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
246 Builder.DAG.getTargetFrameIndex(*Index, Builder.getFrameIndexTy());
247 Builder.StatepointLowering.setLocation(Incoming, Loc);
250 /// Remove any duplicate (as SDValues) from the derived pointer pairs. This
251 /// is not required for correctness. It's purpose is to reduce the size of
252 /// StackMap section. It has no effect on the number of spill slots required
253 /// or the actual lowering.
255 removeDuplicateGCPtrs(SmallVectorImpl<const Value *> &Bases,
256 SmallVectorImpl<const Value *> &Ptrs,
257 SmallVectorImpl<const GCRelocateInst *> &Relocs,
258 SelectionDAGBuilder &Builder,
259 FunctionLoweringInfo::StatepointSpillMap &SSM) {
260 DenseMap<SDValue, const Value *> Seen;
262 SmallVector<const Value *, 64> NewBases, NewPtrs;
263 SmallVector<const GCRelocateInst *, 64> NewRelocs;
264 for (size_t i = 0, e = Ptrs.size(); i < e; i++) {
265 SDValue SD = Builder.getValue(Ptrs[i]);
266 auto SeenIt = Seen.find(SD);
268 if (SeenIt == Seen.end()) {
269 // Only add non-duplicates
270 NewBases.push_back(Bases[i]);
271 NewPtrs.push_back(Ptrs[i]);
272 NewRelocs.push_back(Relocs[i]);
275 // Duplicate pointer found, note in SSM and move on:
276 SSM.DuplicateMap[Ptrs[i]] = SeenIt->second;
279 assert(Bases.size() >= NewBases.size());
280 assert(Ptrs.size() >= NewPtrs.size());
281 assert(Relocs.size() >= NewRelocs.size());
285 assert(Ptrs.size() == Bases.size());
286 assert(Ptrs.size() == Relocs.size());
289 /// Extract call from statepoint, lower it and return pointer to the
290 /// call node. Also update NodeMap so that getValue(statepoint) will
291 /// reference lowered call result
292 static std::pair<SDValue, SDNode *> lowerCallFromStatepointLoweringInfo(
293 SelectionDAGBuilder::StatepointLoweringInfo &SI,
294 SelectionDAGBuilder &Builder, SmallVectorImpl<SDValue> &PendingExports) {
296 SDValue ReturnValue, CallEndVal;
297 std::tie(ReturnValue, CallEndVal) =
298 Builder.lowerInvokable(SI.CLI, SI.EHPadBB);
299 SDNode *CallEnd = CallEndVal.getNode();
301 // Get a call instruction from the call sequence chain. Tail calls are not
302 // allowed. The following code is essentially reverse engineering X86's
305 // We are expecting DAG to have the following form:
307 // ch = eh_label (only in case of invoke statepoint)
308 // ch, glue = callseq_start ch
309 // ch, glue = X86::Call ch, glue
310 // ch, glue = callseq_end ch, glue
311 // get_return_value ch, glue
313 // get_return_value can either be a sequence of CopyFromReg instructions
314 // to grab the return value from the return register(s), or it can be a LOAD
315 // to load a value returned by reference via a stack slot.
