1//===- StatepointLowering.cpp - SDAGBuilder's statepoint code -------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file includes support code use by SelectionDAGBuilder when lowering a
10// statepoint sequence in SelectionDAG IR.
11//
12//===----------------------------------------------------------------------===//
13
14#include "StatepointLowering.h"
15#include "SelectionDAGBuilder.h"
16#include "llvm/ADT/ArrayRef.h"
17#include "llvm/ADT/STLExtras.h"
18#include "llvm/ADT/SetVector.h"
19#include "llvm/ADT/SmallBitVector.h"
20#include "llvm/ADT/SmallSet.h"
21#include "llvm/ADT/SmallVector.h"
22#include "llvm/ADT/Statistic.h"
23#include "llvm/CodeGen/FunctionLoweringInfo.h"
24#include "llvm/CodeGen/GCMetadata.h"
25#include "llvm/CodeGen/ISDOpcodes.h"
26#include "llvm/CodeGen/MachineFrameInfo.h"
27#include "llvm/CodeGen/MachineFunction.h"
28#include "llvm/CodeGen/MachineMemOperand.h"
29#include "llvm/CodeGen/RuntimeLibcallUtil.h"
30#include "llvm/CodeGen/SelectionDAG.h"
31#include "llvm/CodeGen/SelectionDAGNodes.h"
32#include "llvm/CodeGen/StackMaps.h"
33#include "llvm/CodeGen/TargetLowering.h"
34#include "llvm/CodeGen/TargetOpcodes.h"
35#include "llvm/CodeGenTypes/MachineValueType.h"
36#include "llvm/IR/CallingConv.h"
37#include "llvm/IR/DerivedTypes.h"
38#include "llvm/IR/GCStrategy.h"
39#include "llvm/IR/Instruction.h"
40#include "llvm/IR/Instructions.h"
41#include "llvm/IR/LLVMContext.h"
42#include "llvm/IR/Statepoint.h"
43#include "llvm/IR/Type.h"
44#include "llvm/Support/Casting.h"
45#include "llvm/Support/CommandLine.h"
46#include "llvm/Target/TargetMachine.h"
47#include "llvm/Target/TargetOptions.h"
48#include <cassert>
49#include <cstddef>
50#include <cstdint>
51#include <iterator>
52#include <tuple>
53#include <utility>
54
55using namespace llvm;
56
57#define DEBUG_TYPE "statepoint-lowering"
58
59STATISTIC(NumSlotsAllocatedForStatepoints,
60 "Number of stack slots allocated for statepoints");
61STATISTIC(NumOfStatepoints, "Number of statepoint nodes encountered");
62STATISTIC(StatepointMaxSlotsRequired,
63 "Maximum number of stack slots required for a singe statepoint");
64
65static cl::opt<bool> UseRegistersForDeoptValues(
66 "use-registers-for-deopt-values", cl::Hidden, cl::init(Val: false),
67 cl::desc("Allow using registers for non pointer deopt args"));
68
69static cl::opt<bool> UseRegistersForGCPointersInLandingPad(
70 "use-registers-for-gc-values-in-landing-pad", cl::Hidden, cl::init(Val: false),
71 cl::desc("Allow using registers for gc pointer in landing pad"));
72
73static cl::opt<unsigned> MaxRegistersForGCPointers(
74 "max-registers-for-gc-values", cl::Hidden, cl::init(Val: 0),
75 cl::desc("Max number of VRegs allowed to pass GC pointer meta args in"));
76
77typedef FunctionLoweringInfo::StatepointRelocationRecord RecordType;
78
79static void pushStackMapConstant(SmallVectorImpl<SDValue>& Ops,
80 SelectionDAGBuilder &Builder, uint64_t Value) {
81 SDLoc L = Builder.getCurSDLoc();
82 Ops.push_back(Elt: Builder.DAG.getTargetConstant(Val: StackMaps::ConstantOp, DL: L,
83 VT: MVT::i64));
84 Ops.push_back(Elt: Builder.DAG.getTargetConstant(Val: Value, DL: L, VT: MVT::i64));
85}
86
87void StatepointLoweringState::startNewStatepoint(SelectionDAGBuilder &Builder) {
88 // Consistency check
89 assert(PendingGCRelocateCalls.empty() &&
90 "Trying to visit statepoint before finished processing previous one");
91 Locations.clear();
92 NextSlotToAllocate = 0;
93 // Need to resize this on each safepoint - we need the two to stay in sync and
94 // the clear patterns of a SelectionDAGBuilder have no relation to
95 // FunctionLoweringInfo. Also need to ensure used bits get cleared.
96 AllocatedStackSlots.clear();
97 AllocatedStackSlots.resize(N: Builder.FuncInfo.StatepointStackSlots.size());
98}
99
100void StatepointLoweringState::clear() {
101 Locations.clear();
102 AllocatedStackSlots.clear();
103 assert(PendingGCRelocateCalls.empty() &&
104 "cleared before statepoint sequence completed");
105}
106
107SDValue
108StatepointLoweringState::allocateStackSlot(EVT ValueType,
109 SelectionDAGBuilder &Builder) {
110 NumSlotsAllocatedForStatepoints++;
111 MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
112
113 unsigned SpillSize = ValueType.getStoreSize();
114 assert((SpillSize * 8) ==
115 (-8u & (7 + ValueType.getSizeInBits())) && // Round up modulo 8.
116 "Size not in bytes?");
117
118 // First look for a previously created stack slot which is not in
119 // use (accounting for the fact arbitrary slots may already be
120 // reserved), or to create a new stack slot and use it.
121
122 const size_t NumSlots = AllocatedStackSlots.size();
123 assert(NextSlotToAllocate <= NumSlots && "Broken invariant");
124
125 assert(AllocatedStackSlots.size() ==
126 Builder.FuncInfo.StatepointStackSlots.size() &&
127 "Broken invariant");
128
129 for (; NextSlotToAllocate < NumSlots; NextSlotToAllocate++) {
130 if (!AllocatedStackSlots.test(Idx: NextSlotToAllocate)) {
131 const int FI = Builder.FuncInfo.StatepointStackSlots[NextSlotToAllocate];
132 if (MFI.getObjectSize(ObjectIdx: FI) == SpillSize) {
133 AllocatedStackSlots.set(NextSlotToAllocate);
134 // TODO: Is ValueType the right thing to use here?
135 return Builder.DAG.getFrameIndex(FI, VT: ValueType);
136 }
137 }
138 }
139
140 // Couldn't find a free slot, so create a new one:
141
142 SDValue SpillSlot = Builder.DAG.CreateStackTemporary(VT: ValueType);
143 const unsigned FI = cast<FrameIndexSDNode>(Val&: SpillSlot)->getIndex();
144 MFI.markAsStatepointSpillSlotObjectIndex(ObjectIdx: FI);
145
146 Builder.FuncInfo.StatepointStackSlots.push_back(Elt: FI);
147 AllocatedStackSlots.resize(N: AllocatedStackSlots.size()+1, t: true);
148 assert(AllocatedStackSlots.size() ==
149 Builder.FuncInfo.StatepointStackSlots.size() &&
150 "Broken invariant");
151
152 StatepointMaxSlotsRequired.updateMax(
153 V: Builder.FuncInfo.StatepointStackSlots.size());
154
155 return SpillSlot;
156}
157
158/// Utility function for reservePreviousStackSlotForValue. Tries to find
159/// stack slot index to which we have spilled value for previous statepoints.
160/// LookUpDepth specifies maximum DFS depth this function is allowed to look.
