push_back(Chain);
639 // Add GC transition arguments
640 for (const Value *V : SI.GCTransitionArgs) {
641 TSOps.push_back(getValue(V));
642 if (V->getType()->isPointerTy())
643 TSOps.push_back(DAG.getSrcValue(V));
646 // Add glue if necessary
647 if (CallHasIncomingGlue)
648 TSOps.push_back(Glue);
650 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
652 SDValue GCTransitionStart =
653 DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps);
655 Chain = GCTransitionStart.getValue(0);
656 Glue = GCTransitionStart.getValue(1);
659 // TODO: Currently, all of these operands are being marked as read/write in
660 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
661 // and flags to be read-only.
662 SmallVector<SDValue, 40> Ops;
664 // Add the <id> and <numBytes> constants.
665 Ops.push_back(DAG.getTargetConstant(SI.ID, getCurSDLoc(), MVT::i64));
667 DAG.getTargetConstant(SI.NumPatchBytes, getCurSDLoc(), MVT::i32));
669 // Calculate and push starting position of vmstate arguments
670 // Get number of arguments incoming directly into call node
671 unsigned NumCallRegArgs =
672 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
673 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32));
676 SDValue CallTarget = SDValue(CallNode->getOperand(1).getNode(), 0);
677 Ops.push_back(CallTarget);
679 // Add call arguments
680 // Get position of register mask in the call
681 SDNode::op_iterator RegMaskIt;
682 if (CallHasIncomingGlue)
683 RegMaskIt = CallNode->op_end() - 2;
685 RegMaskIt = CallNode->op_end() - 1;
686 Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt);
688 // Add a constant argument for the calling convention
689 pushStackMapConstant(Ops, *this, SI.CLI.CallConv);
691 // Add a constant argument for the flags
692 uint64_t Flags = SI.StatepointFlags;
693 assert(((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0) &&
694 "Unknown flag used");
695 pushStackMapConstant(Ops, *this, Flags);
697 // Insert all vmstate and gcstate arguments
698 Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end());
700 // Add register mask from call node
701 Ops.push_back(*RegMaskIt);
704 Ops.push_back(Chain);
706 // Same for the glue, but we add it only if original call had it
710 // Compute return values. Provide a glue output since we consume one as
711 // input. This allows someone else to chain off us as needed.
712 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
714 SDNode *StatepointMCNode =
715 DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
717 SDNode *SinkNode = StatepointMCNode;
719 // Build the GC_TRANSITION_END node if necessary.
721 // See the comment above regarding GC_TRANSITION_START for the layout of
722 // the operands to the GC_TRANSITION_END node.
723 if (IsGCTransition) {
724 SmallVector<SDValue, 8> TEOps;
727 TEOps.push_back(SDValue(StatepointMCNode, 0));
729 // Add GC transition arguments
730 for (const Value *V : SI.GCTransitionArgs) {
731 TEOps.push_back(getValue(V));
732 if (V->getType()->isPointerTy())
733 TEOps.push_back(DAG.getSrcValue(V));
737 TEOps.push_back(SDValue(StatepointMCNode, 1));
739 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
741 SDValue GCTransitionStart =
742 DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps);
744 SinkNode = GCTransitionStart.getNode();
747 // Replace original call
748 DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root
749 // Remove original call node
750 DAG.DeleteNode(CallNode);
752 // DON'T set the root - under the assumption that it's already set past the
753 // inserted node we created.
755 // TODO: A better future implementation would be to emit a single variable
756 // argument, variable return value STATEPOINT node here and then hookup the
757 // return value of each gc.relocate to the respective output of the
758 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
759 // to actually be possible today.
765 SelectionDAGBuilder::LowerStatepoint(ImmutableStatepoint ISP,
766 const BasicBlock *EHPadBB /*= nullptr*/) {
767 assert(ISP.getCallSite().getCallingConv() != CallingConv::AnyReg &&
768 "anyregcc is not supported on statepoints!");
771 // If this is a malformed statepoint, report it early to simplify debugging.
772 // This should catch any IR level mistake that's made when constructing or
773 // transforming statepoints.
776 // Check that the associated GCStrategy expects to encounter statepoints.