317 bool HasDef = !SI.CLI.RetTy->isVoidTy();
319 if (CallEnd->getOpcode() == ISD::LOAD)
320 CallEnd = CallEnd->getOperand(0).getNode();
322 while (CallEnd->getOpcode() == ISD::CopyFromReg)
323 CallEnd = CallEnd->getOperand(0).getNode();
326 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
327 return std::make_pair(ReturnValue, CallEnd->getOperand(0).getNode());
330 /// Spill a value incoming to the statepoint. It might be either part of
332 /// or gcstate. In both cases unconditionally spill it on the stack unless it
333 /// is a null constant. Return pair with first element being frame index
334 /// containing saved value and second element with outgoing chain from the
336 static std::pair<SDValue, SDValue>
337 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
338 SelectionDAGBuilder &Builder) {
339 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
341 // Emit new store if we didn't do it for this ptr before
342 if (!Loc.getNode()) {
343 Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
345 int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
346 // We use TargetFrameIndex so that isel will not select it into LEA
347 Loc = Builder.DAG.getTargetFrameIndex(Index, Builder.getFrameIndexTy());
349 // TODO: We can create TokenFactor node instead of
350 // chaining stores one after another, this may allow
351 // a bit more optimal scheduling for them
354 // Right now we always allocate spill slots that are of the same
355 // size as the value we're about to spill (the size of spillee can
356 // vary since we spill vectors of pointers too). At some point we
357 // can consider allowing spills of smaller values to larger slots
358 // (i.e. change the '==' in the assert below to a '>=').
359 MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
360 assert((MFI.getObjectSize(Index) * 8) == Incoming.getValueSizeInBits() &&
361 "Bad spill: stack slot does not match!");
364 Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
365 MachinePointerInfo::getFixedStack(
366 Builder.DAG.getMachineFunction(), Index));
368 Builder.StatepointLowering.setLocation(Incoming, Loc);
371 assert(Loc.getNode());
372 return std::make_pair(Loc, Chain);
375 /// Lower a single value incoming to a statepoint node. This value can be
376 /// either a deopt value or a gc value, the handling is the same. We special
377 /// case constants and allocas, then fall back to spilling if required.
378 static void lowerIncomingStatepointValue(SDValue Incoming, bool LiveInOnly,
379 SmallVectorImpl<SDValue> &Ops,
380 SelectionDAGBuilder &Builder) {
381 SDValue Chain = Builder.getRoot();
383 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
384 // If the original value was a constant, make sure it gets recorded as
385 // such in the stackmap. This is required so that the consumer can
386 // parse any internal format to the deopt state. It also handles null
387 // pointers and other constant pointers in GC states. Note the constant
388 // vectors do not appear to actually hit this path and that anything larger
389 // than an i64 value (not type!) will fail asserts here.
390 pushStackMapConstant(Ops, Builder, C->getSExtValue());
391 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
392 // This handles allocas as arguments to the statepoint (this is only
393 // really meaningful for a deopt value. For GC, we'd be trying to
394 // relocate the address of the alloca itself?)
395 assert(Incoming.getValueType() == Builder.getFrameIndexTy() &&
396 "Incoming value is a frame index!");
397 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
398 Builder.getFrameIndexTy()));
399 } else if (LiveInOnly) {
400 // If this value is live in (not live-on-return, or live-through), we can
401 // treat it the same way patchpoint treats it's "live in" values. We'll
402 // end up folding some of these into stack references, but they'll be
403 // handled by the register allocator. Note that we do not have the notion
404 // of a late use so these values might be placed in registers which are
405 // clobbered by the call. This is fine for live-in.
406 Ops.push_back(Incoming);
408 // Otherwise, locate a spill slot and explicitly spill it so it
409 // can be found by the runtime later. We currently do not support
410 // tracking values through callee saved registers to their eventual
411 // spill location. This would be a useful optimization, but would
412 // need to be optional since it requires a lot of complexity on the
413 // runtime side which not all would support.
414 auto Res = spillIncomingStatepointValue(Incoming, Chain, Builder);
415 Ops.push_back(Res.first);
419 Builder.DAG.setRoot(Chain);
422 /// Lower deopt state and gc pointer arguments of the statepoint. The actual
423 /// lowering is described in lowerIncomingStatepointValue. This function is
424 /// responsible for lowering everything in the right position and playing some
425 /// tricks to avoid redundant stack manipulation where possible. On
426 /// completion, 'Ops' will contain ready to use operands for machine code
427 /// statepoint. The chain nodes will have already been created and the DAG root
428 /// will be set to the last value spilled (if any were).