161static std::optional<int> findPreviousSpillSlot(const Value *Val,
162 SelectionDAGBuilder &Builder,
163 int LookUpDepth) {
164 // Can not look any further - give up now
165 if (LookUpDepth <= 0)
166 return std::nullopt;
167
168 // Spill location is known for gc relocates
169 if (const auto *Relocate = dyn_cast<GCRelocateInst>(Val)) {
170 const Value *Statepoint = Relocate->getStatepoint();
171 assert((isa<GCStatepointInst>(Statepoint) || isa<UndefValue>(Statepoint)) &&
172 "GetStatepoint must return one of two types");
173 if (isa<UndefValue>(Val: Statepoint))
174 return std::nullopt;
175
176 const auto &RelocationMap = Builder.FuncInfo.StatepointRelocationMaps
177 [cast<GCStatepointInst>(Val: Statepoint)];
178
179 auto It = RelocationMap.find(Val: Relocate);
180 if (It == RelocationMap.end())
181 return std::nullopt;
182
183 auto &Record = It->second;
184 if (Record.type != RecordType::Spill)
185 return std::nullopt;
186
187 return Record.payload.FI;
188 }
189
190 // Look through bitcast instructions.
191 if (const BitCastInst *Cast = dyn_cast<BitCastInst>(Val))
192 return findPreviousSpillSlot(Val: Cast->getOperand(i_nocapture: 0), Builder, LookUpDepth: LookUpDepth - 1);
193
194 // Look through phi nodes
195 // All incoming values should have same known stack slot, otherwise result
196 // is unknown.
197 if (const PHINode *Phi = dyn_cast<PHINode>(Val)) {
198 std::optional<int> MergedResult;
199
200 for (const auto &IncomingValue : Phi->incoming_values()) {
201 std::optional<int> SpillSlot =
202 findPreviousSpillSlot(Val: IncomingValue, Builder, LookUpDepth: LookUpDepth - 1);
203 if (!SpillSlot)
204 return std::nullopt;
205
206 if (MergedResult && *MergedResult != *SpillSlot)
207 return std::nullopt;
208
209 MergedResult = SpillSlot;
210 }
211 return MergedResult;
212 }
213
214 // TODO: We can do better for PHI nodes. In cases like this:
215 // ptr = phi(relocated_pointer, not_relocated_pointer)
216 // statepoint(ptr)
217 // We will return that stack slot for ptr is unknown. And later we might
218 // assign different stack slots for ptr and relocated_pointer. This limits
219 // llvm's ability to remove redundant stores.
220 // Unfortunately it's hard to accomplish in current infrastructure.
221 // We use this function to eliminate spill store completely, while
222 // in example we still need to emit store, but instead of any location
223 // we need to use special "preferred" location.
224
225 // TODO: handle simple updates. If a value is modified and the original
226 // value is no longer live, it would be nice to put the modified value in the
227 // same slot. This allows folding of the memory accesses for some
228 // instructions types (like an increment).
229 // statepoint (i)
230 // i1 = i+1
231 // statepoint (i1)
232 // However we need to be careful for cases like this:
233 // statepoint(i)
234 // i1 = i+1
235 // statepoint(i, i1)
236 // Here we want to reserve spill slot for 'i', but not for 'i+1'. If we just
237 // put handling of simple modifications in this function like it's done
238 // for bitcasts we might end up reserving i's slot for 'i+1' because order in
239 // which we visit values is unspecified.
240
241 // Don't know any information about this instruction
242 return std::nullopt;
243}
244
245/// Return true if-and-only-if the given SDValue can be lowered as either a
246/// constant argument or a stack reference. The key point is that the value
247/// doesn't need to be spilled or tracked as a vreg use.
248static bool willLowerDirectly(SDValue Incoming) {
249 // We are making an unchecked assumption that the frame size <= 2^16 as that
250 // is the largest offset which can be encoded in the stackmap format.
251 if (isa<FrameIndexSDNode>(Val: Incoming))
252 return true;
253
254 // The largest constant describeable in the StackMap format is 64 bits.
255 // Potential Optimization: Constants values are sign extended by consumer,
256 // and thus there are many constants of static type > 64 bits whose value
257 // happens to be sext(Con64) and could thus be lowered directly.
258 if (Incoming.getValueType().getSizeInBits() > 64)
259 return false;
260
261 return isIntOrFPConstant(V: Incoming) || Incoming.isUndef();
262}
263
264/// Try to find existing copies of the incoming values in stack slots used for
265/// statepoint spilling. If we can find a spill slot for the incoming value,
266/// mark that slot as allocated, and reuse the same slot for this safepoint.
267/// This helps to avoid series of loads and stores that only serve to reshuffle
268/// values on the stack between calls.
269static void reservePreviousStackSlotForValue(const Value *IncomingValue,
270 SelectionDAGBuilder &Builder) {
271 SDValue Incoming = Builder.getValue(V: IncomingValue);
272
273 // If we won't spill this, we don't need to check for previously allocated
274 // stack slots.
275 if (willLowerDirectly(Incoming))
276 return;
277
278 SDValue OldLocation = Builder.StatepointLowering.getLocation(Val: Incoming);
279 if (OldLocation.getNode())
280 // Duplicates in input
281 return;
282
283 const int LookUpDepth = 6;
284 std::optional<int> Index =
285 findPreviousSpillSlot(Val: IncomingValue, Builder, LookUpDepth);
286 if (!Index)
287 return;
288
289 const auto &StatepointSlots = Builder.FuncInfo.StatepointStackSlots;
290
291 auto SlotIt = find(Range: StatepointSlots, Val: *Index);
292 assert(SlotIt != StatepointSlots.end() &&
293 "Value spilled to the unknown stack slot");
294
295 // This is one of our dedicated lowering slots
296 const int Offset = std::distance(first: StatepointSlots.begin(), last: SlotIt);
297 if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) {
298 // stack slot already assigned to someone else, can't use it!
299 // TODO: currently we reserve space for gc arguments after doing
300 // normal allocation for deopt arguments. We should reserve for
301 // _all_ deopt and gc arguments, then start allocating. This
302 // will prevent some moves being inserted when vm state changes,
303 // but gc state doesn't between two calls.
304 return;
305 }
306 // Reserve this stack slot
307 Builder.StatepointLowering.reserveStackSlot(Offset);
308
309 // Cache this slot so we find it when going through the normal
310 // assignment loop.
311 SDValue Loc =
312 Builder.DAG.getTargetFrameIndex(FI: *Index, VT: Builder.getFrameIndexTy());
313 Builder.StatepointLowering.setLocation(Val: Incoming, Location: Loc);
314}
315
316/// Extract call from statepoint, lower it and return pointer to the
317/// call node. Also update NodeMap so that getValue(statepoint) will
318/// reference lowered call result
319static std::pair<SDValue, SDNode *> lowerCallFromStatepointLoweringInfo(
320 SelectionDAGBuilder::StatepointLoweringInfo &SI,
321 SelectionDAGBuilder &Builder) {
322 SDValue ReturnValue, CallEndVal;
323 std::tie(args&: ReturnValue, args&: CallEndVal) =
324 Builder.lowerInvokable(CLI&: SI.CLI, EHPadBB: SI.EHPadBB);
325 SDNode *CallEnd = CallEndVal.getNode();
326
327 // Get a call instruction from the call sequence chain. Tail calls are not
328 // allowed. The following code is essentially reverse engineering X86's
329 // LowerCallTo.
330 //
331 // We are expecting DAG to have the following form:
332 //
333 // ch = eh_label (only in case of invoke statepoint)
334 // ch, glue = callseq_start ch
335 // ch, glue = X86::Call ch, glue
336 // ch, glue = callseq_end ch, glue
337 // get_return_value ch, glue
338 //
339 // get_return_value can either be a sequence of CopyFromReg instructions
340 // to grab the return value from the return register(s), or it can be a LOAD
341 // to load a value returned by reference via a stack slot.
