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