430 lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
431 SelectionDAGBuilder::StatepointLoweringInfo &SI,
432 SelectionDAGBuilder &Builder) {
433 // Lower the deopt and gc arguments for this statepoint. Layout will be:
434 // deopt argument length, deopt arguments.., gc arguments...
436 if (auto *GFI = Builder.GFI) {
437 // Check that each of the gc pointer and bases we've gotten out of the
438 // safepoint is something the strategy thinks might be a pointer (or vector
439 // of pointers) into the GC heap. This is basically just here to help catch
440 // errors during statepoint insertion. TODO: This should actually be in the
441 // Verifier, but we can't get to the GCStrategy from there (yet).
442 GCStrategy &S = GFI->getStrategy();
443 for (const Value *V : SI.Bases) {
444 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
445 if (Opt.hasValue()) {
446 assert(Opt.getValue() &&
447 "non gc managed base pointer found in statepoint");
450 for (const Value *V : SI.Ptrs) {
451 auto Opt = S.isGCManagedPointer(V->getType()->getScalarType());
452 if (Opt.hasValue()) {
453 assert(Opt.getValue() &&
454 "non gc managed derived pointer found in statepoint");
457 assert(SI.Bases.size() == SI.Ptrs.size() && "Pointer without base!");
459 assert(SI.Bases.empty() && "No gc specified, so cannot relocate pointers!");
460 assert(SI.Ptrs.empty() && "No gc specified, so cannot relocate pointers!");
464 // Figure out what lowering strategy we're going to use for each part
465 // Note: Is is conservatively correct to lower both "live-in" and "live-out"
466 // as "live-through". A "live-through" variable is one which is "live-in",
467 // "live-out", and live throughout the lifetime of the call (i.e. we can find
468 // it from any PC within the transitive callee of the statepoint). In
469 // particular, if the callee spills callee preserved registers we may not
470 // be able to find a value placed in that register during the call. This is
471 // fine for live-out, but not for live-through. If we were willing to make
472 // assumptions about the code generator producing the callee, we could
473 // potentially allow live-through values in callee saved registers.
474 const bool LiveInDeopt =
475 SI.StatepointFlags & (uint64_t)StatepointFlags::DeoptLiveIn;
477 auto isGCValue =[&](const Value *V) {
478 return is_contained(SI.Ptrs, V) || is_contained(SI.Bases, V);
481 // Before we actually start lowering (and allocating spill slots for values),
482 // reserve any stack slots which we judge to be profitable to reuse for a
483 // particular value. This is purely an optimization over the code below and
484 // doesn't change semantics at all. It is important for performance that we
485 // reserve slots for both deopt and gc values before lowering either.
486 for (const Value *V : SI.DeoptState) {
487 if (!LiveInDeopt || isGCValue(V))
488 reservePreviousStackSlotForValue(V, Builder);
490 for (unsigned i = 0; i < SI.Bases.size(); ++i) {
491 reservePreviousStackSlotForValue(SI.Bases[i], Builder);
492 reservePreviousStackSlotForValue(SI.Ptrs[i], Builder);
495 // First, prefix the list with the number of unique values to be
496 // lowered. Note that this is the number of *Values* not the
497 // number of SDValues required to lower them.
498 const int NumVMSArgs = SI.DeoptState.size();
499 pushStackMapConstant(Ops, Builder, NumVMSArgs);
501 // The vm state arguments are lowered in an opaque manner. We do not know
502 // what type of values are contained within.
503 for (const Value *V : SI.DeoptState) {
504 SDValue Incoming = Builder.getValue(V);
505 const bool LiveInValue = LiveInDeopt && !isGCValue(V);
506 lowerIncomingStatepointValue(Incoming, LiveInValue, Ops, Builder);
509 // Finally, go ahead and lower all the gc arguments. There's no prefixed
510 // length for this one. After lowering, we'll have the base and pointer
511 // arrays interwoven with each (lowered) base pointer immediately followed by
512 // it's (lowered) derived pointer. i.e
513 // (base[0], ptr[0], base[1], ptr[1], ...)
