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/SelectionDAG.h"
30#include "llvm/CodeGen/SelectionDAGNodes.h"
31#include "llvm/CodeGen/StackMaps.h"
32#include "llvm/CodeGen/TargetLowering.h"
33#include "llvm/CodeGen/TargetOpcodes.h"
34#include "llvm/CodeGenTypes/MachineValueType.h"
35#include "llvm/IR/CallingConv.h"
36#include "llvm/IR/DerivedTypes.h"
37#include "llvm/IR/GCStrategy.h"
38#include "llvm/IR/Instruction.h"
39#include "llvm/IR/Instructions.h"
40#include "llvm/IR/LLVMContext.h"
41#include "llvm/IR/Statepoint.h"
42#include "llvm/IR/Type.h"
43#include "llvm/Support/Casting.h"
44#include "llvm/Support/CommandLine.h"
45#include "llvm/Target/TargetMachine.h"
46#include "llvm/Target/TargetOptions.h"
47#include <cassert>
48#include <cstddef>
49#include <cstdint>
50#include <iterator>
51#include <tuple>
52#include <utility>
53
54using namespace llvm;
55
56#define DEBUG_TYPE "statepoint-lowering"
57
58STATISTIC(NumSlotsAllocatedForStatepoints,
59 "Number of stack slots allocated for statepoints");
60STATISTIC(NumOfStatepoints, "Number of statepoint nodes encountered");
61STATISTIC(StatepointMaxSlotsRequired,
62 "Maximum number of stack slots required for a singe statepoint");
63
64static cl::opt<bool> UseRegistersForDeoptValues(
65 "use-registers-for-deopt-values", cl::Hidden, cl::init(Val: false),
66 cl::desc("Allow using registers for non pointer deopt args"));
67
68static cl::opt<bool> UseRegistersForGCPointersInLandingPad(
69 "use-registers-for-gc-values-in-landing-pad", cl::Hidden, cl::init(Val: false),
70 cl::desc("Allow using registers for gc pointer in landing pad"));
71
72static cl::opt<unsigned> MaxRegistersForGCPointers(
73 "max-registers-for-gc-values", cl::Hidden, cl::init(Val: 0),
74 cl::desc("Max number of VRegs allowed to pass GC pointer meta args in"));
75
76typedef FunctionLoweringInfo::StatepointRelocationRecord RecordType;
77
78static void pushStackMapConstant(SmallVectorImpl<SDValue>& Ops,
79 SelectionDAGBuilder &Builder, uint64_t Value) {
80 SDLoc L = Builder.getCurSDLoc();
81 Ops.push_back(Elt: Builder.DAG.getTargetConstant(Val: StackMaps::ConstantOp, DL: L,
82 VT: MVT::i64));
83 Ops.push_back(Elt: Builder.DAG.getTargetConstant(Val: Value, DL: L, VT: MVT::i64));
84}
85
86void StatepointLoweringState::startNewStatepoint(SelectionDAGBuilder &Builder) {
87 // Consistency check
88 assert(PendingGCRelocateCalls.empty() &&
89 "Trying to visit statepoint before finished processing previous one");
90 Locations.clear();
91 NextSlotToAllocate = 0;
92 // Need to resize this on each safepoint - we need the two to stay in sync and
93 // the clear patterns of a SelectionDAGBuilder have no relation to
94 // FunctionLoweringInfo. Also need to ensure used bits get cleared.
95 AllocatedStackSlots.clear();
96 AllocatedStackSlots.resize(N: Builder.FuncInfo.StatepointStackSlots.size());
97}
98
99void StatepointLoweringState::clear() {
100 Locations.clear();
101 AllocatedStackSlots.clear();
102 assert(PendingGCRelocateCalls.empty() &&
103 "cleared before statepoint sequence completed");
104}
105
106SDValue
107StatepointLoweringState::allocateStackSlot(EVT ValueType,
108 SelectionDAGBuilder &Builder) {
109 NumSlotsAllocatedForStatepoints++;
110 MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
111
112 unsigned SpillSize = ValueType.getStoreSize();
113 assert((SpillSize * 8) ==
114 (-8u & (7 + ValueType.getSizeInBits())) && // Round up modulo 8.
115 "Size not in bytes?");
116
117 // First look for a previously created stack slot which is not in
118 // use (accounting for the fact arbitrary slots may already be
119 // reserved), or to create a new stack slot and use it.
120
121 const size_t NumSlots = AllocatedStackSlots.size();
122 assert(NextSlotToAllocate <= NumSlots && "Broken invariant");
123
124 assert(AllocatedStackSlots.size() ==
125 Builder.FuncInfo.StatepointStackSlots.size() &&
126 "Broken invariant");
127
128 for (; NextSlotToAllocate < NumSlots; NextSlotToAllocate++) {
129 if (!AllocatedStackSlots.test(Idx: NextSlotToAllocate)) {
130 const int FI = Builder.FuncInfo.StatepointStackSlots[NextSlotToAllocate];
131 if (MFI.getObjectSize(ObjectIdx: FI) == SpillSize) {
132 AllocatedStackSlots.set(NextSlotToAllocate);
133 // TODO: Is ValueType the right thing to use here?
134 return Builder.DAG.getFrameIndex(FI, VT: ValueType);
135 }
136 }
137 }
138
139 // Couldn't find a free slot, so create a new one:
140
141 SDValue SpillSlot = Builder.DAG.CreateStackTemporary(VT: ValueType);
142 const unsigned FI = cast<FrameIndexSDNode>(Val&: SpillSlot)->getIndex();
143 MFI.markAsStatepointSpillSlotObjectIndex(ObjectIdx: FI);
144
145 Builder.FuncInfo.StatepointStackSlots.push_back(Elt: FI);
146 AllocatedStackSlots.resize(N: AllocatedStackSlots.size()+1, t: true);
147 assert(AllocatedStackSlots.size() ==
148 Builder.FuncInfo.StatepointStackSlots.size() &&
149 "Broken invariant");
150
151 StatepointMaxSlotsRequired.updateMax(
152 V: Builder.FuncInfo.StatepointStackSlots.size());
153
154 return SpillSlot;
155}
156
157/// Utility function for reservePreviousStackSlotForValue. Tries to find
158/// stack slot index to which we have spilled value for previous statepoints.
159/// LookUpDepth specifies maximum DFS depth this function is allowed to look.
160static std::optional<int> findPreviousSpillSlot(const Value *Val,
161 SelectionDAGBuilder &Builder,
162 int LookUpDepth) {
163 // Can not look any further - give up now
164 if (LookUpDepth <= 0)
165 return std::nullopt;
166
167 // Spill location is known for gc relocates
168 if (const auto *Relocate = dyn_cast<GCRelocateInst>(Val)) {
169 const Value *Statepoint = Relocate->getStatepoint();
170 assert((isa<GCStatepointInst>(Statepoint) || isa<UndefValue>(Statepoint)) &&
171 "GetStatepoint must return one of two types");
172 if (isa<UndefValue>(Val: Statepoint))
173 return std::nullopt;
174
175 const auto &RelocationMap = Builder.FuncInfo.StatepointRelocationMaps
176 [cast<GCStatepointInst>(Val: Statepoint)];
177
178 auto It = RelocationMap.find(Val: Relocate);
179 if (It == RelocationMap.end())
180 return std::nullopt;
181
182 auto &Record = It->second;
183 if (Record.type != RecordType::Spill)
184 return std::nullopt;
185
186 return Record.payload.FI;
187 }
188
189 // Look through bitcast instructions.
190 if (const BitCastInst *Cast = dyn_cast<BitCastInst>(Val))
191 return findPreviousSpillSlot(Val: Cast->getOperand(i_nocapture: 0), Builder, LookUpDepth: LookUpDepth - 1);
192
193 // Look through phi nodes
194 // All incoming values should have same known stack slot, otherwise result
195 // is unknown.