342
343 if (CallEnd->getOpcode() == ISD::EH_LABEL)
344 CallEnd = CallEnd->getOperand(Num: 0).getNode();
345
346 bool HasDef = !SI.CLI.RetTy->isVoidTy();
347 if (HasDef) {
348 if (CallEnd->getOpcode() == ISD::LOAD)
349 CallEnd = CallEnd->getOperand(Num: 0).getNode();
350 else
351 while (CallEnd->getOpcode() == ISD::CopyFromReg)
352 CallEnd = CallEnd->getOperand(Num: 0).getNode();
353 }
354
355 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
356 return std::make_pair(x&: ReturnValue, y: CallEnd->getOperand(Num: 0).getNode());
357}
358
359static MachineMemOperand* getMachineMemOperand(MachineFunction &MF,
360 FrameIndexSDNode &FI) {
361 auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FI.getIndex());
362 auto MMOFlags = MachineMemOperand::MOStore |
363 MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile;
364 auto &MFI = MF.getFrameInfo();
365 return MF.getMachineMemOperand(PtrInfo, F: MMOFlags,
366 Size: MFI.getObjectSize(ObjectIdx: FI.getIndex()),
367 BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI.getIndex()));
368}
369
370/// Spill a value incoming to the statepoint. It might be either part of
371/// vmstate
372/// or gcstate. In both cases unconditionally spill it on the stack unless it
373/// is a null constant. Return pair with first element being frame index
374/// containing saved value and second element with outgoing chain from the
375/// emitted store
376static std::tuple<SDValue, SDValue, MachineMemOperand*>
377spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
378 SelectionDAGBuilder &Builder) {
379 SDValue Loc = Builder.StatepointLowering.getLocation(Val: Incoming);
380 MachineMemOperand* MMO = nullptr;
381
382 // Emit new store if we didn't do it for this ptr before
383 if (!Loc.getNode()) {
384 Loc = Builder.StatepointLowering.allocateStackSlot(ValueType: Incoming.getValueType(),
385 Builder);
386 int Index = cast<FrameIndexSDNode>(Val&: Loc)->getIndex();
387 // We use TargetFrameIndex so that isel will not select it into LEA
388 Loc = Builder.DAG.getTargetFrameIndex(FI: Index, VT: Builder.getFrameIndexTy());
389
390 // Right now we always allocate spill slots that are of the same
391 // size as the value we're about to spill (the size of spillee can
392 // vary since we spill vectors of pointers too). At some point we
393 // can consider allowing spills of smaller values to larger slots
394 // (i.e. change the '==' in the assert below to a '>=').
395 MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
396 assert((MFI.getObjectSize(Index) * 8) ==
397 (-8 & (7 + // Round up modulo 8.
398 (int64_t)Incoming.getValueSizeInBits())) &&
399 "Bad spill: stack slot does not match!");
400
401 // Note: Using the alignment of the spill slot (rather than the abi or
402 // preferred alignment) is required for correctness when dealing with spill
403 // slots with preferred alignments larger than frame alignment..
404 auto &MF = Builder.DAG.getMachineFunction();
405 auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: Index);
406 auto *StoreMMO = MF.getMachineMemOperand(
407 PtrInfo, F: MachineMemOperand::MOStore, Size: MFI.getObjectSize(ObjectIdx: Index),
408 BaseAlignment: MFI.getObjectAlign(ObjectIdx: Index));
409 Chain = Builder.DAG.getStore(Chain, dl: Builder.getCurSDLoc(), Val: Incoming, Ptr: Loc,
410 MMO: StoreMMO);
411
412 MMO = getMachineMemOperand(MF, FI&: *cast<FrameIndexSDNode>(Val&: Loc));
413
414 Builder.StatepointLowering.setLocation(Val: Incoming, Location: Loc);
415 }
416
417 assert(Loc.getNode());
418 return std::make_tuple(args&: Loc, args&: Chain, args&: MMO);
419}
420
421/// Lower a single value incoming to a statepoint node. This value can be
422/// either a deopt value or a gc value, the handling is the same. We special
423/// case constants and allocas, then fall back to spilling if required.
424static void
425lowerIncomingStatepointValue(SDValue Incoming, bool RequireSpillSlot,
426 SmallVectorImpl<SDValue> &Ops,
427 SmallVectorImpl<MachineMemOperand *> &MemRefs,
428 SelectionDAGBuilder &Builder) {
429
430 if (willLowerDirectly(Incoming)) {
431 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Incoming)) {
432 // This handles allocas as arguments to the statepoint (this is only
433 // really meaningful for a deopt value. For GC, we'd be trying to
434 // relocate the address of the alloca itself?)
435 assert(Incoming.getValueType() == Builder.getFrameIndexTy() &&
436 "Incoming value is a frame index!");
437 Ops.push_back(Elt: Builder.DAG.getTargetFrameIndex(FI: FI->getIndex(),
438 VT: Builder.getFrameIndexTy()));
439
440 auto &MF = Builder.DAG.getMachineFunction();
441 auto *MMO = getMachineMemOperand(MF, FI&: *FI);
442 MemRefs.push_back(Elt: MMO);
443 return;
444 }
445
446 assert(Incoming.getValueType().getSizeInBits() <= 64);
447
448 if (Incoming.isUndef()) {
449 // Put an easily recognized constant that's unlikely to be a valid
450 // value so that uses of undef by the consumer of the stackmap is
451 // easily recognized. This is legal since the compiler is always
452 // allowed to chose an arbitrary value for undef.
453 pushStackMapConstant(Ops, Builder, Value: 0xFEFEFEFE);
454 return;
455 }
456
457 // If the original value was a constant, make sure it gets recorded as
458 // such in the stackmap. This is required so that the consumer can
459 // parse any internal format to the deopt state. It also handles null
460 // pointers and other constant pointers in GC states.
461 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Incoming)) {
462 pushStackMapConstant(Ops, Builder, Value: C->getSExtValue());
463 return;
464 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val&: Incoming)) {
465 pushStackMapConstant(Ops, Builder,
466 Value: C->getValueAPF().bitcastToAPInt().getZExtValue());
467 return;
468 }
469
470 llvm_unreachable("unhandled direct lowering case");
471 }
472
473
474
475 if (!RequireSpillSlot) {
476 // If this value is live in (not live-on-return, or live-through), we can
477 // treat it the same way patchpoint treats it's "live in" values. We'll
478 // end up folding some of these into stack references, but they'll be
479 // handled by the register allocator. Note that we do not have the notion
480 // of a late use so these values might be placed in registers which are
481 // clobbered by the call. This is fine for live-in. For live-through
482 // fix-up pass should be executed to force spilling of such registers.
483 Ops.push_back(Elt: Incoming);
484 } else {
485 // Otherwise, locate a spill slot and explicitly spill it so it can be
486 // found by the runtime later. Note: We know all of these spills are
487 // independent, but don't bother to exploit that chain wise. DAGCombine
488 // will happily do so as needed, so doing it here would be a small compile
489 // time win at most.
490 SDValue Chain = Builder.getRoot();
491 auto Res = spillIncomingStatepointValue(Incoming, Chain, Builder);
492 Ops.push_back(Elt: std::get<0>(t&: Res));
493 if (auto *MMO = std::get<2>(t&: Res))
494 MemRefs.push_back(Elt: MMO);
495 Chain = std::get<1>(t&: Res);
496 Builder.DAG.setRoot(Chain);
497 }
498
499}
500
501/// Return true if value V represents the GC value. The behavior is conservative
502/// in case it is not sure that value is not GC the function returns true.
503static bool isGCValue(const Value *V, SelectionDAGBuilder &Builder) {
504 auto *Ty = V->getType();
505 if (!Ty->isPtrOrPtrVectorTy())
506 return false;
507 if (auto *GFI = Builder.GFI)
508 if (auto IsManaged = GFI->getStrategy().isGCManagedPointer(Ty))
509 return *IsManaged;
510 return true; // conservative
511}
512
513/// Lower deopt state and gc pointer arguments of the statepoint. The actual
514/// lowering is described in lowerIncomingStatepointValue. This function is
515/// responsible for lowering everything in the right position and playing some
516/// tricks to avoid redundant stack manipulation where possible. On
517/// completion, 'Ops' will contain ready to use operands for machine code
518/// statepoint. The chain nodes will have already been created and the DAG root
519/// will be set to the last value spilled (if any were).