514 for (unsigned i = 0; i < SI.Bases.size(); ++i) {
515 const Value *Base = SI.Bases[i];
516 lowerIncomingStatepointValue(Builder.getValue(Base), /*LiveInOnly*/ false,
519 const Value *Ptr = SI.Ptrs[i];
520 lowerIncomingStatepointValue(Builder.getValue(Ptr), /*LiveInOnly*/ false,
524 // If there are any explicit spill slots passed to the statepoint, record
525 // them, but otherwise do not do anything special. These are user provided
526 // allocas and give control over placement to the consumer. In this case,
527 // it is the contents of the slot which may get updated, not the pointer to
529 for (Value *V : SI.GCArgs) {
530 SDValue Incoming = Builder.getValue(V);
531 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
532 // This handles allocas as arguments to the statepoint
533 assert(Incoming.getValueType() == Builder.getFrameIndexTy() &&
534 "Incoming value is a frame index!");
535 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
536 Builder.getFrameIndexTy()));
540 // Record computed locations for all lowered values.
541 // This can not be embedded in lowering loops as we need to record *all*
542 // values, while previous loops account only values with unique SDValues.
543 const Instruction *StatepointInstr = SI.StatepointInstr;
544 auto &SpillMap = Builder.FuncInfo.StatepointSpillMaps[StatepointInstr];
546 for (const GCRelocateInst *Relocate : SI.GCRelocates) {
547 const Value *V = Relocate->getDerivedPtr();
548 SDValue SDV = Builder.getValue(V);
549 SDValue Loc = Builder.StatepointLowering.getLocation(SDV);
552 SpillMap.SlotMap[V] = cast<FrameIndexSDNode>(Loc)->getIndex();
554 // Record value as visited, but not spilled. This is case for allocas
555 // and constants. For this values we can avoid emitting spill load while
556 // visiting corresponding gc_relocate.
557 // Actually we do not need to record them in this map at all.
558 // We do this only to check that we are not relocating any unvisited
560 SpillMap.SlotMap[V] = None;
562 // Default llvm mechanisms for exporting values which are used in
563 // different basic blocks does not work for gc relocates.
564 // Note that it would be incorrect to teach llvm that all relocates are
565 // uses of the corresponding values so that it would automatically
566 // export them. Relocates of the spilled values does not use original
568 if (Relocate->getParent() != StatepointInstr->getParent())
569 Builder.ExportFromCurrentBlock(V);
574 SDValue SelectionDAGBuilder::LowerAsSTATEPOINT(
575 SelectionDAGBuilder::StatepointLoweringInfo &SI) {
576 // The basic scheme here is that information about both the original call and
577 // the safepoint is encoded in the CallInst. We create a temporary call and
578 // lower it, then reverse engineer the calling sequence.
582 StatepointLowering.startNewStatepoint(*this);
585 // We schedule gc relocates before removeDuplicateGCPtrs since we _will_
586 // encounter the duplicate gc relocates we elide in removeDuplicateGCPtrs.
587 for (auto *Reloc : SI.GCRelocates)
588 if (Reloc->getParent() == SI.StatepointInstr->getParent())
589 StatepointLowering.scheduleRelocCall(*Reloc);
592 // Remove any redundant llvm::Values which map to the same SDValue as another
593 // input. Also has the effect of removing duplicates in the original
594 // llvm::Value input list as well. This is a useful optimization for
595 // reducing the size of the StackMap section. It has no other impact.
596 removeDuplicateGCPtrs(SI.Bases, SI.Ptrs, SI.GCRelocates, *this,
597 FuncInfo.StatepointSpillMaps[SI.StatepointInstr]);
598 assert(SI.Bases.size() == SI.Ptrs.size() &&
599 SI.Ptrs.size() == SI.GCRelocates.size());
601 // Lower statepoint vmstate and gcstate arguments
602 SmallVector<SDValue, 10> LoweredMetaArgs;
603 lowerStatepointMetaArgs(LoweredMetaArgs, SI, *this);
605 // Now that we've emitted the spills, we need to update the root so that the
606 // call sequence is ordered correctly.