196 if (const PHINode *Phi = dyn_cast<PHINode>(Val)) {
197 std::optional<int> MergedResult;
198
199 for (const auto &IncomingValue : Phi->incoming_values()) {
200 std::optional<int> SpillSlot =
201 findPreviousSpillSlot(Val: IncomingValue, Builder, LookUpDepth: LookUpDepth - 1);
202 if (!SpillSlot)
203 return std::nullopt;
204
205 if (MergedResult && *MergedResult != *SpillSlot)
206 return std::nullopt;
207
208 MergedResult = SpillSlot;
209 }
210 return MergedResult;
211 }
212
213 // TODO: We can do better for PHI nodes. In cases like this:
214 // ptr = phi(relocated_pointer, not_relocated_pointer)
215 // statepoint(ptr)
216 // We will return that stack slot for ptr is unknown. And later we might
217 // assign different stack slots for ptr and relocated_pointer. This limits
218 // llvm's ability to remove redundant stores.
219 // Unfortunately it's hard to accomplish in current infrastructure.
220 // We use this function to eliminate spill store completely, while
221 // in example we still need to emit store, but instead of any location
222 // we need to use special "preferred" location.
223
224 // TODO: handle simple updates. If a value is modified and the original
225 // value is no longer live, it would be nice to put the modified value in the
226 // same slot. This allows folding of the memory accesses for some
227 // instructions types (like an increment).
228 // statepoint (i)
229 // i1 = i+1
230 // statepoint (i1)
231 // However we need to be careful for cases like this:
232 // statepoint(i)
233 // i1 = i+1
234 // statepoint(i, i1)
235 // Here we want to reserve spill slot for 'i', but not for 'i+1'. If we just
236 // put handling of simple modifications in this function like it's done
237 // for bitcasts we might end up reserving i's slot for 'i+1' because order in
238 // which we visit values is unspecified.
239
240 // Don't know any information about this instruction
241 return std::nullopt;
242}
243
244/// Return true if-and-only-if the given SDValue can be lowered as either a
245/// constant argument or a stack reference. The key point is that the value
246/// doesn't need to be spilled or tracked as a vreg use.
247static bool willLowerDirectly(SDValue Incoming) {
248 // We are making an unchecked assumption that the frame size <= 2^16 as that
249 // is the largest offset which can be encoded in the stackmap format.
250 if (isa<FrameIndexSDNode>(Val: Incoming))
251 return true;
252
253 // The largest constant describeable in the StackMap format is 64 bits.
254 // Potential Optimization: Constants values are sign extended by consumer,
255 // and thus there are many constants of static type > 64 bits whose value
256 // happens to be sext(Con64) and could thus be lowered directly.
257 if (Incoming.getValueType().getSizeInBits() > 64)
258 return false;
259
260 return isIntOrFPConstant(V: Incoming) || Incoming.isUndef();
261}
262
263/// Try to find existing copies of the incoming values in stack slots used for
264/// statepoint spilling. If we can find a spill slot for the incoming value,
265/// mark that slot as allocated, and reuse the same slot for this safepoint.
266/// This helps to avoid series of loads and stores that only serve to reshuffle
267/// values on the stack between calls.
268static void reservePreviousStackSlotForValue(const Value *IncomingValue,
269 SelectionDAGBuilder &Builder) {
270 SDValue Incoming = Builder.getValue(V: IncomingValue);
271
272 // If we won't spill this, we don't need to check for previously allocated
273 // stack slots.
274 if (willLowerDirectly(Incoming))
275 return;
276
277 SDValue OldLocation = Builder.StatepointLowering.getLocation(Val: Incoming);
278 if (OldLocation.getNode())
279 // Duplicates in input
280 return;
281
282 const int LookUpDepth = 6;
283 std::optional<int> Index =
284 findPreviousSpillSlot(Val: IncomingValue, Builder, LookUpDepth);
285 if (!Index)
286 return;
287
288 const auto &StatepointSlots = Builder.FuncInfo.StatepointStackSlots;
289
290 auto SlotIt = find(Range: StatepointSlots, Val: *Index);
291 assert(SlotIt != StatepointSlots.end() &&
292 "Value spilled to the unknown stack slot");
293
294 // This is one of our dedicated lowering slots
295 const int Offset = std::distance(first: StatepointSlots.begin(), last: SlotIt);
296 if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) {
297 // stack slot already assigned to someone else, can't use it!
298 // TODO: currently we reserve space for gc arguments after doing
299 // normal allocation for deopt arguments. We should reserve for
300 // _all_ deopt and gc arguments, then start allocating. This
301 // will prevent some moves being inserted when vm state changes,
302 // but gc state doesn't between two calls.
303 return;
304 }
305 // Reserve this stack slot
306 Builder.StatepointLowering.reserveStackSlot(Offset);
307
308 // Cache this slot so we find it when going through the normal
309 // assignment loop.
310 SDValue Loc =
311 Builder.DAG.getTargetFrameIndex(FI: *Index, VT: Builder.getFrameIndexTy());
312 Builder.StatepointLowering.setLocation(Val: Incoming, Location: Loc);
313}
314
315/// Extract call from statepoint, lower it and return pointer to the
316/// call node. Also update NodeMap so that getValue(statepoint) will
317/// reference lowered call result
318static std::pair<SDValue, SDNode *> lowerCallFromStatepointLoweringInfo(
319 SelectionDAGBuilder::StatepointLoweringInfo &SI,
320 SelectionDAGBuilder &Builder) {
321 SDValue ReturnValue, CallEndVal;
322 std::tie(args&: ReturnValue, args&: CallEndVal) =
323 Builder.lowerInvokable(CLI&: SI.CLI, EHPadBB: SI.EHPadBB);
324 SDNode *CallEnd = CallEndVal.getNode();
325
326 // Get a call instruction from the call sequence chain. Tail calls are not
327 // allowed. The following code is essentially reverse engineering X86's
328 // LowerCallTo.
329 //
330 // We are expecting DAG to have the following form:
331 //
332 // ch = eh_label (only in case of invoke statepoint)
333 // ch, glue = callseq_start ch
334 // ch, glue = X86::Call ch, glue
335 // ch, glue = callseq_end ch, glue
336 // get_return_value ch, glue
337 //
338 // get_return_value can either be a sequence of CopyFromReg instructions
339 // to grab the return value from the return register(s), or it can be a LOAD
340 // to load a value returned by reference via a stack slot.