777 assert(GFI->getStrategy().useStatepoints() &&
778 "GCStrategy does not expect to encounter statepoints");
781 SDValue ActualCallee;
783 if (ISP.getNumPatchBytes() > 0) {
784 // If we've been asked to emit a nop sequence instead of a call instruction
785 // for this statepoint then don't lower the call target, but use a constant
786 // `null` instead. Not lowering the call target lets statepoint clients get
787 // away without providing a physical address for the symbolic call target at
790 const auto &TLI = DAG.getTargetLoweringInfo();
791 const auto &DL = DAG.getDataLayout();
793 unsigned AS = ISP.getCalledValue()->getType()->getPointerAddressSpace();
794 ActualCallee = DAG.getConstant(0, getCurSDLoc(), TLI.getPointerTy(DL, AS));
796 ActualCallee = getValue(ISP.getCalledValue());
799 StatepointLoweringInfo SI(DAG);
800 populateCallLoweringInfo(SI.CLI, ISP.getCallSite(),
801 ImmutableStatepoint::CallArgsBeginPos,
802 ISP.getNumCallArgs(), ActualCallee,
803 ISP.getActualReturnType(), false /* IsPatchPoint */);
805 for (const GCRelocateInst *Relocate : ISP.getRelocates()) {
806 SI.GCRelocates.push_back(Relocate);
807 SI.Bases.push_back(Relocate->getBasePtr());
808 SI.Ptrs.push_back(Relocate->getDerivedPtr());
811 SI.GCArgs = ArrayRef<const Use>(ISP.gc_args_begin(), ISP.gc_args_end());
812 SI.StatepointInstr = ISP.getInstruction();
813 SI.GCTransitionArgs =
814 ArrayRef<const Use>(ISP.gc_args_begin(), ISP.gc_args_end());
816 SI.DeoptState = ArrayRef<const Use>(ISP.vm_state_begin(), ISP.vm_state_end());
817 SI.StatepointFlags = ISP.getFlags();
818 SI.NumPatchBytes = ISP.getNumPatchBytes();
819 SI.EHPadBB = EHPadBB;
821 SDValue ReturnValue = LowerAsSTATEPOINT(SI);
823 // Export the result value if needed
824 const GCResultInst *GCResult = ISP.getGCResult();
825 Type *RetTy = ISP.getActualReturnType();
826 if (!RetTy->isVoidTy() && GCResult) {
827 if (GCResult->getParent() != ISP.getCallSite().getParent()) {
828 // Result value will be used in a different basic block so we need to
829 // export it now. Default exporting mechanism will not work here because
830 // statepoint call has a different type than the actual call. It means
831 // that by default llvm will create export register of the wrong type
832 // (always i32 in our case). So instead we need to create export register
833 // with correct type manually.
834 // TODO: To eliminate this problem we can remove gc.result intrinsics
835 // completely and make statepoint call to return a tuple.
836 unsigned Reg = FuncInfo.CreateRegs(RetTy);
837 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
838 DAG.getDataLayout(), Reg, RetTy);
839 SDValue Chain = DAG.getEntryNode();
841 RFV.getCopyToRegs(ReturnValue, DAG, getCurSDLoc(), Chain, nullptr);
842 PendingExports.push_back(Chain);
843 FuncInfo.ValueMap[ISP.getInstruction()] = Reg;
845 // Result value will be used in a same basic block. Don't export it or
846 // perform any explicit register copies.
847 // We'll replace the actuall call node shortly. gc_result will grab
849 setValue(ISP.getInstruction(), ReturnValue);
852 // The token value is never used from here on, just generate a poison value
853 setValue(ISP.getInstruction(), DAG.getIntPtrConstant(-1, getCurSDLoc()));
857 void SelectionDAGBuilder::LowerCallSiteWithDeoptBundleImpl(
858 ImmutableCallSite CS, SDValue Callee, const BasicBlock *EHPadBB,
859 bool VarArgDisallowed, bool ForceVoidReturnTy) {
860 StatepointLoweringInfo SI(DAG);
861 unsigned ArgBeginIndex = CS.arg_begin() - CS.getInstruction()->op_begin();
862 populateCallLoweringInfo(
863 SI.CLI, CS, ArgBeginIndex, CS.getNumArgOperands(), Callee,
864 ForceVoidReturnTy ? Type::getVoidTy(*DAG.getContext()) : CS.getType(),
866 if (!VarArgDisallowed)
867 SI.CLI.IsVarArg = CS.getFunctionType()->isVarArg();
869 auto DeoptBundle = *CS.getOperandBundle(LLVMContext::OB_deopt);
871 unsigned DefaultID = StatepointDirectives::DeoptBundleStatepointID;
873 auto SD = parseStatepointDirectivesFromAttrs(CS.getAttributes());
874 SI.ID = SD.StatepointID.getValueOr(DefaultID);
875 SI.NumPatchBytes = SD.NumPatchBytes.getValueOr(0);
878 ArrayRef<const Use>(DeoptBundle.Inputs.begin(), DeoptBundle.Inputs.end());
879 SI.StatepointFlags = static_cast<uint64_t>(StatepointFlags::None);
880 SI.EHPadBB = EHPadBB;
882 // NB! The GC arguments are deliberately left empty.