607 SI.CLI.setChain(getRoot());
609 // Get call node, we will replace it later with statepoint
612 std::tie(ReturnVal, CallNode) =
613 lowerCallFromStatepointLoweringInfo(SI, *this, PendingExports);
615 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
616 // nodes with all the appropriate arguments and return values.
618 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
619 SDValue Chain = CallNode->getOperand(0);
622 bool CallHasIncomingGlue = CallNode->getGluedNode();
623 if (CallHasIncomingGlue) {
624 // Glue is always last operand
625 Glue = CallNode->getOperand(CallNode->getNumOperands() - 1);
628 // Build the GC_TRANSITION_START node if necessary.
630 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
631 // order in which they appear in the call to the statepoint intrinsic. If
632 // any of the operands is a pointer-typed, that operand is immediately
633 // followed by a SRCVALUE for the pointer that may be used during lowering
634 // (e.g. to form MachinePointerInfo values for loads/stores).
635 const bool IsGCTransition =
636 (SI.StatepointFlags & (uint64_t)StatepointFlags::GCTransition) ==
637 (uint64_t)StatepointFlags::GCTransition;
638 if (IsGCTransition) {
639 SmallVector<SDValue, 8> TSOps;
642 TSOps.push_back(Chain);
644 // Add GC transition arguments
645 for (const Value *V : SI.GCTransitionArgs) {
646 TSOps.push_back(getValue(V));
647 if (V->getType()->isPointerTy())
648 TSOps.push_back(DAG.getSrcValue(V));
651 // Add glue if necessary
652 if (CallHasIncomingGlue)
653 TSOps.push_back(Glue);
655 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
657 SDValue GCTransitionStart =
658 DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps);
660 Chain = GCTransitionStart.getValue(0);
661 Glue = GCTransitionStart.getValue(1);
664 // TODO: Currently, all of these operands are being marked as read/write in
665 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
666 // and flags to be read-only.
667 SmallVector<SDValue, 40> Ops;
669 // Add the <id> and <numBytes> constants.
670 Ops.push_back(DAG.getTargetConstant(SI.ID, getCurSDLoc(), MVT::i64));
672 DAG.getTargetConstant(SI.NumPatchBytes, getCurSDLoc(), MVT::i32));
674 // Calculate and push starting position of vmstate arguments
675 // Get number of arguments incoming directly into call node
676 unsigned NumCallRegArgs =
677 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
678 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32));
681 SDValue CallTarget = SDValue(CallNode->getOperand(1).getNode(), 0);
682 Ops.push_back(CallTarget);
684 // Add call arguments
685 // Get position of register mask in the call
686 SDNode::op_iterator RegMaskIt;
687 if (CallHasIncomingGlue)
688 RegMaskIt = CallNode->op_end() - 2;
690 RegMaskIt = CallNode->op_end() - 1;
691 Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt);
693 // Add a constant argument for the calling convention
694 pushStackMapConstant(Ops, *this, SI.CLI.CallConv);
696 // Add a constant argument for the flags
697 uint64_t Flags = SI.StatepointFlags;
698 assert(((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0) &&
699 "Unknown flag used");
700 pushStackMapConstant(Ops, *this, Flags);
702 // Insert all vmstate and gcstate arguments
703 Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end());
705 // Add register mask from call node
706 Ops.push_back(*RegMaskIt);
709 Ops.push_back(Chain);
711 // Same for the glue, but we add it only if original call had it
715 // Compute return values. Provide a glue output since we consume one as
716 // input. This allows someone else to chain off us as needed.