341
342 if (CallEnd->getOpcode() == ISD::EH_LABEL)
343 CallEnd = CallEnd->getOperand(Num: 0).getNode();
344
345 bool HasDef = !SI.CLI.RetTy->isVoidTy();
346 if (HasDef) {
347 if (CallEnd->getOpcode() == ISD::LOAD)
348 CallEnd = CallEnd->getOperand(Num: 0).getNode();
349 else
350 while (CallEnd->getOpcode() == ISD::CopyFromReg)
351 CallEnd = CallEnd->getOperand(Num: 0).getNode();
352 }
353
354 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
355 return std::make_pair(x&: ReturnValue, y: CallEnd->getOperand(Num: 0).getNode());
356}
357
358static MachineMemOperand* getMachineMemOperand(MachineFunction &MF,
359 FrameIndexSDNode &FI) {
360 auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FI.getIndex());
361 auto MMOFlags = MachineMemOperand::MOStore |
362 MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile;
363 auto &MFI = MF.getFrameInfo();
364 return MF.getMachineMemOperand(PtrInfo, F: MMOFlags,
365 Size: MFI.getObjectSize(ObjectIdx: FI.getIndex()),
366 BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI.getIndex()));
367}
368
369/// Spill a value incoming to the statepoint. It might be either part of
370/// vmstate
371/// or gcstate. In both cases unconditionally spill it on the stack unless it
372/// is a null constant. Return pair with first element being frame index
373/// containing saved value and second element with outgoing chain from the
374/// emitted store
375static std::tuple<SDValue, SDValue, MachineMemOperand*>
376spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
377 SelectionDAGBuilder &Builder) {
378 SDValue Loc = Builder.StatepointLowering.getLocation(Val: Incoming);
379 MachineMemOperand* MMO = nullptr;
380
381 // Emit new store if we didn't do it for this ptr before
382 if (!Loc.getNode()) {
383 Loc = Builder.StatepointLowering.allocateStackSlot(ValueType: Incoming.getValueType(),
384 Builder);
385 int Index = cast<FrameIndexSDNode>(Val&: Loc)->getIndex();
386 // We use TargetFrameIndex so that isel will not select it into LEA
387 Loc = Builder.DAG.getTargetFrameIndex(FI: Index, VT: Builder.getFrameIndexTy());
388
389 // Right now we always allocate spill slots that are of the same
390 // size as the value we're about to spill (the size of spillee can
391 // vary since we spill vectors of pointers too). At some point we
392 // can consider allowing spills of smaller values to larger slots
393 // (i.e. change the '==' in the assert below to a '>=').
394 MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
395 assert((MFI.getObjectSize(Index) * 8) ==
396 (-8 & (7 + // Round up modulo 8.
397 (int64_t)Incoming.getValueSizeInBits())) &&
398 "Bad spill: stack slot does not match!");
399
400 // Note: Using the alignment of the spill slot (rather than the abi or
401 // preferred alignment) is required for correctness when dealing with spill
402 // slots with preferred alignments larger than frame alignment..
403 auto &MF = Builder.DAG.getMachineFunction();
404 auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: Index);
405 auto *StoreMMO = MF.getMachineMemOperand(
406 PtrInfo, F: MachineMemOperand::MOStore, Size: MFI.getObjectSize(ObjectIdx: Index),
407 BaseAlignment: MFI.getObjectAlign(ObjectIdx: Index));
408 Chain = Builder.DAG.getStore(Chain, dl: Builder.getCurSDLoc(), Val: Incoming, Ptr: Loc,
409 MMO: StoreMMO);
410
411 MMO = getMachineMemOperand(MF, FI&: *cast<FrameIndexSDNode>(Val&: Loc));
412
413 Builder.StatepointLowering.setLocation(Val: Incoming, Location: Loc);
414 }
415
416 assert(Loc.getNode());
417 return std::make_tuple(args&: Loc, args&: Chain, args&: MMO);
418}
419
420/// Lower a single value incoming to a statepoint node. This value can be
421/// either a deopt value or a gc value, the handling is the same. We special
422/// case constants and allocas, then fall back to spilling if required.
423static void
424lowerIncomingStatepointValue(SDValue Incoming, bool RequireSpillSlot,
425 SmallVectorImpl<SDValue> &Ops,
426 SmallVectorImpl<MachineMemOperand *> &MemRefs,
427 SelectionDAGBuilder &Builder) {
428
429 if (willLowerDirectly(Incoming)) {
430 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Incoming)) {
431 // This handles allocas as arguments to the statepoint (this is only
432 // really meaningful for a deopt value. For GC, we'd be trying to
433 // relocate the address of the alloca itself?)
434 assert(Incoming.getValueType() == Builder.getFrameIndexTy() &&
435 "Incoming value is a frame index!");
436 Ops.push_back(Elt: Builder.DAG.getTargetFrameIndex(FI: FI->getIndex(),
437 VT: Builder.getFrameIndexTy()));
438
439 auto &MF = Builder.DAG.getMachineFunction();
440 auto *MMO = getMachineMemOperand(MF, FI&: *FI);
441 MemRefs.push_back(Elt: MMO);
442 return;
443 }
444
445 assert(Incoming.getValueType().getSizeInBits() <= 64);
446
447 if (Incoming.isUndef()) {
448 // Put an easily recognized constant that's unlikely to be a valid
449 // value so that uses of undef by the consumer of the stackmap is
450 // easily recognized. This is legal since the compiler is always
451 // allowed to chose an arbitrary value for undef.
452 pushStackMapConstant(Ops, Builder, Value: 0xFEFEFEFE);
453 return;
454 }
455
456 // If the original value was a constant, make sure it gets recorded as
457 // such in the stackmap. This is required so that the consumer can
458 // parse any internal format to the deopt state. It also handles null
459 // pointers and other constant pointers in GC states.
460 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Incoming)) {
461 pushStackMapConstant(Ops, Builder, Value: C->getSExtValue());
462 return;
463 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val&: Incoming)) {
464 pushStackMapConstant(Ops, Builder,
465 Value: C->getValueAPF().bitcastToAPInt().getZExtValue());
466 return;
467 }
468
469 llvm_unreachable("unhandled direct lowering case");
470 }
471
472
473
474 if (!RequireSpillSlot) {
475 // If this value is live in (not live-on-return, or live-through), we can
476 // treat it the same way patchpoint treats it's "live in" values. We'll
477 // end up folding some of these into stack references, but they'll be
478 // handled by the register allocator. Note that we do not have the notion
479 // of a late use so these values might be placed in registers which are
480 // clobbered by the call. This is fine for live-in. For live-through
481 // fix-up pass should be executed to force spilling of such registers.
482 Ops.push_back(Elt: Incoming);
483 } else {
484 // Otherwise, locate a spill slot and explicitly spill it so it can be
485 // found by the runtime later. Note: We know all of these spills are
486 // independent, but don't bother to exploit that chain wise. DAGCombine
487 // will happily do so as needed, so doing it here would be a small compile
488 // time win at most.
489 SDValue Chain = Builder.getRoot();
490 auto Res = spillIncomingStatepointValue(Incoming, Chain, Builder);
491 Ops.push_back(Elt: std::get<0>(t&: Res));
492 if (auto *MMO = std::get<2>(t&: Res))
493 MemRefs.push_back(Elt: MMO);
494 Chain = std::get<1>(t&: Res);
495 Builder.DAG.setRoot(Chain);
496 }
497
498}
499
500/// Return true if value V represents the GC value. The behavior is conservative
501/// in case it is not sure that value is not GC the function returns true.
502static bool isGCValue(const Value *V, SelectionDAGBuilder &Builder) {
503 auto *Ty = V->getType();
504 if (!Ty->isPtrOrPtrVectorTy())
505 return false;
506 if (auto *GFI = Builder.GFI)
507 if (auto IsManaged = GFI->getStrategy().isGCManagedPointer(Ty))
508 return *IsManaged;
509 return true; // conservative
510}
511
512/// Lower deopt state and gc pointer arguments of the statepoint. The actual
513/// lowering is described in lowerIncomingStatepointValue. This function is
514/// responsible for lowering everything in the right position and playing some
515/// tricks to avoid redundant stack manipulation where possible. On
516/// completion, 'Ops' will contain ready to use operands for machine code
517/// statepoint. The chain nodes will have already been created and the DAG root
518/// will be set to the last value spilled (if any were).
519static void
520lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
521 SmallVectorImpl<MachineMemOperand *> &MemRefs,
522 SmallVectorImpl<SDValue> &GCPtrs,
523 DenseMap<SDValue, int> &LowerAsVReg,
524 SelectionDAGBuilder::StatepointLoweringInfo &SI,
525 SelectionDAGBuilder &Builder) {
526 // Lower the deopt and gc arguments for this statepoint. Layout will be:
527 // deopt argument length, deopt arguments.., gc arguments...