520static void
521lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
522 SmallVectorImpl<MachineMemOperand *> &MemRefs,
523 SmallVectorImpl<SDValue> &GCPtrs,
524 DenseMap<SDValue, int> &LowerAsVReg,
525 SelectionDAGBuilder::StatepointLoweringInfo &SI,
526 SelectionDAGBuilder &Builder) {
527 // Lower the deopt and gc arguments for this statepoint. Layout will be:
528 // deopt argument length, deopt arguments.., gc arguments...
529
530 // Figure out what lowering strategy we're going to use for each part
531 // Note: It is conservatively correct to lower both "live-in" and "live-out"
532 // as "live-through". A "live-through" variable is one which is "live-in",
533 // "live-out", and live throughout the lifetime of the call (i.e. we can find
534 // it from any PC within the transitive callee of the statepoint). In
535 // particular, if the callee spills callee preserved registers we may not
536 // be able to find a value placed in that register during the call. This is
537 // fine for live-out, but not for live-through. If we were willing to make
538 // assumptions about the code generator producing the callee, we could
539 // potentially allow live-through values in callee saved registers.
540 const bool LiveInDeopt =
541 SI.StatepointFlags & (uint64_t)StatepointFlags::DeoptLiveIn;
542
543 // Decide which deriver pointers will go on VRegs
544 unsigned MaxVRegPtrs = MaxRegistersForGCPointers.getValue();
545
546 // Pointers used on exceptional path of invoke statepoint.
547 // We cannot assing them to VRegs.
548 SmallSet<SDValue, 8> LPadPointers;
549 if (!UseRegistersForGCPointersInLandingPad)
550 if (const auto *StInvoke =
551 dyn_cast_or_null<InvokeInst>(Val: SI.StatepointInstr)) {
552 LandingPadInst *LPI = StInvoke->getLandingPadInst();
553 for (const auto *Relocate : SI.GCRelocates)
554 if (Relocate->getOperand(i_nocapture: 0) == LPI) {
555 LPadPointers.insert(V: Builder.getValue(V: Relocate->getBasePtr()));
556 LPadPointers.insert(V: Builder.getValue(V: Relocate->getDerivedPtr()));
557 }
558 }
559
560 LLVM_DEBUG(dbgs() << "Deciding how to lower GC Pointers:\n");
561
562 // List of unique lowered GC Pointer values.
563 SmallSetVector<SDValue, 16> LoweredGCPtrs;
564 // Map lowered GC Pointer value to the index in above vector
565 DenseMap<SDValue, unsigned> GCPtrIndexMap;
566
567 unsigned CurNumVRegs = 0;
568
569 auto canPassGCPtrOnVReg = [&](SDValue SD) {
570 if (SD.getValueType().isVector())
571 return false;
572 if (LPadPointers.count(V: SD))
573 return false;
574 return !willLowerDirectly(Incoming: SD);
575 };
576
577 auto processGCPtr = [&](const Value *V) {
578 SDValue PtrSD = Builder.getValue(V);
579 if (!LoweredGCPtrs.insert(X: PtrSD))
580 return; // skip duplicates
581 GCPtrIndexMap[PtrSD] = LoweredGCPtrs.size() - 1;
582
583 assert(!LowerAsVReg.count(PtrSD) && "must not have been seen");
584 if (LowerAsVReg.size() == MaxVRegPtrs)
585 return;
586 assert(V->getType()->isVectorTy() == PtrSD.getValueType().isVector() &&
587 "IR and SD types disagree");
588 if (!canPassGCPtrOnVReg(PtrSD)) {
589 LLVM_DEBUG(dbgs() << "direct/spill "; PtrSD.dump(&Builder.DAG));
590 return;
591 }
592 LLVM_DEBUG(dbgs() << "vreg "; PtrSD.dump(&Builder.DAG));
593 LowerAsVReg[PtrSD] = CurNumVRegs++;
594 };
595
596 // Process derived pointers first to give them more chance to go on VReg.
597 for (const Value *V : SI.Ptrs)
598 processGCPtr(V);
599 for (const Value *V : SI.Bases)
600 processGCPtr(V);
601
602 LLVM_DEBUG(dbgs() << LowerAsVReg.size() << " pointers will go in vregs\n");
603
604 auto requireSpillSlot = [&](const Value *V) {
605 if (!Builder.DAG.getTargetLoweringInfo().isTypeLegal(
606 VT: Builder.getValue(V).getValueType()))
607 return true;
608 if (isGCValue(V, Builder))
609 return !LowerAsVReg.count(Val: Builder.getValue(V));
610 return !(LiveInDeopt || UseRegistersForDeoptValues);
611 };
612
613 // Before we actually start lowering (and allocating spill slots for values),
614 // reserve any stack slots which we judge to be profitable to reuse for a
615 // particular value. This is purely an optimization over the code below and
616 // doesn't change semantics at all. It is important for performance that we
617 // reserve slots for both deopt and gc values before lowering either.
618 for (const Value *V : SI.DeoptState) {
619 if (requireSpillSlot(V))
620 reservePreviousStackSlotForValue(IncomingValue: V, Builder);
621 }
622
623 for (const Value *V : SI.Ptrs) {
624 SDValue SDV = Builder.getValue(V);
625 if (!LowerAsVReg.count(Val: SDV))
626 reservePreviousStackSlotForValue(IncomingValue: V, Builder);
627 }
628
629 for (const Value *V : SI.Bases) {
630 SDValue SDV = Builder.getValue(V);
631 if (!LowerAsVReg.count(Val: SDV))
632 reservePreviousStackSlotForValue(IncomingValue: V, Builder);
633 }
634
635 // First, prefix the list with the number of unique values to be
636 // lowered. Note that this is the number of *Values* not the
637 // number of SDValues required to lower them.
638 const int NumVMSArgs = SI.DeoptState.size();
639 pushStackMapConstant(Ops, Builder, Value: NumVMSArgs);
640
641 // The vm state arguments are lowered in an opaque manner. We do not know
642 // what type of values are contained within.
643 LLVM_DEBUG(dbgs() << "Lowering deopt state\n");
644 for (const Value *V : SI.DeoptState) {
645 SDValue Incoming;
646 // If this is a function argument at a static frame index, generate it as
647 // the frame index.
648 if (const Argument *Arg = dyn_cast<Argument>(Val: V)) {
649 int FI = Builder.FuncInfo.getArgumentFrameIndex(A: Arg);
650 if (FI != INT_MAX)
651 Incoming = Builder.DAG.getFrameIndex(FI, VT: Builder.getFrameIndexTy());
652 }
653 if (!Incoming.getNode())
654 Incoming = Builder.getValue(V);
655 LLVM_DEBUG(dbgs() << "Value " << *V
656 << " requireSpillSlot = " << requireSpillSlot(V) << "\n");
657 lowerIncomingStatepointValue(Incoming, RequireSpillSlot: requireSpillSlot(V), Ops, MemRefs,
658 Builder);
659 }
660
661 // Finally, go ahead and lower all the gc arguments.
662 pushStackMapConstant(Ops, Builder, Value: LoweredGCPtrs.size());
663 for (SDValue SDV : LoweredGCPtrs)
664 lowerIncomingStatepointValue(Incoming: SDV, RequireSpillSlot: !LowerAsVReg.count(Val: SDV), Ops, MemRefs,
665 Builder);
666
667 // Copy to out vector. LoweredGCPtrs will be empty after this point.