717 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
719 SDNode *StatepointMCNode =
720 DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
722 SDNode *SinkNode = StatepointMCNode;
724 // Build the GC_TRANSITION_END node if necessary.
726 // See the comment above regarding GC_TRANSITION_START for the layout of
727 // the operands to the GC_TRANSITION_END node.
728 if (IsGCTransition) {
729 SmallVector<SDValue, 8> TEOps;
732 TEOps.push_back(SDValue(StatepointMCNode, 0));
734 // Add GC transition arguments
735 for (const Value *V : SI.GCTransitionArgs) {
736 TEOps.push_back(getValue(V));
737 if (V->getType()->isPointerTy())
738 TEOps.push_back(DAG.getSrcValue(V));
742 TEOps.push_back(SDValue(StatepointMCNode, 1));
744 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
746 SDValue GCTransitionStart =
747 DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps);
749 SinkNode = GCTransitionStart.getNode();
752 // Replace original call
753 DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root
754 // Remove original call node
755 DAG.DeleteNode(CallNode);
757 // DON'T set the root - under the assumption that it's already set past the
758 // inserted node we created.
760 // TODO: A better future implementation would be to emit a single variable
761 // argument, variable return value STATEPOINT node here and then hookup the
762 // return value of each gc.relocate to the respective output of the
763 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
764 // to actually be possible today.
770 SelectionDAGBuilder::LowerStatepoint(ImmutableStatepoint ISP,
771 const BasicBlock *EHPadBB /*= nullptr*/) {
772 assert(ISP.getCallSite().getCallingConv() != CallingConv::AnyReg &&
773 "anyregcc is not supported on statepoints!");
776 // If this is a malformed statepoint, report it early to simplify debugging.
777 // This should catch any IR level mistake that's made when constructing or
778 // transforming statepoints.
781 // Check that the associated GCStrategy expects to encounter statepoints.
782 assert(GFI->getStrategy().useStatepoints() &&
783 "GCStrategy does not expect to encounter statepoints");
786 SDValue ActualCallee;
788 if (ISP.getNumPatchBytes() > 0) {
789 // If we've been asked to emit a nop sequence instead of a call instruction
790 // for this statepoint then don't lower the call target, but use a constant
791 // `null` instead. Not lowering the call target lets statepoint clients get
792 // away without providing a physical address for the symbolic call target at
795 const auto &TLI = DAG.getTargetLoweringInfo();
796 const auto &DL = DAG.getDataLayout();
798 unsigned AS = ISP.getCalledValue()->getType()->getPointerAddressSpace();
799 ActualCallee = DAG.getConstant(0, getCurSDLoc(), TLI.getPointerTy(DL, AS));
801 ActualCallee = getValue(ISP.getCalledValue());
804 StatepointLoweringInfo SI(DAG);
805 populateCallLoweringInfo(SI.CLI, ISP.getCallSite(),
806 ImmutableStatepoint::CallArgsBeginPos,
807 ISP.getNumCallArgs(), ActualCallee,
808 ISP.getActualReturnType(), false /* IsPatchPoint */);
810 for (const GCRelocateInst *Relocate : ISP.getRelocates()) {
811 SI.GCRelocates.push_back(Relocate);
812 SI.Bases.push_back(Relocate->getBasePtr());
813 SI.Ptrs.push_back(Relocate->getDerivedPtr());
816 SI.GCArgs = ArrayRef<const Use>(ISP.gc_args_begin(), ISP.gc_args_end());
817 SI.StatepointInstr = ISP.getInstruction();
818 SI.GCTransitionArgs =
819 ArrayRef<const Use>(ISP.gc_args_begin(), ISP.gc_args_end());
821 SI.DeoptState = ArrayRef<const Use>(ISP.vm_state_begin(), ISP.vm_state_end());
822 SI.StatepointFlags = ISP.getFlags();
823 SI.NumPatchBytes = ISP.getNumPatchBytes();
824 SI.EHPadBB = EHPadBB;
826 SDValue ReturnValue = LowerAsSTATEPOINT(SI);
828 // Export the result value if needed
829 const GCResultInst *GCResult = ISP.getGCResult();
830 Type *RetTy = ISP.getActualReturnType();
831 if (!RetTy->isVoidTy() && GCResult) {
832 if (GCResult->getParent() != ISP.getCallSite().getParent()) {
833 // Result value will be used in a different basic block so we need to
834 // export it now. Default exporting mechanism will not work here because
835 // statepoint call has a different type than the actual call. It means
836 // that by default llvm will create export register of the wrong type
837 // (always i32 in our case). So instead we need to create export register
838 // with correct type manually.