528
529 // Figure out what lowering strategy we're going to use for each part
530 // Note: It is conservatively correct to lower both "live-in" and "live-out"
531 // as "live-through". A "live-through" variable is one which is "live-in",
532 // "live-out", and live throughout the lifetime of the call (i.e. we can find
533 // it from any PC within the transitive callee of the statepoint). In
534 // particular, if the callee spills callee preserved registers we may not
535 // be able to find a value placed in that register during the call. This is
536 // fine for live-out, but not for live-through. If we were willing to make
537 // assumptions about the code generator producing the callee, we could
538 // potentially allow live-through values in callee saved registers.
539 const bool LiveInDeopt =
540 SI.StatepointFlags & (uint64_t)StatepointFlags::DeoptLiveIn;
541
542 // Decide which deriver pointers will go on VRegs
543 unsigned MaxVRegPtrs = MaxRegistersForGCPointers.getValue();
544
545 // Pointers used on exceptional path of invoke statepoint.
546 // We cannot assing them to VRegs.
547 SmallSet<SDValue, 8> LPadPointers;
548 if (!UseRegistersForGCPointersInLandingPad)
549 if (const auto *StInvoke =
550 dyn_cast_or_null<InvokeInst>(Val: SI.StatepointInstr)) {
551 LandingPadInst *LPI = StInvoke->getLandingPadInst();
552 for (const auto *Relocate : SI.GCRelocates)
553 if (Relocate->getOperand(i_nocapture: 0) == LPI) {
554 LPadPointers.insert(V: Builder.getValue(V: Relocate->getBasePtr()));
555 LPadPointers.insert(V: Builder.getValue(V: Relocate->getDerivedPtr()));
556 }
557 }
558
559 LLVM_DEBUG(dbgs() << "Deciding how to lower GC Pointers:\n");
560
561 // List of unique lowered GC Pointer values.
562 SmallSetVector<SDValue, 16> LoweredGCPtrs;
563 // Map lowered GC Pointer value to the index in above vector
564 DenseMap<SDValue, unsigned> GCPtrIndexMap;
565
566 unsigned CurNumVRegs = 0;
567
568 auto canPassGCPtrOnVReg = [&](SDValue SD) {
569 if (SD.getValueType().isVector())
570 return false;
571 if (LPadPointers.count(V: SD))
572 return false;
573 return !willLowerDirectly(Incoming: SD);
574 };
575
576 auto processGCPtr = [&](const Value *V) {
577 SDValue PtrSD = Builder.getValue(V);
578 if (!LoweredGCPtrs.insert(X: PtrSD))
579 return; // skip duplicates
580 GCPtrIndexMap[PtrSD] = LoweredGCPtrs.size() - 1;
581
582 assert(!LowerAsVReg.count(PtrSD) && "must not have been seen");
583 if (LowerAsVReg.size() == MaxVRegPtrs)
584 return;
585 assert(V->getType()->isVectorTy() == PtrSD.getValueType().isVector() &&
586 "IR and SD types disagree");
587 if (!canPassGCPtrOnVReg(PtrSD)) {
588 LLVM_DEBUG(dbgs() << "direct/spill "; PtrSD.dump(&Builder.DAG));
589 return;
590 }
591 LLVM_DEBUG(dbgs() << "vreg "; PtrSD.dump(&Builder.DAG));
592 LowerAsVReg[PtrSD] = CurNumVRegs++;
593 };
594
595 // Process derived pointers first to give them more chance to go on VReg.
596 for (const Value *V : SI.Ptrs)
597 processGCPtr(V);
598 for (const Value *V : SI.Bases)
599 processGCPtr(V);
600
601 LLVM_DEBUG(dbgs() << LowerAsVReg.size() << " pointers will go in vregs\n");
602
603 auto requireSpillSlot = [&](const Value *V) {
604 if (!Builder.DAG.getTargetLoweringInfo().isTypeLegal(
605 VT: Builder.getValue(V).getValueType()))
606 return true;
607 if (isGCValue(V, Builder))
608 return !LowerAsVReg.count(Val: Builder.getValue(V));
609 return !(LiveInDeopt || UseRegistersForDeoptValues);
610 };
611
612 // Before we actually start lowering (and allocating spill slots for values),
613 // reserve any stack slots which we judge to be profitable to reuse for a
614 // particular value. This is purely an optimization over the code below and
615 // doesn't change semantics at all. It is important for performance that we
616 // reserve slots for both deopt and gc values before lowering either.
617 for (const Value *V : SI.DeoptState) {
618 if (requireSpillSlot(V))
619 reservePreviousStackSlotForValue(IncomingValue: V, Builder);
620 }
621
622 for (const Value *V : SI.Ptrs) {
623 SDValue SDV = Builder.getValue(V);
624 if (!LowerAsVReg.count(Val: SDV))
625 reservePreviousStackSlotForValue(IncomingValue: V, Builder);
626 }
627
628 for (const Value *V : SI.Bases) {
629 SDValue SDV = Builder.getValue(V);
630 if (!LowerAsVReg.count(Val: SDV))
631 reservePreviousStackSlotForValue(IncomingValue: V, Builder);
632 }
633
634 // First, prefix the list with the number of unique values to be
635 // lowered. Note that this is the number of *Values* not the
636 // number of SDValues required to lower them.
637 const int NumVMSArgs = SI.DeoptState.size();
638 pushStackMapConstant(Ops, Builder, Value: NumVMSArgs);
639
640 // The vm state arguments are lowered in an opaque manner. We do not know
641 // what type of values are contained within.
642 LLVM_DEBUG(dbgs() << "Lowering deopt state\n");
643 for (const Value *V : SI.DeoptState) {
644 SDValue Incoming;
645 // If this is a function argument at a static frame index, generate it as
646 // the frame index.
647 if (const Argument *Arg = dyn_cast<Argument>(Val: V)) {
648 int FI = Builder.FuncInfo.getArgumentFrameIndex(A: Arg);
649 if (FI != INT_MAX)
650 Incoming = Builder.DAG.getFrameIndex(FI, VT: Builder.getFrameIndexTy());
651 }
652 if (!Incoming.getNode())
653 Incoming = Builder.getValue(V);
654 LLVM_DEBUG(dbgs() << "Value " << *V
655 << " requireSpillSlot = " << requireSpillSlot(V) << "\n");
656 lowerIncomingStatepointValue(Incoming, RequireSpillSlot: requireSpillSlot(V), Ops, MemRefs,
657 Builder);
658 }
659
660 // Finally, go ahead and lower all the gc arguments.
661 pushStackMapConstant(Ops, Builder, Value: LoweredGCPtrs.size());
662 for (SDValue SDV : LoweredGCPtrs)
663 lowerIncomingStatepointValue(Incoming: SDV, RequireSpillSlot: !LowerAsVReg.count(Val: SDV), Ops, MemRefs,
664 Builder);
665
666 // Copy to out vector. LoweredGCPtrs will be empty after this point.