668 GCPtrs = LoweredGCPtrs.takeVector();
669
670 // If there are any explicit spill slots passed to the statepoint, record
671 // them, but otherwise do not do anything special. These are user provided
672 // allocas and give control over placement to the consumer. In this case,
673 // it is the contents of the slot which may get updated, not the pointer to
674 // the alloca
675 SmallVector<SDValue, 4> Allocas;
676 for (Value *V : SI.GCArgs) {
677 SDValue Incoming = Builder.getValue(V);
678 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Incoming)) {
679 // This handles allocas as arguments to the statepoint
680 assert(Incoming.getValueType() == Builder.getFrameIndexTy() &&
681 "Incoming value is a frame index!");
682 Allocas.push_back(Elt: Builder.DAG.getTargetFrameIndex(
683 FI: FI->getIndex(), VT: Builder.getFrameIndexTy()));
684
685 auto &MF = Builder.DAG.getMachineFunction();
686 auto *MMO = getMachineMemOperand(MF, FI&: *FI);
687 MemRefs.push_back(Elt: MMO);
688 }
689 }
690 pushStackMapConstant(Ops, Builder, Value: Allocas.size());
691 Ops.append(in_start: Allocas.begin(), in_end: Allocas.end());
692
693 // Now construct GC base/derived map;
694 pushStackMapConstant(Ops, Builder, Value: SI.Ptrs.size());
695 SDLoc L = Builder.getCurSDLoc();
696 for (unsigned i = 0; i < SI.Ptrs.size(); ++i) {
697 SDValue Base = Builder.getValue(V: SI.Bases[i]);
698 assert(GCPtrIndexMap.count(Base) && "base not found in index map");
699 Ops.push_back(
700 Elt: Builder.DAG.getTargetConstant(Val: GCPtrIndexMap[Base], DL: L, VT: MVT::i64));
701 SDValue Derived = Builder.getValue(V: SI.Ptrs[i]);
702 assert(GCPtrIndexMap.count(Derived) && "derived not found in index map");
703 Ops.push_back(
704 Elt: Builder.DAG.getTargetConstant(Val: GCPtrIndexMap[Derived], DL: L, VT: MVT::i64));
705 }
706}
707
708SDValue SelectionDAGBuilder::LowerAsSTATEPOINT(
709 SelectionDAGBuilder::StatepointLoweringInfo &SI) {
710 // The basic scheme here is that information about both the original call and
711 // the safepoint is encoded in the CallInst. We create a temporary call and
712 // lower it, then reverse engineer the calling sequence.
713
714 NumOfStatepoints++;
715 // Clear state
716 StatepointLowering.startNewStatepoint(Builder&: *this);
717 assert(SI.Bases.size() == SI.Ptrs.size() && "Pointer without base!");
718 assert((GFI || SI.Bases.empty()) &&
719 "No gc specified, so cannot relocate pointers!");
720
721 LLVM_DEBUG(if (SI.StatepointInstr) dbgs()
722 << "Lowering statepoint " << *SI.StatepointInstr << "\n");
723#ifndef NDEBUG
724 for (const auto *Reloc : SI.GCRelocates)
725 if (Reloc->getParent() == SI.StatepointInstr->getParent())
726 StatepointLowering.scheduleRelocCall(*Reloc);
727#endif
728
729 // Lower statepoint vmstate and gcstate arguments
730
731 // All lowered meta args.
732 SmallVector<SDValue, 10> LoweredMetaArgs;
733 // Lowered GC pointers (subset of above).
734 SmallVector<SDValue, 16> LoweredGCArgs;
735 SmallVector<MachineMemOperand*, 16> MemRefs;
736 // Maps derived pointer SDValue to statepoint result of relocated pointer.
737 DenseMap<SDValue, int> LowerAsVReg;
738 lowerStatepointMetaArgs(Ops&: LoweredMetaArgs, MemRefs, GCPtrs&: LoweredGCArgs, LowerAsVReg,
739 SI, Builder&: *this);
740
741 // Now that we've emitted the spills, we need to update the root so that the
742 // call sequence is ordered correctly.
743 SI.CLI.setChain(getRoot());
744
745 // Get call node, we will replace it later with statepoint
746 SDValue ReturnVal;
747 SDNode *CallNode;
748 std::tie(args&: ReturnVal, args&: CallNode) = lowerCallFromStatepointLoweringInfo(SI, Builder&: *this);
749
750 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
751 // nodes with all the appropriate arguments and return values.
752
753 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
754 SDValue Chain = CallNode->getOperand(Num: 0);
755
756 SDValue Glue;
757 bool CallHasIncomingGlue = CallNode->getGluedNode();
758 if (CallHasIncomingGlue) {
759 // Glue is always last operand
760 Glue = CallNode->getOperand(Num: CallNode->getNumOperands() - 1);
761 }
762
763 // Build the GC_TRANSITION_START node if necessary.
764 //
765 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
766 // order in which they appear in the call to the statepoint intrinsic. If
767 // any of the operands is a pointer-typed, that operand is immediately
768 // followed by a SRCVALUE for the pointer that may be used during lowering
769 // (e.g. to form MachinePointerInfo values for loads/stores).
770 const bool IsGCTransition =
771 (SI.StatepointFlags & (uint64_t)StatepointFlags::GCTransition) ==
772 (uint64_t)StatepointFlags::GCTransition;
773 if (IsGCTransition) {
774 SmallVector<SDValue, 8> TSOps;
775
776 // Add chain
777 TSOps.push_back(Elt: Chain);
778
779 // Add GC transition arguments
780 for (const Value *V : SI.GCTransitionArgs) {
781 TSOps.push_back(Elt: getValue(V));
782 if (V->getType()->isPointerTy())
783 TSOps.push_back(Elt: DAG.getSrcValue(v: V));
784 }
785
786 // Add glue if necessary
787 if (CallHasIncomingGlue)
788 TSOps.push_back(Elt: Glue);
789
790 SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue);
791
792 SDValue GCTransitionStart =
793 DAG.getNode(Opcode: ISD::GC_TRANSITION_START, DL: getCurSDLoc(), VTList: NodeTys, Ops: TSOps);
794
795 Chain = GCTransitionStart.getValue(R: 0);
796 Glue = GCTransitionStart.getValue(R: 1);
797 }
798
799 // TODO: Currently, all of these operands are being marked as read/write in
800 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
801 // and flags to be read-only.
802 SmallVector<SDValue, 40> Ops;
803
804 // Add the <id> and <numBytes> constants.
805 Ops.push_back(Elt: DAG.getTargetConstant(Val: SI.ID, DL: getCurSDLoc(), VT: MVT::i64));
806 Ops.push_back(
807 Elt: DAG.getTargetConstant(Val: SI.NumPatchBytes, DL: getCurSDLoc(), VT: MVT::i32));
808
809 // Calculate and push starting position of vmstate arguments
810 // Get number of arguments incoming directly into call node
811 unsigned NumCallRegArgs =
812 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
813 Ops.push_back(Elt: DAG.getTargetConstant(Val: NumCallRegArgs, DL: getCurSDLoc(), VT: MVT::i32));
814
815 // Add call target
816 SDValue CallTarget = SDValue(CallNode->getOperand(Num: 1).getNode(), 0);
817 Ops.push_back(Elt: CallTarget);
818
819 // Add call arguments
820 // Get position of register mask in the call
821 SDNode::op_iterator RegMaskIt;
822 if (CallHasIncomingGlue)
823 RegMaskIt = CallNode->op_end() - 2;
824 else
825 RegMaskIt = CallNode->op_end() - 1;
826 Ops.insert(I: Ops.end(), From: CallNode->op_begin() + 2, To: RegMaskIt);
827
828 // Add a constant argument for the calling convention
829 pushStackMapConstant(Ops, Builder&: *this, Value: SI.CLI.CallConv);
830
831 // Add a constant argument for the flags
832 uint64_t Flags = SI.StatepointFlags;
833 assert(((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0) &&
834 "Unknown flag used");
835 pushStackMapConstant(Ops, Builder&: *this, Value: Flags);
836
837 // Insert all vmstate and gcstate arguments
838 llvm::append_range(C&: Ops, R&: LoweredMetaArgs);
839
840 // Add register mask from call node
841 Ops.push_back(Elt: *RegMaskIt);
842
843 // Add chain
844 Ops.push_back(Elt: Chain);
845
846 // Same for the glue, but we add it only if original call had it
847 if (Glue.getNode())
848 Ops.push_back(Elt: Glue);
849
850 // Compute return values. Provide a glue output since we consume one as
851 // input. This allows someone else to chain off us as needed.