839 // TODO: To eliminate this problem we can remove gc.result intrinsics
840 // completely and make statepoint call to return a tuple.
841 unsigned Reg = FuncInfo.CreateRegs(RetTy);
842 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
843 DAG.getDataLayout(), Reg, RetTy);
844 SDValue Chain = DAG.getEntryNode();
846 RFV.getCopyToRegs(ReturnValue, DAG, getCurSDLoc(), Chain, nullptr);
847 PendingExports.push_back(Chain);
848 FuncInfo.ValueMap[ISP.getInstruction()] = Reg;
850 // Result value will be used in a same basic block. Don't export it or
851 // perform any explicit register copies.
852 // We'll replace the actuall call node shortly. gc_result will grab
854 setValue(ISP.getInstruction(), ReturnValue);
857 // The token value is never used from here on, just generate a poison value
858 setValue(ISP.getInstruction(), DAG.getIntPtrConstant(-1, getCurSDLoc()));
862 void SelectionDAGBuilder::LowerCallSiteWithDeoptBundleImpl(
863 ImmutableCallSite CS, SDValue Callee, const BasicBlock *EHPadBB,
864 bool VarArgDisallowed, bool ForceVoidReturnTy) {
865 StatepointLoweringInfo SI(DAG);
866 unsigned ArgBeginIndex = CS.arg_begin() - CS.getInstruction()->op_begin();
867 populateCallLoweringInfo(
868 SI.CLI, CS, ArgBeginIndex, CS.getNumArgOperands(), Callee,
869 ForceVoidReturnTy ? Type::getVoidTy(*DAG.getContext()) : CS.getType(),
871 if (!VarArgDisallowed)
872 SI.CLI.IsVarArg = CS.getFunctionType()->isVarArg();
874 auto DeoptBundle = *CS.getOperandBundle(LLVMContext::OB_deopt);
876 unsigned DefaultID = StatepointDirectives::DeoptBundleStatepointID;
878 auto SD = parseStatepointDirectivesFromAttrs(CS.getAttributes());
879 SI.ID = SD.StatepointID.getValueOr(DefaultID);
880 SI.NumPatchBytes = SD.NumPatchBytes.getValueOr(0);
883 ArrayRef<const Use>(DeoptBundle.Inputs.begin(), DeoptBundle.Inputs.end());
884 SI.StatepointFlags = static_cast<uint64_t>(StatepointFlags::None);
885 SI.EHPadBB = EHPadBB;
887 // NB! The GC arguments are deliberately left empty.
889 if (SDValue ReturnVal = LowerAsSTATEPOINT(SI)) {
890 const Instruction *Inst = CS.getInstruction();
891 ReturnVal = lowerRangeToAssertZExt(DAG, *Inst, ReturnVal);
892 setValue(Inst, ReturnVal);
896 void SelectionDAGBuilder::LowerCallSiteWithDeoptBundle(
897 ImmutableCallSite CS, SDValue Callee, const BasicBlock *EHPadBB) {
898 LowerCallSiteWithDeoptBundleImpl(CS, Callee, EHPadBB,
899 /* VarArgDisallowed = */ false,
900 /* ForceVoidReturnTy = */ false);
903 void SelectionDAGBuilder::visitGCResult(const GCResultInst &CI) {
904 // The result value of the gc_result is simply the result of the actual
905 // call. We've already emitted this, so just grab the value.