667 GCPtrs = LoweredGCPtrs.takeVector();
668
669 // If there are any explicit spill slots passed to the statepoint, record
670 // them, but otherwise do not do anything special. These are user provided
671 // allocas and give control over placement to the consumer. In this case,
672 // it is the contents of the slot which may get updated, not the pointer to
673 // the alloca
674 SmallVector<SDValue, 4> Allocas;
675 for (Value *V : SI.GCLives) {
676 SDValue Incoming = Builder.getValue(V);
677 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Incoming)) {
678 // This handles allocas as arguments to the statepoint
679 assert(Incoming.getValueType() == Builder.getFrameIndexTy() &&
680 "Incoming value is a frame index!");
681 Allocas.push_back(Elt: Builder.DAG.getTargetFrameIndex(
682 FI: FI->getIndex(), VT: Builder.getFrameIndexTy()));
683
684 auto &MF = Builder.DAG.getMachineFunction();
685 auto *MMO = getMachineMemOperand(MF, FI&: *FI);
686 MemRefs.push_back(Elt: MMO);
687 }
688 }
689 pushStackMapConstant(Ops, Builder, Value: Allocas.size());
690 Ops.append(in_start: Allocas.begin(), in_end: Allocas.end());
691
692 // Now construct GC base/derived map;
693 pushStackMapConstant(Ops, Builder, Value: SI.Ptrs.size());
694 SDLoc L = Builder.getCurSDLoc();
695 for (unsigned i = 0; i < SI.Ptrs.size(); ++i) {
696 SDValue Base = Builder.getValue(V: SI.Bases[i]);
697 assert(GCPtrIndexMap.count(Base) && "base not found in index map");
698 Ops.push_back(
699 Elt: Builder.DAG.getTargetConstant(Val: GCPtrIndexMap[Base], DL: L, VT: MVT::i64));
700 SDValue Derived = Builder.getValue(V: SI.Ptrs[i]);
701 assert(GCPtrIndexMap.count(Derived) && "derived not found in index map");
702 Ops.push_back(
703 Elt: Builder.DAG.getTargetConstant(Val: GCPtrIndexMap[Derived], DL: L, VT: MVT::i64));
704 }
705}
706
707SDValue SelectionDAGBuilder::LowerAsSTATEPOINT(
708 SelectionDAGBuilder::StatepointLoweringInfo &SI) {
709 // The basic scheme here is that information about both the original call and
710 // the safepoint is encoded in the CallInst. We create a temporary call and
711 // lower it, then reverse engineer the calling sequence.
712
713 NumOfStatepoints++;
714 // Clear state
715 StatepointLowering.startNewStatepoint(Builder&: *this);
716 assert(SI.Bases.size() == SI.Ptrs.size() && "Pointer without base!");
717 assert((GFI || SI.Bases.empty()) &&
718 "No gc specified, so cannot relocate pointers!");
719
720 LLVM_DEBUG(if (SI.StatepointInstr) dbgs()
721 << "Lowering statepoint " << *SI.StatepointInstr << "\n");
722#ifndef NDEBUG
723 for (const auto *Reloc : SI.GCRelocates)
724 if (Reloc->getParent() == SI.StatepointInstr->getParent())
725 StatepointLowering.scheduleRelocCall(*Reloc);
726#endif
727
728 // Lower statepoint vmstate and gcstate arguments
729
730 // All lowered meta args.
731 SmallVector<SDValue, 10> LoweredMetaArgs;
732 // Lowered GC pointers (subset of above).
733 SmallVector<SDValue, 16> LoweredGCArgs;
734 SmallVector<MachineMemOperand*, 16> MemRefs;
735 // Maps derived pointer SDValue to statepoint result of relocated pointer.
736 DenseMap<SDValue, int> LowerAsVReg;
737 lowerStatepointMetaArgs(Ops&: LoweredMetaArgs, MemRefs, GCPtrs&: LoweredGCArgs, LowerAsVReg,
738 SI, Builder&: *this);
739
740 // Now that we've emitted the spills, we need to update the root so that the
741 // call sequence is ordered correctly.
742 SI.CLI.setChain(getRoot());
743
744 // Get call node, we will replace it later with statepoint
745 SDValue ReturnVal;
746 SDNode *CallNode;
747 std::tie(args&: ReturnVal, args&: CallNode) = lowerCallFromStatepointLoweringInfo(SI, Builder&: *this);
748
749 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
750 // nodes with all the appropriate arguments and return values.
751
752 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
753 SDValue Chain = CallNode->getOperand(Num: 0);
754
755 SDValue Glue;
756 bool CallHasIncomingGlue = CallNode->getGluedNode();
757 if (CallHasIncomingGlue) {
758 // Glue is always last operand
759 Glue = CallNode->getOperand(Num: CallNode->getNumOperands() - 1);
760 }
761
762 // Build the GC_TRANSITION_START node if necessary.
763 //
764 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
765 // order in which they appear in the call to the statepoint intrinsic. If
766 // any of the operands is a pointer-typed, that operand is immediately
767 // followed by a SRCVALUE for the pointer that may be used during lowering
768 // (e.g. to form MachinePointerInfo values for loads/stores).
769 const bool IsGCTransition =
770 (SI.StatepointFlags & (uint64_t)StatepointFlags::GCTransition) ==
771 (uint64_t)StatepointFlags::GCTransition;
772 if (IsGCTransition) {
773 SmallVector<SDValue, 8> TSOps;
774
775 // Add chain
776 TSOps.push_back(Elt: Chain);
777
778 // Add GC transition arguments
779 for (const Value *V : SI.GCTransitionArgs) {
780 TSOps.push_back(Elt: getValue(V));
781 if (V->getType()->isPointerTy())
782 TSOps.push_back(Elt: DAG.getSrcValue(v: V));
783 }
784
785 // Add glue if necessary
786 if (CallHasIncomingGlue)
787 TSOps.push_back(Elt: Glue);
788
789 SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue);
790
791 SDValue GCTransitionStart =
792 DAG.getNode(Opcode: ISD::GC_TRANSITION_START, DL: getCurSDLoc(), VTList: NodeTys, Ops: TSOps);
793
794 Chain = GCTransitionStart.getValue(R: 0);
795 Glue = GCTransitionStart.getValue(R: 1);
796 }
797
798 // TODO: Currently, all of these operands are being marked as read/write in
799 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
800 // and flags to be read-only.
801 SmallVector<SDValue, 40> Ops;
802
803 // Add the <id> and <numBytes> constants.
804 Ops.push_back(Elt: DAG.getTargetConstant(Val: SI.ID, DL: getCurSDLoc(), VT: MVT::i64));
805 Ops.push_back(
806 Elt: DAG.getTargetConstant(Val: SI.NumPatchBytes, DL: getCurSDLoc(), VT: MVT::i32));
807
808 // Calculate and push starting position of vmstate arguments
809 // Get number of arguments incoming directly into call node
810 unsigned NumCallRegArgs =
811 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
812 Ops.push_back(Elt: DAG.getTargetConstant(Val: NumCallRegArgs, DL: getCurSDLoc(), VT: MVT::i32));
813
814 // Add call target
815 SDValue CallTarget = SDValue(CallNode->getOperand(Num: 1).getNode(), 0);
816 Ops.push_back(Elt: CallTarget);
817
818 // Add call arguments
819 // Get position of register mask in the call
820 SDNode::op_iterator RegMaskIt;
821 if (CallHasIncomingGlue)
822 RegMaskIt = CallNode->op_end() - 2;
823 else
824 RegMaskIt = CallNode->op_end() - 1;
825 Ops.insert(I: Ops.end(), From: CallNode->op_begin() + 2, To: RegMaskIt);
826
827 // Add a constant argument for the calling convention
828 pushStackMapConstant(Ops, Builder&: *this, Value: SI.CLI.CallConv);
829
830 // Add a constant argument for the flags
831 uint64_t Flags = SI.StatepointFlags;
832 assert(((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0) &&
833 "Unknown flag used");
834 pushStackMapConstant(Ops, Builder&: *this, Value: Flags);
835
836 // Insert all vmstate and gcstate arguments
837 llvm::append_range(C&: Ops, R&: LoweredMetaArgs);
838
839 // Add register mask from call node
840 Ops.push_back(Elt: *RegMaskIt);
841
842 // Add chain
843 Ops.push_back(Elt: Chain);
844
845 // Same for the glue, but we add it only if original call had it
846 if (Glue.getNode())
847 Ops.push_back(Elt: Glue);
848
849 // Compute return values. Provide a glue output since we consume one as
850 // input. This allows someone else to chain off us as needed.
