1 //===-- StatepointLowering.cpp - SDAGBuilder's statepoint code -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file includes support code use by SelectionDAGBuilder when lowering a
11 // statepoint sequence in SelectionDAG IR.
13 //===----------------------------------------------------------------------===//
15 #include "StatepointLowering.h"
16 #include "SelectionDAGBuilder.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/CodeGen/FunctionLoweringInfo.h"
20 #include "llvm/CodeGen/GCMetadata.h"
21 #include "llvm/CodeGen/GCStrategy.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/StackMaps.h"
24 #include "llvm/IR/CallingConv.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/Statepoint.h"
29 #include "llvm/Target/TargetLowering.h"
33 #define DEBUG_TYPE "statepoint-lowering"
35 STATISTIC(NumSlotsAllocatedForStatepoints,
36 "Number of stack slots allocated for statepoints");
37 STATISTIC(NumOfStatepoints, "Number of statepoint nodes encountered");
38 STATISTIC(StatepointMaxSlotsRequired,
39 "Maximum number of stack slots required for a singe statepoint");
41 void StatepointLoweringState::startNewStatepoint(SelectionDAGBuilder &Builder) {
43 assert(PendingGCRelocateCalls.empty() &&
44 "Trying to visit statepoint before finished processing previous one");
46 RelocLocations.clear();
47 NextSlotToAllocate = 0;
48 // Need to resize this on each safepoint - we need the two to stay in
49 // sync and the clear patterns of a SelectionDAGBuilder have no relation
50 // to FunctionLoweringInfo.
51 AllocatedStackSlots.resize(Builder.FuncInfo.StatepointStackSlots.size());
52 for (size_t i = 0; i < AllocatedStackSlots.size(); i++) {
53 AllocatedStackSlots[i] = false;
56 void StatepointLoweringState::clear() {
58 RelocLocations.clear();
59 AllocatedStackSlots.clear();
60 assert(PendingGCRelocateCalls.empty() &&
61 "cleared before statepoint sequence completed");
65 StatepointLoweringState::allocateStackSlot(EVT ValueType,
66 SelectionDAGBuilder &Builder) {
68 NumSlotsAllocatedForStatepoints++;
70 // The basic scheme here is to first look for a previously created stack slot
71 // which is not in use (accounting for the fact arbitrary slots may already
72 // be reserved), or to create a new stack slot and use it.
74 // If this doesn't succeed in 40000 iterations, something is seriously wrong
75 for (int i = 0; i < 40000; i++) {
76 assert(Builder.FuncInfo.StatepointStackSlots.size() ==
77 AllocatedStackSlots.size() &&
79 const size_t NumSlots = AllocatedStackSlots.size();
80 assert(NextSlotToAllocate <= NumSlots && "broken invariant");
82 if (NextSlotToAllocate >= NumSlots) {
83 assert(NextSlotToAllocate == NumSlots);
85 if (NumSlots + 1 > StatepointMaxSlotsRequired) {
86 StatepointMaxSlotsRequired = NumSlots + 1;
89 SDValue SpillSlot = Builder.DAG.CreateStackTemporary(ValueType);
90 const unsigned FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
91 Builder.FuncInfo.StatepointStackSlots.push_back(FI);
92 AllocatedStackSlots.push_back(true);
95 if (!AllocatedStackSlots[NextSlotToAllocate]) {
96 const int FI = Builder.FuncInfo.StatepointStackSlots[NextSlotToAllocate];
97 AllocatedStackSlots[NextSlotToAllocate] = true;
98 return Builder.DAG.getFrameIndex(FI, ValueType);
100 // Note: We deliberately choose to advance this only on the failing path.
101 // Doing so on the suceeding path involes a bit of complexity that caused a
102 // minor bug previously. Unless performance shows this matters, please
103 // keep this code as simple as possible.
104 NextSlotToAllocate++;
106 llvm_unreachable("infinite loop?");
109 /// Try to find existing copies of the incoming values in stack slots used for
110 /// statepoint spilling. If we can find a spill slot for the incoming value,
111 /// mark that slot as allocated, and reuse the same slot for this safepoint.
112 /// This helps to avoid series of loads and stores that only serve to resuffle
113 /// values on the stack between calls.