906 const Instruction *I = CI.getStatepoint();
908 if (I->getParent() != CI.getParent()) {
909 // Statepoint is in different basic block so we should have stored call
910 // result in a virtual register.
911 // We can not use default getValue() functionality to copy value from this
912 // register because statepoint and actual call return types can be
913 // different, and getValue() will use CopyFromReg of the wrong type,
914 // which is always i32 in our case.
915 PointerType *CalleeType = cast<PointerType>(
916 ImmutableStatepoint(I).getCalledValue()->getType());
918 cast<FunctionType>(CalleeType->getElementType())->getReturnType();
919 SDValue CopyFromReg = getCopyFromRegs(I, RetTy);
921 assert(CopyFromReg.getNode());
922 setValue(&CI, CopyFromReg);
924 setValue(&CI, getValue(I));
928 void SelectionDAGBuilder::visitGCRelocate(const GCRelocateInst &Relocate) {
931 // We skip this check for relocates not in the same basic block as their
932 // statepoint. It would be too expensive to preserve validation info through
933 // different basic blocks.
934 if (Relocate.getStatepoint()->getParent() == Relocate.getParent())
935 StatepointLowering.relocCallVisited(Relocate);
937 auto *Ty = Relocate.getType()->getScalarType();
938 if (auto IsManaged = GFI->getStrategy().isGCManagedPointer(Ty))
939 assert(*IsManaged && "Non gc managed pointer relocated!");
942 const Value *DerivedPtr = Relocate.getDerivedPtr();
943 SDValue SD = getValue(DerivedPtr);
945 auto &SpillMap = FuncInfo.StatepointSpillMaps[Relocate.getStatepoint()];
946 auto SlotIt = SpillMap.find(DerivedPtr);
947 assert(SlotIt != SpillMap.end() && "Relocating not lowered gc value");
948 Optional<int> DerivedPtrLocation = SlotIt->second;
950 // We didn't need to spill these special cases (constants and allocas).
951 // See the handling in spillIncomingValueForStatepoint for detail.
952 if (!DerivedPtrLocation) {
953 setValue(&Relocate, SD);
958 DAG.getTargetFrameIndex(*DerivedPtrLocation, getFrameIndexTy());
960 // Be conservative: flush all pending loads
961 // TODO: Probably we can be less restrictive on this,
962 // it may allow more scheduling opportunities.
963 SDValue Chain = getRoot();
966 DAG.getLoad(DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
968 getCurSDLoc(), Chain, SpillSlot,
969 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
970 *DerivedPtrLocation));
972 // Again, be conservative, don't emit pending loads
973 DAG.setRoot(SpillLoad.getValue(1));
975 assert(SpillLoad.getNode());
976 setValue(&Relocate, SpillLoad);
979 void SelectionDAGBuilder::LowerDeoptimizeCall(const CallInst *CI) {
980 const auto &TLI = DAG.getTargetLoweringInfo();
981 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(RTLIB::DEOPTIMIZE),
982 TLI.getPointerTy(DAG.getDataLayout()));
984 // We don't lower calls to __llvm_deoptimize as varargs, but as a regular
985 // call. We also do not lower the return value to any virtual register, and
986 // change the immediately following return to a trap instruction.
987 LowerCallSiteWithDeoptBundleImpl(CI, Callee, /* EHPadBB = */ nullptr,
988 /* VarArgDisallowed = */ true,
989 /* ForceVoidReturnTy = */ true);
992 void SelectionDAGBuilder::LowerDeoptimizingReturn() {
993 // We do not lower the return value from llvm.deoptimize to any virtual
994 // register, and change the immediately following return to a trap
996 if (DAG.getTarget().Options.TrapUnreachable)
998 DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));