851 SmallVector<EVT, 8> NodeTys;
852 for (auto SD : LoweredGCArgs) {
853 if (!LowerAsVReg.count(Val: SD))
854 continue;
855 NodeTys.push_back(Elt: SD.getValueType());
856 }
857 LLVM_DEBUG(dbgs() << "Statepoint has " << NodeTys.size() << " results\n");
858 assert(NodeTys.size() == LowerAsVReg.size() && "Inconsistent GC Ptr lowering");
859 NodeTys.push_back(Elt: MVT::Other);
860 NodeTys.push_back(Elt: MVT::Glue);
861
862 unsigned NumResults = NodeTys.size();
863 MachineSDNode *StatepointMCNode =
864 DAG.getMachineNode(Opcode: TargetOpcode::STATEPOINT, dl: getCurSDLoc(), ResultTys: NodeTys, Ops);
865 DAG.setNodeMemRefs(N: StatepointMCNode, NewMemRefs: MemRefs);
866
867 // For values lowered to tied-defs, create the virtual registers if used
868 // in other blocks. For local gc.relocate record appropriate statepoint
869 // result in StatepointLoweringState.
870 DenseMap<SDValue, Register> VirtRegs;
871 for (const auto *Relocate : SI.GCRelocates) {
872 Value *Derived = Relocate->getDerivedPtr();
873 SDValue SD = getValue(V: Derived);
874 auto It = LowerAsVReg.find(Val: SD);
875 if (It == LowerAsVReg.end())
876 continue;
877
878 SDValue Relocated = SDValue(StatepointMCNode, It->second);
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 auto [VRegIt, Inserted] = VirtRegs.try_emplace(Key: SD);
893 if (!Inserted)
894 continue;
895
896 auto *RetTy = Relocate->getType();
897 Register Reg = FuncInfo.CreateRegs(Ty: RetTy);
898 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
899 DAG.getDataLayout(), Reg, RetTy, std::nullopt);
900 SDValue Chain = DAG.getRoot();
901 RFV.getCopyToRegs(Val: Relocated, DAG, dl: getCurSDLoc(), Chain, Glue: nullptr);
902 PendingExports.push_back(Elt: Chain);
903
904 VRegIt->second = Reg;
905 }
906
907 // Record for later use how each relocation was lowered. This is needed to
908 // allow later gc.relocates to mirror the lowering chosen.
909 const Instruction *StatepointInstr = SI.StatepointInstr;
910 auto &RelocationMap = FuncInfo.StatepointRelocationMaps[StatepointInstr];
911 for (const GCRelocateInst *Relocate : SI.GCRelocates) {
912 const Value *V = Relocate->getDerivedPtr();
913 SDValue SDV = getValue(V);
914 SDValue Loc = StatepointLowering.getLocation(Val: SDV);
915
916 bool IsLocal = (Relocate->getParent() == StatepointInstr->getParent());
917
918 RecordType Record;
919 if (LowerAsVReg.count(Val: SDV)) {
920 if (IsLocal) {
921 // Result is already stored in StatepointLowering
922 Record.type = RecordType::SDValueNode;
923 } else {
924 Record.type = RecordType::VReg;
925 auto It = VirtRegs.find(Val: SDV);
926 assert(It != VirtRegs.end());
927 Record.payload.Reg = It->second;
928 }
929 } else if (Loc.getNode()) {
930 Record.type = RecordType::Spill;
931 Record.payload.FI = cast<FrameIndexSDNode>(Val&: Loc)->getIndex();
932 } else {
933 Record.type = RecordType::NoRelocate;
934 // If we didn't relocate a value, we'll essentialy end up inserting an
935 // additional use of the original value when lowering the gc.relocate.
936 // We need to make sure the value is available at the new use, which
937 // might be in another block.
938 if (Relocate->getParent() != StatepointInstr->getParent())
939 ExportFromCurrentBlock(V);
940 }
941 RelocationMap[Relocate] = Record;
942 }
943
944
945
946 SDNode *SinkNode = StatepointMCNode;
947
948 // Build the GC_TRANSITION_END node if necessary.
949 //
950 // See the comment above regarding GC_TRANSITION_START for the layout of
951 // the operands to the GC_TRANSITION_END node.
952 if (IsGCTransition) {
953 SmallVector<SDValue, 8> TEOps;
954
955 // Add chain
956 TEOps.push_back(Elt: SDValue(StatepointMCNode, NumResults - 2));
957
958 // Add GC transition arguments
959 for (const Value *V : SI.GCTransitionArgs) {
960 TEOps.push_back(Elt: getValue(V));
961 if (V->getType()->isPointerTy())
962 TEOps.push_back(Elt: DAG.getSrcValue(v: V));
963 }
964
965 // Add glue
966 TEOps.push_back(Elt: SDValue(StatepointMCNode, NumResults - 1));
967
968 SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue);
969
970 SDValue GCTransitionStart =
971 DAG.getNode(Opcode: ISD::GC_TRANSITION_END, DL: getCurSDLoc(), VTList: NodeTys, Ops: TEOps);
972
973 SinkNode = GCTransitionStart.getNode();
974 }
975
976 // Replace original call
977 // Call: ch,glue = CALL ...
978 // Statepoint: [gc relocates],ch,glue = STATEPOINT ...
979 unsigned NumSinkValues = SinkNode->getNumValues();
980 SDValue StatepointValues[2] = {SDValue(SinkNode, NumSinkValues - 2),
981 SDValue(SinkNode, NumSinkValues - 1)};
982 DAG.ReplaceAllUsesWith(From: CallNode, To: StatepointValues);
983 // Remove original call node
984 DAG.DeleteNode(N: CallNode);
985
986 // Since we always emit CopyToRegs (even for local relocates), we must
987 // update root, so that they are emitted before any local uses.
988 (void)getControlRoot();
989
990 // TODO: A better future implementation would be to emit a single variable
991 // argument, variable return value STATEPOINT node here and then hookup the
992 // return value of each gc.relocate to the respective output of the
993 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
994 // to actually be possible today.
995
996 return ReturnVal;
997}
998
999/// Return two gc.results if present. First result is a block local
1000/// gc.result, second result is a non-block local gc.result. Corresponding
1001/// entry will be nullptr if not present.
1002static std::pair<const GCResultInst*, const GCResultInst*>
1003getGCResultLocality(const GCStatepointInst &S) {
1004 std::pair<const GCResultInst *, const GCResultInst*> Res(nullptr, nullptr);
1005 for (const auto *U : S.users()) {
1006 auto *GRI = dyn_cast<GCResultInst>(Val: U);
1007 if (!GRI)
1008 continue;
1009 if (GRI->getParent() == S.getParent())
1010 Res.first = GRI;
1011 else
1012 Res.second = GRI;
1013 }
1014 return Res;
1015}
1016
1017void
1018SelectionDAGBuilder::LowerStatepoint(const GCStatepointInst &I,
1019 const BasicBlock *EHPadBB /*= nullptr*/) {
1020 assert(I.getCallingConv() != CallingConv::AnyReg &&
1021 "anyregcc is not supported on statepoints!");
1022
1023#ifndef NDEBUG
1024 // Check that the associated GCStrategy expects to encounter statepoints.
1025 assert(GFI->getStrategy().useStatepoints() &&
1026 "GCStrategy does not expect to encounter statepoints");
1027#endif
1028
1029 SDValue ActualCallee;
1030 SDValue Callee = getValue(V: I.getActualCalledOperand());
1031
1032 if (I.getNumPatchBytes() > 0) {
1033 // If we've been asked to emit a nop sequence instead of a call instruction
1034 // for this statepoint then don't lower the call target, but use a constant
1035 // `undef` instead. Not lowering the call target lets statepoint clients
1036 // get away without providing a physical address for the symbolic call
1037 // target at link time.