852 SmallVector<EVT, 8> NodeTys;
853 for (auto SD : LoweredGCArgs) {
854 if (!LowerAsVReg.count(Val: SD))
855 continue;
856 NodeTys.push_back(Elt: SD.getValueType());
857 }
858 LLVM_DEBUG(dbgs() << "Statepoint has " << NodeTys.size() << " results\n");
859 assert(NodeTys.size() == LowerAsVReg.size() && "Inconsistent GC Ptr lowering");
860 NodeTys.push_back(Elt: MVT::Other);
861 NodeTys.push_back(Elt: MVT::Glue);
862
863 unsigned NumResults = NodeTys.size();
864 MachineSDNode *StatepointMCNode =
865 DAG.getMachineNode(Opcode: TargetOpcode::STATEPOINT, dl: getCurSDLoc(), ResultTys: NodeTys, Ops);
866 DAG.setNodeMemRefs(N: StatepointMCNode, NewMemRefs: MemRefs);
867
868 // For values lowered to tied-defs, create the virtual registers if used
869 // in other blocks. For local gc.relocate record appropriate statepoint
870 // result in StatepointLoweringState.
871 DenseMap<SDValue, Register> VirtRegs;
872 for (const auto *Relocate : SI.GCRelocates) {
873 Value *Derived = Relocate->getDerivedPtr();
874 SDValue SD = getValue(V: Derived);
875 if (!LowerAsVReg.count(Val: SD))
876 continue;
877
878 SDValue Relocated = SDValue(StatepointMCNode, LowerAsVReg[SD]);
879
880 // Handle local relocate. Note that different relocates might
881 // map to the same SDValue.
882 if (SI.StatepointInstr->getParent() == Relocate->getParent()) {
883 SDValue Res = StatepointLowering.getLocation(Val: SD);
884 if (Res)
885 assert(Res == Relocated);
886 else
887 StatepointLowering.setLocation(Val: SD, Location: Relocated);
888 continue;
889 }
890
891 // Handle multiple gc.relocates of the same input efficiently.
892 if (VirtRegs.count(Val: SD))
893 continue;
894
895 auto *RetTy = Relocate->getType();
896 Register Reg = FuncInfo.CreateRegs(Ty: RetTy);
897 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
898 DAG.getDataLayout(), Reg, RetTy, std::nullopt);
899 SDValue Chain = DAG.getRoot();
900 RFV.getCopyToRegs(Val: Relocated, DAG, dl: getCurSDLoc(), Chain, Glue: nullptr);
901 PendingExports.push_back(Elt: Chain);
902
903 VirtRegs[SD] = Reg;
904 }
905
906 // Record for later use how each relocation was lowered. This is needed to
907 // allow later gc.relocates to mirror the lowering chosen.
908 const Instruction *StatepointInstr = SI.StatepointInstr;
909 auto &RelocationMap = FuncInfo.StatepointRelocationMaps[StatepointInstr];
910 for (const GCRelocateInst *Relocate : SI.GCRelocates) {
911 const Value *V = Relocate->getDerivedPtr();
912 SDValue SDV = getValue(V);
913 SDValue Loc = StatepointLowering.getLocation(Val: SDV);
914
915 bool IsLocal = (Relocate->getParent() == StatepointInstr->getParent());
916
917 RecordType Record;
918 if (IsLocal && LowerAsVReg.count(Val: SDV)) {
919 // Result is already stored in StatepointLowering
920 Record.type = RecordType::SDValueNode;
921 } else if (LowerAsVReg.count(Val: SDV)) {
922 Record.type = RecordType::VReg;
923 assert(VirtRegs.count(SDV));
924 Record.payload.Reg = VirtRegs[SDV];
925 } else if (Loc.getNode()) {
926 Record.type = RecordType::Spill;
927 Record.payload.FI = cast<FrameIndexSDNode>(Val&: Loc)->getIndex();
928 } else {
929 Record.type = RecordType::NoRelocate;
930 // If we didn't relocate a value, we'll essentialy end up inserting an
931 // additional use of the original value when lowering the gc.relocate.
932 // We need to make sure the value is available at the new use, which
933 // might be in another block.
934 if (Relocate->getParent() != StatepointInstr->getParent())
935 ExportFromCurrentBlock(V);
936 }
937 RelocationMap[Relocate] = Record;
938 }
939
940
941
942 SDNode *SinkNode = StatepointMCNode;
943
944 // Build the GC_TRANSITION_END node if necessary.
945 //
946 // See the comment above regarding GC_TRANSITION_START for the layout of
947 // the operands to the GC_TRANSITION_END node.
948 if (IsGCTransition) {
949 SmallVector<SDValue, 8> TEOps;
950
951 // Add chain
952 TEOps.push_back(Elt: SDValue(StatepointMCNode, NumResults - 2));
953
954 // Add GC transition arguments
955 for (const Value *V : SI.GCTransitionArgs) {
956 TEOps.push_back(Elt: getValue(V));
957 if (V->getType()->isPointerTy())
958 TEOps.push_back(Elt: DAG.getSrcValue(v: V));
959 }
960
961 // Add glue
962 TEOps.push_back(Elt: SDValue(StatepointMCNode, NumResults - 1));
963
964 SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue);
965
966 SDValue GCTransitionStart =
967 DAG.getNode(Opcode: ISD::GC_TRANSITION_END, DL: getCurSDLoc(), VTList: NodeTys, Ops: TEOps);
968
969 SinkNode = GCTransitionStart.getNode();
970 }
971
972 // Replace original call
973 // Call: ch,glue = CALL ...
974 // Statepoint: [gc relocates],ch,glue = STATEPOINT ...
975 unsigned NumSinkValues = SinkNode->getNumValues();
976 SDValue StatepointValues[2] = {SDValue(SinkNode, NumSinkValues - 2),
977 SDValue(SinkNode, NumSinkValues - 1)};
978 DAG.ReplaceAllUsesWith(From: CallNode, To: StatepointValues);
979 // Remove original call node
980 DAG.DeleteNode(N: CallNode);
981
982 // Since we always emit CopyToRegs (even for local relocates), we must
983 // update root, so that they are emitted before any local uses.
984 (void)getControlRoot();
985
986 // TODO: A better future implementation would be to emit a single variable
987 // argument, variable return value STATEPOINT node here and then hookup the
988 // return value of each gc.relocate to the respective output of the
989 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
990 // to actually be possible today.
991
992 return ReturnVal;
993}
994
995/// Return two gc.results if present. First result is a block local
996/// gc.result, second result is a non-block local gc.result. Corresponding
997/// entry will be nullptr if not present.
998static std::pair<const GCResultInst*, const GCResultInst*>
999getGCResultLocality(const GCStatepointInst &S) {
1000 std::pair<const GCResultInst *, const GCResultInst*> Res(nullptr, nullptr);
1001 for (const auto *U : S.users()) {
1002 auto *GRI = dyn_cast<GCResultInst>(Val: U);
1003 if (!GRI)
1004 continue;
1005 if (GRI->getParent() == S.getParent())
1006 Res.first = GRI;
1007 else
1008 Res.second = GRI;
1009 }
1010 return Res;
1011}
1012
1013void
1014SelectionDAGBuilder::LowerStatepoint(const GCStatepointInst &I,
1015 const BasicBlock *EHPadBB /*= nullptr*/) {
1016 assert(I.getCallingConv() != CallingConv::AnyReg &&
1017 "anyregcc is not supported on statepoints!");
1018
1019#ifndef NDEBUG
1020 // Check that the associated GCStrategy expects to encounter statepoints.