884 if (SDValue ReturnVal = LowerAsSTATEPOINT(SI)) {
885 const Instruction *Inst = CS.getInstruction();
886 ReturnVal = lowerRangeToAssertZExt(DAG, *Inst, ReturnVal);
887 setValue(Inst, ReturnVal);
891 void SelectionDAGBuilder::LowerCallSiteWithDeoptBundle(
892 ImmutableCallSite CS, SDValue Callee, const BasicBlock *EHPadBB) {
893 LowerCallSiteWithDeoptBundleImpl(CS, Callee, EHPadBB,
894 /* VarArgDisallowed = */ false,
895 /* ForceVoidReturnTy = */ false);
898 void SelectionDAGBuilder::visitGCResult(const GCResultInst &CI) {
899 // The result value of the gc_result is simply the result of the actual
900 // call. We've already emitted this, so just grab the value.
901 const Instruction *I = CI.getStatepoint();
903 if (I->getParent() != CI.getParent()) {
904 // Statepoint is in different basic block so we should have stored call
905 // result in a virtual register.
906 // We can not use default getValue() functionality to copy value from this
907 // register because statepoint and actual call return types can be
908 // different, and getValue() will use CopyFromReg of the wrong type,
909 // which is always i32 in our case.
910 PointerType *CalleeType = cast<PointerType>(
911 ImmutableStatepoint(I).getCalledValue()->getType());
913 cast<FunctionType>(CalleeType->getElementType())->getReturnType();
914 SDValue CopyFromReg = getCopyFromRegs(I, RetTy);
916 assert(CopyFromReg.getNode());
917 setValue(&CI, CopyFromReg);
919 setValue(&CI, getValue(I));
923 void SelectionDAGBuilder::visitGCRelocate(const GCRelocateInst &Relocate) {
926 // We skip this check for relocates not in the same basic block as their
927 // statepoint. It would be too expensive to preserve validation info through
928 // different basic blocks.
929 if (Relocate.getStatepoint()->getParent() == Relocate.getParent())
930 StatepointLowering.relocCallVisited(Relocate);
932 auto *Ty = Relocate.getType()->getScalarType();
933 if (auto IsManaged = GFI->getStrategy().isGCManagedPointer(Ty))
934 assert(*IsManaged && "Non gc managed pointer relocated!");
937 const Value *DerivedPtr = Relocate.getDerivedPtr();
938 SDValue SD = getValue(DerivedPtr);
940 auto &SpillMap = FuncInfo.StatepointSpillMaps[Relocate.getStatepoint()];
941 auto SlotIt = SpillMap.find(DerivedPtr);
942 assert(SlotIt != SpillMap.end() && "Relocating not lowered gc value");
943 Optional<int> DerivedPtrLocation = SlotIt->second;
945 // We didn't need to spill these special cases (constants and allocas).
946 // See the handling in spillIncomingValueForStatepoint for detail.
947 if (!DerivedPtrLocation) {
948 setValue(&Relocate, SD);
952 SDValue SpillSlot = DAG.getTargetFrameIndex(*DerivedPtrLocation,
955 // Be conservative: flush all pending loads
956 // TODO: Probably we can be less restrictive on this,
957 // it may allow more scheduling opportunities.
958 SDValue Chain = getRoot();
961 DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, SpillSlot,
962 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
963 *DerivedPtrLocation));
965 // Again, be conservative, don't emit pending loads
966 DAG.setRoot(SpillLoad.getValue(1));
968 assert(SpillLoad.getNode());
969 setValue(&Relocate, SpillLoad);
972 void SelectionDAGBuilder::LowerDeoptimizeCall(const CallInst *CI) {
973 const auto &TLI = DAG.getTargetLoweringInfo();
974 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(RTLIB::DEOPTIMIZE),
975 TLI.getPointerTy(DAG.getDataLayout()));
977 // We don't lower calls to __llvm_deoptimize as varargs, but as a regular
978 // call. We also do not lower the return value to any virtual register, and
979 // change the immediately following return to a trap instruction.
980 LowerCallSiteWithDeoptBundleImpl(CI, Callee, /* EHPadBB = */ nullptr,
981 /* VarArgDisallowed = */ true,
982 /* ForceVoidReturnTy = */ true);
985 void SelectionDAGBuilder::LowerDeoptimizingReturn() {
986 // We do not lower the return value from llvm.deoptimize to any virtual
987 // register, and change the immediately following return to a trap
989 if (DAG.getTarget().Options.TrapUnreachable)
991 DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));