1038 ActualCallee = DAG.getUNDEF(VT: Callee.getValueType());
1039 } else {
1040 ActualCallee = Callee;
1041 }
1042
1043 const auto GCResultLocality = getGCResultLocality(S: I);
1044 AttributeSet retAttrs;
1045 if (GCResultLocality.first)
1046 retAttrs = GCResultLocality.first->getAttributes().getRetAttrs();
1047
1048 StatepointLoweringInfo SI(DAG);
1049 populateCallLoweringInfo(CLI&: SI.CLI, Call: &I, ArgIdx: GCStatepointInst::CallArgsBeginPos,
1050 NumArgs: I.getNumCallArgs(), Callee: ActualCallee,
1051 ReturnTy: I.getActualReturnType(), RetAttrs: retAttrs,
1052 /*IsPatchPoint=*/false);
1053
1054 // There may be duplication in the gc.relocate list; such as two copies of
1055 // each relocation on normal and exceptional path for an invoke. We only
1056 // need to spill once and record one copy in the stackmap, but we need to
1057 // reload once per gc.relocate. (Dedupping gc.relocates is trickier and best
1058 // handled as a CSE problem elsewhere.)
1059 // TODO: There a couple of major stackmap size optimizations we could do
1060 // here if we wished.
1061 // 1) If we've encountered a derived pair {B, D}, we don't need to actually
1062 // record {B,B} if it's seen later.
1063 // 2) Due to rematerialization, actual derived pointers are somewhat rare;
1064 // given that, we could change the format to record base pointer relocations
1065 // separately with half the space. This would require a format rev and a
1066 // fairly major rework of the STATEPOINT node though.
1067 SmallSet<SDValue, 8> Seen;
1068 for (const GCRelocateInst *Relocate : I.getGCRelocates()) {
1069 SI.GCRelocates.push_back(Elt: Relocate);
1070
1071 SDValue DerivedSD = getValue(V: Relocate->getDerivedPtr());
1072 if (Seen.insert(V: DerivedSD).second) {
1073 SI.Bases.push_back(Elt: Relocate->getBasePtr());
1074 SI.Ptrs.push_back(Elt: Relocate->getDerivedPtr());
1075 }
1076 }
1077
1078 // If we find a deopt value which isn't explicitly added, we need to
1079 // ensure it gets lowered such that gc cycles occurring before the
1080 // deoptimization event during the lifetime of the call don't invalidate
1081 // the pointer we're deopting with. Note that we assume that all
1082 // pointers passed to deopt are base pointers; relaxing that assumption
1083 // would require relatively large changes to how we represent relocations.
1084 for (Value *V : I.deopt_operands()) {
1085 if (!isGCValue(V, Builder&: *this))
1086 continue;
1087 if (Seen.insert(V: getValue(V)).second) {
1088 SI.Bases.push_back(Elt: V);
1089 SI.Ptrs.push_back(Elt: V);
1090 }
1091 }
1092
1093 SI.GCLives = ArrayRef<const Use>(I.gc_live_begin(), I.gc_live_end());
1094 SI.StatepointInstr = &I;
1095 SI.ID = I.getID();
1096
1097 SI.DeoptState = ArrayRef<const Use>(I.deopt_begin(), I.deopt_end());
1098 SI.GCTransitionArgs = ArrayRef<const Use>(I.gc_transition_args_begin(),
1099 I.gc_transition_args_end());
1100
1101 SI.StatepointFlags = I.getFlags();
1102 SI.NumPatchBytes = I.getNumPatchBytes();
1103 SI.EHPadBB = EHPadBB;
1104
1105 SDValue ReturnValue = LowerAsSTATEPOINT(SI);
1106
1107 // Export the result value if needed
1108 if (!GCResultLocality.first && !GCResultLocality.second) {
1109 // The return value is not needed, just generate a poison value.
1110 // Note: This covers the void return case.
1111 setValue(V: &I, NewN: DAG.getIntPtrConstant(Val: -1, DL: getCurSDLoc()));
1112 return;
1113 }
1114
1115 if (GCResultLocality.first) {
1116 // Result value will be used in a same basic block. Don't export it or
1117 // perform any explicit register copies. The gc_result will simply grab
1118 // this value.
1119 setValue(V: &I, NewN: ReturnValue);
1120 }
1121
1122 if (!GCResultLocality.second)
1123 return;
1124 // Result value will be used in a different basic block so we need to export
1125 // it now. Default exporting mechanism will not work here because statepoint
1126 // call has a different type than the actual call. It means that by default
1127 // llvm will create export register of the wrong type (always i32 in our
1128 // case). So instead we need to create export register with correct type
1129 // manually.
1130 // TODO: To eliminate this problem we can remove gc.result intrinsics
1131 // completely and make statepoint call to return a tuple.
1132 Type *RetTy = GCResultLocality.second->getType();
1133 Register Reg = FuncInfo.CreateRegs(Ty: RetTy);
1134 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
1135 DAG.getDataLayout(), Reg, RetTy,
1136 I.getCallingConv());
1137 SDValue Chain = DAG.getEntryNode();
1138
1139 RFV.getCopyToRegs(Val: ReturnValue, DAG, dl: getCurSDLoc(), Chain, Glue: nullptr);
1140 PendingExports.push_back(Elt: Chain);
1141 FuncInfo.ValueMap[&I] = Reg;
1142}
1143
1144void SelectionDAGBuilder::LowerCallSiteWithDeoptBundleImpl(
1145 const CallBase *Call, SDValue Callee, const BasicBlock *EHPadBB,
1146 bool VarArgDisallowed, bool ForceVoidReturnTy) {
1147 StatepointLoweringInfo SI(DAG);
1148 unsigned ArgBeginIndex = Call->arg_begin() - Call->op_begin();
1149 populateCallLoweringInfo(
1150 CLI&: SI.CLI, Call, ArgIdx: ArgBeginIndex, NumArgs: Call->arg_size(), Callee,
1151 ReturnTy: ForceVoidReturnTy ? Type::getVoidTy(C&: *DAG.getContext()) : Call->getType(),
1152 RetAttrs: Call->getAttributes().getRetAttrs(), /*IsPatchPoint=*/false);
1153 if (!VarArgDisallowed)
1154 SI.CLI.IsVarArg = Call->getFunctionType()->isVarArg();
1155
1156 auto DeoptBundle = *Call->getOperandBundle(ID: LLVMContext::OB_deopt);
1157
1158 unsigned DefaultID = StatepointDirectives::DeoptBundleStatepointID;
1159
1160 auto SD = parseStatepointDirectivesFromAttrs(AS: Call->getAttributes());
1161 SI.ID = SD.StatepointID.value_or(u&: DefaultID);
1162 SI.NumPatchBytes = SD.NumPatchBytes.value_or(u: 0);
1163
1164 SI.DeoptState =
1165 ArrayRef<const Use>(DeoptBundle.Inputs.begin(), DeoptBundle.Inputs.end());
1166 SI.StatepointFlags = static_cast<uint64_t>(StatepointFlags::None);
1167 SI.EHPadBB = EHPadBB;
1168
1169 // NB! The GC arguments are deliberately left empty.
1170
1171 LLVM_DEBUG(dbgs() << "Lowering call with deopt bundle " << *Call << "\n");
1172 if (SDValue ReturnVal = LowerAsSTATEPOINT(SI)) {
1173 ReturnVal = lowerRangeToAssertZExt(DAG, I: *Call, Op: ReturnVal);
1174 setValue(V: Call, NewN: ReturnVal);
1175 }
1176}
1177
1178void SelectionDAGBuilder::LowerCallSiteWithDeoptBundle(
1179 const CallBase *Call, SDValue Callee, const BasicBlock *EHPadBB) {
1180 LowerCallSiteWithDeoptBundleImpl(Call, Callee, EHPadBB,
1181 /* VarArgDisallowed = */ false,
1182 /* ForceVoidReturnTy = */ false);
1183}
1184
1185void SelectionDAGBuilder::visitGCResult(const GCResultInst &CI) {
1186 // The result value of the gc_result is simply the result of the actual
1187 // call. We've already emitted this, so just grab the value.