1021 assert(GFI->getStrategy().useStatepoints() &&
1022 "GCStrategy does not expect to encounter statepoints");
1023#endif
1024
1025 SDValue ActualCallee;
1026 SDValue Callee = getValue(V: I.getActualCalledOperand());
1027
1028 if (I.getNumPatchBytes() > 0) {
1029 // If we've been asked to emit a nop sequence instead of a call instruction
1030 // for this statepoint then don't lower the call target, but use a constant
1031 // `undef` instead. Not lowering the call target lets statepoint clients
1032 // get away without providing a physical address for the symbolic call
1033 // target at link time.
1034 ActualCallee = DAG.getUNDEF(VT: Callee.getValueType());
1035 } else {
1036 ActualCallee = Callee;
1037 }
1038
1039 const auto GCResultLocality = getGCResultLocality(S: I);
1040 AttributeSet retAttrs;
1041 if (GCResultLocality.first)
1042 retAttrs = GCResultLocality.first->getAttributes().getRetAttrs();
1043
1044 StatepointLoweringInfo SI(DAG);
1045 populateCallLoweringInfo(CLI&: SI.CLI, Call: &I, ArgIdx: GCStatepointInst::CallArgsBeginPos,
1046 NumArgs: I.getNumCallArgs(), Callee: ActualCallee,
1047 ReturnTy: I.getActualReturnType(), RetAttrs: retAttrs,
1048 /*IsPatchPoint=*/false);
1049
1050 // There may be duplication in the gc.relocate list; such as two copies of
1051 // each relocation on normal and exceptional path for an invoke. We only
1052 // need to spill once and record one copy in the stackmap, but we need to
1053 // reload once per gc.relocate. (Dedupping gc.relocates is trickier and best
1054 // handled as a CSE problem elsewhere.)
1055 // TODO: There a couple of major stackmap size optimizations we could do
1056 // here if we wished.
1057 // 1) If we've encountered a derived pair {B, D}, we don't need to actually
1058 // record {B,B} if it's seen later.
1059 // 2) Due to rematerialization, actual derived pointers are somewhat rare;
1060 // given that, we could change the format to record base pointer relocations
1061 // separately with half the space. This would require a format rev and a
1062 // fairly major rework of the STATEPOINT node though.
1063 SmallSet<SDValue, 8> Seen;
1064 for (const GCRelocateInst *Relocate : I.getGCRelocates()) {
1065 SI.GCRelocates.push_back(Elt: Relocate);
1066
1067 SDValue DerivedSD = getValue(V: Relocate->getDerivedPtr());
1068 if (Seen.insert(V: DerivedSD).second) {
1069 SI.Bases.push_back(Elt: Relocate->getBasePtr());
1070 SI.Ptrs.push_back(Elt: Relocate->getDerivedPtr());
1071 }
1072 }
1073
1074 // If we find a deopt value which isn't explicitly added, we need to
1075 // ensure it gets lowered such that gc cycles occurring before the
1076 // deoptimization event during the lifetime of the call don't invalidate
1077 // the pointer we're deopting with. Note that we assume that all
1078 // pointers passed to deopt are base pointers; relaxing that assumption
1079 // would require relatively large changes to how we represent relocations.
1080 for (Value *V : I.deopt_operands()) {
1081 if (!isGCValue(V, Builder&: *this))
1082 continue;
1083 if (Seen.insert(V: getValue(V)).second) {
1084 SI.Bases.push_back(Elt: V);
1085 SI.Ptrs.push_back(Elt: V);
1086 }
1087 }
1088
1089 SI.GCArgs = ArrayRef<const Use>(I.gc_args_begin(), I.gc_args_end());
1090 SI.StatepointInstr = &I;
1091 SI.ID = I.getID();
1092
1093 SI.DeoptState = ArrayRef<const Use>(I.deopt_begin(), I.deopt_end());
1094 SI.GCTransitionArgs = ArrayRef<const Use>(I.gc_transition_args_begin(),
1095 I.gc_transition_args_end());
1096
1097 SI.StatepointFlags = I.getFlags();
1098 SI.NumPatchBytes = I.getNumPatchBytes();
1099 SI.EHPadBB = EHPadBB;
1100
1101 SDValue ReturnValue = LowerAsSTATEPOINT(SI);
1102
1103 // Export the result value if needed
1104 if (!GCResultLocality.first && !GCResultLocality.second) {
1105 // The return value is not needed, just generate a poison value.
1106 // Note: This covers the void return case.
1107 setValue(V: &I, NewN: DAG.getIntPtrConstant(Val: -1, DL: getCurSDLoc()));
1108 return;
1109 }
1110
1111 if (GCResultLocality.first) {
1112 // Result value will be used in a same basic block. Don't export it or
1113 // perform any explicit register copies. The gc_result will simply grab
1114 // this value.
1115 setValue(V: &I, NewN: ReturnValue);
1116 }
1117
1118 if (!GCResultLocality.second)
1119 return;
1120 // Result value will be used in a different basic block so we need to export
1121 // it now. Default exporting mechanism will not work here because statepoint
1122 // call has a different type than the actual call. It means that by default
1123 // llvm will create export register of the wrong type (always i32 in our
1124 // case). So instead we need to create export register with correct type
1125 // manually.
1126 // TODO: To eliminate this problem we can remove gc.result intrinsics
1127 // completely and make statepoint call to return a tuple.
1128 Type *RetTy = GCResultLocality.second->getType();
1129 Register Reg = FuncInfo.CreateRegs(Ty: RetTy);
1130 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
1131 DAG.getDataLayout(), Reg, RetTy,
1132 I.getCallingConv());
1133 SDValue Chain = DAG.getEntryNode();
1134
1135 RFV.getCopyToRegs(Val: ReturnValue, DAG, dl: getCurSDLoc(), Chain, Glue: nullptr);
1136 PendingExports.push_back(Elt: Chain);
1137 FuncInfo.ValueMap[&I] = Reg;
1138}
1139
1140void SelectionDAGBuilder::LowerCallSiteWithDeoptBundleImpl(
1141 const CallBase *Call, SDValue Callee, const BasicBlock *EHPadBB,
1142 bool VarArgDisallowed, bool ForceVoidReturnTy) {
1143 StatepointLoweringInfo SI(DAG);
1144 unsigned ArgBeginIndex = Call->arg_begin() - Call->op_begin();
1145 populateCallLoweringInfo(
1146 CLI&: SI.CLI, Call, ArgIdx: ArgBeginIndex, NumArgs: Call->arg_size(), Callee,
1147 ReturnTy: ForceVoidReturnTy ? Type::getVoidTy(C&: *DAG.getContext()) : Call->getType(),
1148 RetAttrs: Call->getAttributes().getRetAttrs(), /*IsPatchPoint=*/false);
1149 if (!VarArgDisallowed)
1150 SI.CLI.IsVarArg = Call->getFunctionType()->isVarArg();
1151
1152 auto DeoptBundle = *Call->getOperandBundle(ID: LLVMContext::OB_deopt);
1153
1154 unsigned DefaultID = StatepointDirectives::DeoptBundleStatepointID;
1155
1156 auto SD = parseStatepointDirectivesFromAttrs(AS: Call->getAttributes());
1157 SI.ID = SD.StatepointID.value_or(u&: DefaultID);
1158 SI.NumPatchBytes = SD.NumPatchBytes.value_or(u: 0);
1159
1160 SI.DeoptState =
1161 ArrayRef<const Use>(DeoptBundle.Inputs.begin(), DeoptBundle.Inputs.end());
1162 SI.StatepointFlags = static_cast<uint64_t>(StatepointFlags::None);
1163 SI.EHPadBB = EHPadBB;
1164
1165 // NB! The GC arguments are deliberately left empty.