114 static void reservePreviousStackSlotForValue(SDValue Incoming,
115 SelectionDAGBuilder &Builder) {
117 if (isa<ConstantSDNode>(Incoming) || isa<FrameIndexSDNode>(Incoming)) {
118 // We won't need to spill this, so no need to check for previously
119 // allocated stack slots
123 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
125 // duplicates in input
129 // Search back for the load from a stack slot pattern to find the original
130 // slot we allocated for this value. We could extend this to deal with
131 // simple modification patterns, but simple dealing with trivial load/store
132 // sequences helps a lot already.
133 if (LoadSDNode *Load = dyn_cast<LoadSDNode>(Incoming)) {
134 if (auto *FI = dyn_cast<FrameIndexSDNode>(Load->getBasePtr())) {
135 const int Index = FI->getIndex();
136 auto Itr = std::find(Builder.FuncInfo.StatepointStackSlots.begin(),
137 Builder.FuncInfo.StatepointStackSlots.end(), Index);
138 if (Itr == Builder.FuncInfo.StatepointStackSlots.end()) {
139 // not one of the lowering stack slots, can't reuse!
140 // TODO: Actually, we probably could reuse the stack slot if the value
141 // hasn't changed at all, but we'd need to look for intervening writes
144 // This is one of our dedicated lowering slots
146 std::distance(Builder.FuncInfo.StatepointStackSlots.begin(), Itr);
147 if (Builder.StatepointLowering.isStackSlotAllocated(Offset)) {
148 // stack slot already assigned to someone else, can't use it!
149 // TODO: currently we reserve space for gc arguments after doing
150 // normal allocation for deopt arguments. We should reserve for
151 // _all_ deopt and gc arguments, then start allocating. This
152 // will prevent some moves being inserted when vm state changes,
153 // but gc state doesn't between two calls.
156 // Reserve this stack slot
157 Builder.StatepointLowering.reserveStackSlot(Offset);
160 // Cache this slot so we find it when going through the normal
163 Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
165 Builder.StatepointLowering.setLocation(Incoming, Loc);
169 // TODO: handle case where a reloaded value flows through a phi to
170 // another safepoint. e.g.
173 // bb2: % pred: bb1, bb3, bb4, etc.
174 // a_phi = phi(a', ...)
175 // statepoint ... a_phi
176 // NOTE: This will require reasoning about cross basic block values. This is
177 // decidedly non trivial and this might not be the right place to do it. We
178 // don't really have the information we need here...
180 // TODO: handle simple updates. If a value is modified and the original
181 // value is no longer live, it would be nice to put the modified value in the
182 // same slot. This allows folding of the memory accesses for some
183 // instructions types (like an increment).
189 /// Remove any duplicate (as SDValues) from the derived pointer pairs. This
190 /// is not required for correctness. It's purpose is to reduce the size of
191 /// StackMap section. It has no effect on the number of spill slots required
192 /// or the actual lowering.
193 static void removeDuplicatesGCPtrs(SmallVectorImpl<const Value *> &Bases,
194 SmallVectorImpl<const Value *> &Ptrs,
195 SmallVectorImpl<const Value *> &Relocs,
196 SelectionDAGBuilder &Builder) {
198 // This is horribly ineffecient, but I don't care right now
199 SmallSet<SDValue, 64> Seen;
201 SmallVector<const Value *, 64> NewBases, NewPtrs, NewRelocs;
202 for (size_t i = 0; i < Ptrs.size(); i++) {
203 SDValue SD = Builder.getValue(Ptrs[i]);
204 // Only add non-duplicates
205 if (Seen.count(SD) == 0) {
206 NewBases.push_back(Bases[i]);
207 NewPtrs.push_back(Ptrs[i]);
208 NewRelocs.push_back(Relocs[i]);
212 assert(Bases.size() >= NewBases.size());
213 assert(Ptrs.size() >= NewPtrs.size());
214 assert(Relocs.size() >= NewRelocs.size());
218 assert(Ptrs.size() == Bases.size());
219 assert(Ptrs.size() == Relocs.size());
222 /// Extract call from statepoint, lower it and return pointer to the
223 /// call node. Also update NodeMap so that getValue(statepoint) will
224 /// reference lowered call result
226 lowerCallFromStatepoint(ImmutableStatepoint ISP, MachineBasicBlock *LandingPad,
227 SelectionDAGBuilder &Builder,
228 SmallVectorImpl<SDValue> &PendingExports) {
230 ImmutableCallSite CS(ISP.getCallSite());
232 SDValue ActualCallee = Builder.getValue(ISP.getActualCallee());
234 // Handle immediate and symbolic callees.