1188 const Value *SI = CI.getStatepoint();
1189 assert((isa<GCStatepointInst>(SI) || isa<UndefValue>(SI)) &&
1190 "GetStatepoint must return one of two types");
1191 if (isa<UndefValue>(Val: SI))
1192 return;
1193
1194 if (cast<GCStatepointInst>(Val: SI)->getParent() == CI.getParent()) {
1195 setValue(V: &CI, NewN: getValue(V: SI));
1196 return;
1197 }
1198 // Statepoint is in different basic block so we should have stored call
1199 // result in a virtual register.
1200 // We can not use default getValue() functionality to copy value from this
1201 // register because statepoint and actual call return types can be
1202 // different, and getValue() will use CopyFromReg of the wrong type,
1203 // which is always i32 in our case.
1204 Type *RetTy = CI.getType();
1205 SDValue CopyFromReg = getCopyFromRegs(V: SI, Ty: RetTy);
1206
1207 assert(CopyFromReg.getNode());
1208 setValue(V: &CI, NewN: CopyFromReg);
1209}
1210
1211void SelectionDAGBuilder::visitGCRelocate(const GCRelocateInst &Relocate) {
1212 const Value *Statepoint = Relocate.getStatepoint();
1213#ifndef NDEBUG
1214 // Consistency check
1215 // We skip this check for relocates not in the same basic block as their
1216 // statepoint. It would be too expensive to preserve validation info through
1217 // different basic blocks.
1218 assert((isa<GCStatepointInst>(Statepoint) || isa<UndefValue>(Statepoint)) &&
1219 "GetStatepoint must return one of two types");
1220 if (isa<UndefValue>(Statepoint))
1221 return;
1222
1223 if (cast<GCStatepointInst>(Statepoint)->getParent() == Relocate.getParent())
1224 StatepointLowering.relocCallVisited(Relocate);
1225#endif
1226
1227 const Value *DerivedPtr = Relocate.getDerivedPtr();
1228 auto &RelocationMap =
1229 FuncInfo.StatepointRelocationMaps[cast<GCStatepointInst>(Val: Statepoint)];
1230 auto SlotIt = RelocationMap.find(Val: &Relocate);
1231 assert(SlotIt != RelocationMap.end() && "Relocating not lowered gc value");
1232 const RecordType &Record = SlotIt->second;
1233
1234 // If relocation was done via virtual register..
1235 if (Record.type == RecordType::SDValueNode) {
1236 assert(cast<GCStatepointInst>(Statepoint)->getParent() ==
1237 Relocate.getParent() &&
1238 "Nonlocal gc.relocate mapped via SDValue");
1239 SDValue SDV = StatepointLowering.getLocation(Val: getValue(V: DerivedPtr));
1240 assert(SDV.getNode() && "empty SDValue");
1241 setValue(V: &Relocate, NewN: SDV);
1242 return;
1243 }
1244 if (Record.type == RecordType::VReg) {
1245 Register InReg = Record.payload.Reg;
1246 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
1247 DAG.getDataLayout(), InReg, Relocate.getType(),
1248 std::nullopt); // This is not an ABI copy.
1249 // We generate copy to/from regs even for local uses, hence we must
1250 // chain with current root to ensure proper ordering of copies w.r.t.
1251 // statepoint.
1252 SDValue Chain = DAG.getRoot();
1253 SDValue Relocation = RFV.getCopyFromRegs(DAG, FuncInfo, dl: getCurSDLoc(),
1254 Chain, Glue: nullptr, V: nullptr);
1255 setValue(V: &Relocate, NewN: Relocation);
1256 return;
1257 }
1258
1259 if (Record.type == RecordType::Spill) {
1260 unsigned Index = Record.payload.FI;
1261 SDValue SpillSlot = DAG.getTargetFrameIndex(FI: Index, VT: getFrameIndexTy());
1262
1263 // All the reloads are independent and are reading memory only modified by
1264 // statepoints (i.e. no other aliasing stores); informing SelectionDAG of
1265 // this lets CSE kick in for free and allows reordering of
1266 // instructions if possible. The lowering for statepoint sets the root,
1267 // so this is ordering all reloads with the either
1268 // a) the statepoint node itself, or
1269 // b) the entry of the current block for an invoke statepoint.
1270 const SDValue Chain = DAG.getRoot(); // != Builder.getRoot()
1271
1272 auto &MF = DAG.getMachineFunction();
1273 auto &MFI = MF.getFrameInfo();
1274 auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: Index);
1275 auto *LoadMMO = MF.getMachineMemOperand(PtrInfo, F: MachineMemOperand::MOLoad,
1276 Size: MFI.getObjectSize(ObjectIdx: Index),
1277 BaseAlignment: MFI.getObjectAlign(ObjectIdx: Index));
1278
1279 auto LoadVT = DAG.getTargetLoweringInfo().getValueType(DL: DAG.getDataLayout(),
1280 Ty: Relocate.getType());
1281
1282 SDValue SpillLoad =
1283 DAG.getLoad(VT: LoadVT, dl: getCurSDLoc(), Chain, Ptr: SpillSlot, MMO: LoadMMO);
1284 PendingLoads.push_back(Elt: SpillLoad.getValue(R: 1));
1285
1286 assert(SpillLoad.getNode());
1287 setValue(V: &Relocate, NewN: SpillLoad);
1288 return;
1289 }
1290
1291 assert(Record.type == RecordType::NoRelocate);
1292 SDValue SD = getValue(V: DerivedPtr);
1293
1294 if (SD.isUndef() && SD.getValueType().getSizeInBits() <= 64) {
1295 // Lowering relocate(undef) as arbitrary constant. Current constant value
1296 // is chosen such that it's unlikely to be a valid pointer.
1297 setValue(V: &Relocate, NewN: DAG.getConstant(Val: 0xFEFEFEFE, DL: SDLoc(SD), VT: MVT::i64));
1298 return;
1299 }
1300
1301 // We didn't need to spill these special cases (constants and allocas).
1302 // See the handling in spillIncomingValueForStatepoint for detail.
1303 setValue(V: &Relocate, NewN: SD);
1304}
1305
1306void SelectionDAGBuilder::LowerDeoptimizeCall(const CallInst *CI) {
1307 const auto &TLI = DAG.getTargetLoweringInfo();
1308 SDValue Callee = DAG.getExternalSymbol(Sym: TLI.getLibcallName(Call: RTLIB::DEOPTIMIZE),
1309 VT: TLI.getPointerTy(DL: DAG.getDataLayout()));
1310
1311 // We don't lower calls to __llvm_deoptimize as varargs, but as a regular
1312 // call. We also do not lower the return value to any virtual register, and
1313 // change the immediately following return to a trap instruction.
1314 LowerCallSiteWithDeoptBundleImpl(Call: CI, Callee, /* EHPadBB = */ nullptr,
1315 /* VarArgDisallowed = */ true,
1316 /* ForceVoidReturnTy = */ true);
1317}
1318
1319void SelectionDAGBuilder::LowerDeoptimizingReturn() {
1320 // We do not lower the return value from llvm.deoptimize to any virtual
1321 // register, and change the immediately following return to a trap
1322 // instruction.
1323 if (DAG.getTarget().Options.TrapUnreachable)
1324 DAG.setRoot(
1325 DAG.getNode(Opcode: ISD::TRAP, DL: getCurSDLoc(), VT: MVT::Other, Operand: DAG.getRoot()));
1326}
1327