1166
1167 LLVM_DEBUG(dbgs() << "Lowering call with deopt bundle " << *Call << "\n");
1168 if (SDValue ReturnVal = LowerAsSTATEPOINT(SI)) {
1169 ReturnVal = lowerRangeToAssertZExt(DAG, I: *Call, Op: ReturnVal);
1170 setValue(V: Call, NewN: ReturnVal);
1171 }
1172}
1173
1174void SelectionDAGBuilder::LowerCallSiteWithDeoptBundle(
1175 const CallBase *Call, SDValue Callee, const BasicBlock *EHPadBB) {
1176 LowerCallSiteWithDeoptBundleImpl(Call, Callee, EHPadBB,
1177 /* VarArgDisallowed = */ false,
1178 /* ForceVoidReturnTy = */ false);
1179}
1180
1181void SelectionDAGBuilder::visitGCResult(const GCResultInst &CI) {
1182 // The result value of the gc_result is simply the result of the actual
1183 // call. We've already emitted this, so just grab the value.
1184 const Value *SI = CI.getStatepoint();
1185 assert((isa<GCStatepointInst>(SI) || isa<UndefValue>(SI)) &&
1186 "GetStatepoint must return one of two types");
1187 if (isa<UndefValue>(Val: SI))
1188 return;
1189
1190 if (cast<GCStatepointInst>(Val: SI)->getParent() == CI.getParent()) {
1191 setValue(V: &CI, NewN: getValue(V: SI));
1192 return;
1193 }
1194 // Statepoint is in different basic block so we should have stored call
1195 // result in a virtual register.
1196 // We can not use default getValue() functionality to copy value from this
1197 // register because statepoint and actual call return types can be
1198 // different, and getValue() will use CopyFromReg of the wrong type,
1199 // which is always i32 in our case.
1200 Type *RetTy = CI.getType();
1201 SDValue CopyFromReg = getCopyFromRegs(V: SI, Ty: RetTy);
1202
1203 assert(CopyFromReg.getNode());
1204 setValue(V: &CI, NewN: CopyFromReg);
1205}
1206
1207void SelectionDAGBuilder::visitGCRelocate(const GCRelocateInst &Relocate) {
1208 const Value *Statepoint = Relocate.getStatepoint();
1209#ifndef NDEBUG
1210 // Consistency check
1211 // We skip this check for relocates not in the same basic block as their
1212 // statepoint. It would be too expensive to preserve validation info through
1213 // different basic blocks.
1214 assert((isa<GCStatepointInst>(Statepoint) || isa<UndefValue>(Statepoint)) &&
1215 "GetStatepoint must return one of two types");
1216 if (isa<UndefValue>(Statepoint))
1217 return;
1218
1219 if (cast<GCStatepointInst>(Statepoint)->getParent() == Relocate.getParent())
1220 StatepointLowering.relocCallVisited(Relocate);
1221#endif
1222
1223 const Value *DerivedPtr = Relocate.getDerivedPtr();
1224 auto &RelocationMap =
1225 FuncInfo.StatepointRelocationMaps[cast<GCStatepointInst>(Val: Statepoint)];
1226 auto SlotIt = RelocationMap.find(Val: &Relocate);
1227 assert(SlotIt != RelocationMap.end() && "Relocating not lowered gc value");
1228 const RecordType &Record = SlotIt->second;
1229
1230 // If relocation was done via virtual register..
1231 if (Record.type == RecordType::SDValueNode) {
1232 assert(cast<GCStatepointInst>(Statepoint)->getParent() ==
1233 Relocate.getParent() &&
1234 "Nonlocal gc.relocate mapped via SDValue");
1235 SDValue SDV = StatepointLowering.getLocation(Val: getValue(V: DerivedPtr));
1236 assert(SDV.getNode() && "empty SDValue");
1237 setValue(V: &Relocate, NewN: SDV);
1238 return;
1239 }
1240 if (Record.type == RecordType::VReg) {
1241 Register InReg = Record.payload.Reg;
1242 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
1243 DAG.getDataLayout(), InReg, Relocate.getType(),
1244 std::nullopt); // This is not an ABI copy.
1245 // We generate copy to/from regs even for local uses, hence we must
1246 // chain with current root to ensure proper ordering of copies w.r.t.
1247 // statepoint.
1248 SDValue Chain = DAG.getRoot();
1249 SDValue Relocation = RFV.getCopyFromRegs(DAG, FuncInfo, dl: getCurSDLoc(),
1250 Chain, Glue: nullptr, V: nullptr);
1251 setValue(V: &Relocate, NewN: Relocation);
1252 return;
1253 }
1254
1255 if (Record.type == RecordType::Spill) {
1256 unsigned Index = Record.payload.FI;
1257 SDValue SpillSlot = DAG.getTargetFrameIndex(FI: Index, VT: getFrameIndexTy());
1258
1259 // All the reloads are independent and are reading memory only modified by
1260 // statepoints (i.e. no other aliasing stores); informing SelectionDAG of
1261 // this lets CSE kick in for free and allows reordering of
1262 // instructions if possible. The lowering for statepoint sets the root,
1263 // so this is ordering all reloads with the either
1264 // a) the statepoint node itself, or
1265 // b) the entry of the current block for an invoke statepoint.
1266 const SDValue Chain = DAG.getRoot(); // != Builder.getRoot()
1267
1268 auto &MF = DAG.getMachineFunction();
1269 auto &MFI = MF.getFrameInfo();
1270 auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: Index);
1271 auto *LoadMMO = MF.getMachineMemOperand(PtrInfo, F: MachineMemOperand::MOLoad,
1272 Size: MFI.getObjectSize(ObjectIdx: Index),
1273 BaseAlignment: MFI.getObjectAlign(ObjectIdx: Index));
1274
1275 auto LoadVT = DAG.getTargetLoweringInfo().getValueType(DL: DAG.getDataLayout(),
1276 Ty: Relocate.getType());
1277
1278 SDValue SpillLoad =
1279 DAG.getLoad(VT: LoadVT, dl: getCurSDLoc(), Chain, Ptr: SpillSlot, MMO: LoadMMO);
1280 PendingLoads.push_back(Elt: SpillLoad.getValue(R: 1));
1281
1282 assert(SpillLoad.getNode());
1283 setValue(V: &Relocate, NewN: SpillLoad);
1284 return;
1285 }
1286
1287 assert(Record.type == RecordType::NoRelocate);
1288 SDValue SD = getValue(V: DerivedPtr);
1289
1290 if (SD.isUndef() && SD.getValueType().getSizeInBits() <= 64) {
1291 // Lowering relocate(undef) as arbitrary constant. Current constant value
1292 // is chosen such that it's unlikely to be a valid pointer.
1293 setValue(V: &Relocate, NewN: DAG.getConstant(Val: 0xFEFEFEFE, DL: SDLoc(SD), VT: MVT::i64));
1294 return;
1295 }
1296
1297 // We didn't need to spill these special cases (constants and allocas).
1298 // See the handling in spillIncomingValueForStatepoint for detail.
1299 setValue(V: &Relocate, NewN: SD);
1300}
1301
1302void SelectionDAGBuilder::LowerDeoptimizeCall(const CallInst *CI) {
1303 const auto &TLI = DAG.getTargetLoweringInfo();
1304 SDValue Callee = DAG.getExternalSymbol(Sym: TLI.getLibcallName(Call: RTLIB::DEOPTIMIZE),
1305 VT: TLI.getPointerTy(DL: DAG.getDataLayout()));
1306
1307 // We don't lower calls to __llvm_deoptimize as varargs, but as a regular
1308 // call. We also do not lower the return value to any virtual register, and
1309 // change the immediately following return to a trap instruction.
1310 LowerCallSiteWithDeoptBundleImpl(Call: CI, Callee, /* EHPadBB = */ nullptr,
1311 /* VarArgDisallowed = */ true,
1312 /* ForceVoidReturnTy = */ true);
1313}
1314
1315void SelectionDAGBuilder::LowerDeoptimizingReturn() {
1316 // We do not lower the return value from llvm.deoptimize to any virtual
1317 // register, and change the immediately following return to a trap
1318 // instruction.
1319 if (DAG.getTarget().Options.TrapUnreachable)
1320 DAG.setRoot(
1321 DAG.getNode(Opcode: ISD::TRAP, DL: getCurSDLoc(), VT: MVT::Other, Operand: DAG.getRoot()));
1322}
1323