235 if (auto *ConstCallee = dyn_cast<ConstantSDNode>(ActualCallee.getNode()))
236 ActualCallee = Builder.DAG.getIntPtrConstant(ConstCallee->getZExtValue(),
237 Builder.getCurSDLoc(),
239 else if (auto *SymbolicCallee =
240 dyn_cast<GlobalAddressSDNode>(ActualCallee.getNode()))
241 ActualCallee = Builder.DAG.getTargetGlobalAddress(
242 SymbolicCallee->getGlobal(), SDLoc(SymbolicCallee),
243 SymbolicCallee->getValueType(0));
245 assert(CS.getCallingConv() != CallingConv::AnyReg &&
246 "anyregcc is not supported on statepoints!");
248 Type *DefTy = ISP.getActualReturnType();
249 bool HasDef = !DefTy->isVoidTy();
251 SDValue ReturnValue, CallEndVal;
252 std::tie(ReturnValue, CallEndVal) = Builder.lowerCallOperands(
253 ISP.getCallSite(), ImmutableStatepoint::CallArgsBeginPos,
254 ISP.getNumCallArgs(), ActualCallee, DefTy, LandingPad,
255 false /* IsPatchPoint */);
257 SDNode *CallEnd = CallEndVal.getNode();
259 // Get a call instruction from the call sequence chain. Tail calls are not
260 // allowed. The following code is essentially reverse engineering X86's
263 // We are expecting DAG to have the following form:
265 // ch = eh_label (only in case of invoke statepoint)
266 // ch, glue = callseq_start ch
267 // ch, glue = X86::Call ch, glue
268 // ch, glue = callseq_end ch, glue
269 // get_return_value ch, glue
271 // get_return_value can either be a CopyFromReg to grab the return value from
272 // %RAX, or it can be a LOAD to load a value returned by reference via a stack
275 if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg ||
276 CallEnd->getOpcode() == ISD::LOAD))
277 CallEnd = CallEnd->getOperand(0).getNode();
279 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && "expected!");
283 // Result value will be used in different basic block for invokes
284 // so we need to export it now. But statepoint call has a different type
285 // than the actuall call. It means that standart exporting mechanism will
286 // create register of the wrong type. So instead we need to create
287 // register with correct type and save value into it manually.
288 // TODO: To eliminate this problem we can remove gc.result intrinsics
289 // completelly and make statepoint call to return a tuple.
290 unsigned Reg = Builder.FuncInfo.CreateRegs(ISP.getActualReturnType());
291 RegsForValue RFV(*Builder.DAG.getContext(),
292 Builder.DAG.getTargetLoweringInfo(), Reg,
293 ISP.getActualReturnType());
294 SDValue Chain = Builder.DAG.getEntryNode();
296 RFV.getCopyToRegs(ReturnValue, Builder.DAG, Builder.getCurSDLoc(), Chain,
298 PendingExports.push_back(Chain);
299 Builder.FuncInfo.ValueMap[CS.getInstruction()] = Reg;
301 // The value of the statepoint itself will be the value of call itself.
302 // We'll replace the actually call node shortly. gc_result will grab
304 Builder.setValue(CS.getInstruction(), ReturnValue);
307 // The token value is never used from here on, just generate a poison value
308 Builder.setValue(CS.getInstruction(),
309 Builder.DAG.getIntPtrConstant(-1, Builder.getCurSDLoc()));
312 return CallEnd->getOperand(0).getNode();
315 /// Callect all gc pointers coming into statepoint intrinsic, clean them up,
316 /// and return two arrays:
317 /// Bases - base pointers incoming to this statepoint
318 /// Ptrs - derived pointers incoming to this statepoint
319 /// Relocs - the gc_relocate corresponding to each base/ptr pair
320 /// Elements of this arrays should be in one-to-one correspondence with each
321 /// other i.e Bases[i], Ptrs[i] are from the same gcrelocate call
322 static void getIncomingStatepointGCValues(
323 SmallVectorImpl<const Value *> &Bases, SmallVectorImpl<const Value *> &Ptrs,
324 SmallVectorImpl<const Value *> &Relocs, ImmutableStatepoint StatepointSite,
325 SelectionDAGBuilder &Builder) {
326 for (GCRelocateOperands relocateOpers :
327 StatepointSite.getRelocates(StatepointSite)) {
328 Relocs.push_back(relocateOpers.getUnderlyingCallSite().getInstruction());
329 Bases.push_back(relocateOpers.getBasePtr());
330 Ptrs.push_back(relocateOpers.getDerivedPtr());
333 // Remove any redundant llvm::Values which map to the same SDValue as another
334 // input. Also has the effect of removing duplicates in the original
335 // llvm::Value input list as well. This is a useful optimization for
336 // reducing the size of the StackMap section. It has no other impact.
337 removeDuplicatesGCPtrs(Bases, Ptrs, Relocs, Builder);
339 assert(Bases.size() == Ptrs.size() && Ptrs.size() == Relocs.size());
342 /// Spill a value incoming to the statepoint. It might be either part of
344 /// or gcstate. In both cases unconditionally spill it on the stack unless it
345 /// is a null constant. Return pair with first element being frame index
346 /// containing saved value and second element with outgoing chain from the
348 static std::pair<SDValue, SDValue>
349 spillIncomingStatepointValue(SDValue Incoming, SDValue Chain,
350 SelectionDAGBuilder &Builder) {
351 SDValue Loc = Builder.StatepointLowering.getLocation(Incoming);
353 // Emit new store if we didn't do it for this ptr before
354 if (!Loc.getNode()) {
355 Loc = Builder.StatepointLowering.allocateStackSlot(Incoming.getValueType(),
357 assert(isa<FrameIndexSDNode>(Loc));
358 int Index = cast<FrameIndexSDNode>(Loc)->getIndex();
359 // We use TargetFrameIndex so that isel will not select it into LEA
360 Loc = Builder.DAG.getTargetFrameIndex(Index, Incoming.getValueType());
362 // TODO: We can create TokenFactor node instead of
363 // chaining stores one after another, this may allow
364 // a bit more optimal scheduling for them
365 Chain = Builder.DAG.getStore(Chain, Builder.getCurSDLoc(), Incoming, Loc,
366 MachinePointerInfo::getFixedStack(Index),
369 Builder.StatepointLowering.setLocation(Incoming, Loc);
372 assert(Loc.getNode());
373 return std::make_pair(Loc, Chain);
376 /// Lower a single value incoming to a statepoint node. This value can be
377 /// either a deopt value or a gc value, the handling is the same. We special
378 /// case constants and allocas, then fall back to spilling if required.
379 static void lowerIncomingStatepointValue(SDValue Incoming,
380 SmallVectorImpl<SDValue> &Ops,
381 SelectionDAGBuilder &Builder) {
382 SDValue Chain = Builder.getRoot();
384 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Incoming)) {
385 // If the original value was a constant, make sure it gets recorded as
386 // such in the stackmap. This is required so that the consumer can
387 // parse any internal format to the deopt state. It also handles null
388 // pointers and other constant pointers in GC states
389 Ops.push_back(Builder.DAG.getTargetConstant(StackMaps::ConstantOp,
390 Builder.getCurSDLoc(),
392 Ops.push_back(Builder.DAG.getTargetConstant(C->getSExtValue(),
393 Builder.getCurSDLoc(),
395 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
396 // This handles allocas as arguments to the statepoint (this is only
397 // really meaningful for a deopt value. For GC, we'd be trying to
398 // relocate the address of the alloca itself?)
399 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
400 Incoming.getValueType()));
402 // Otherwise, locate a spill slot and explicitly spill it so it
403 // can be found by the runtime later. We currently do not support
404 // tracking values through callee saved registers to their eventual
405 // spill location. This would be a useful optimization, but would
406 // need to be optional since it requires a lot of complexity on the
407 // runtime side which not all would support.
408 std::pair<SDValue, SDValue> Res =
409 spillIncomingStatepointValue(Incoming, Chain, Builder);
410 Ops.push_back(Res.first);
414 Builder.DAG.setRoot(Chain);
417 /// Lower deopt state and gc pointer arguments of the statepoint. The actual
418 /// lowering is described in lowerIncomingStatepointValue. This function is
419 /// responsible for lowering everything in the right position and playing some
420 /// tricks to avoid redundant stack manipulation where possible. On
421 /// completion, 'Ops' will contain ready to use operands for machine code
422 /// statepoint. The chain nodes will have already been created and the DAG root
423 /// will be set to the last value spilled (if any were).
424 static void lowerStatepointMetaArgs(SmallVectorImpl<SDValue> &Ops,
425 ImmutableStatepoint StatepointSite,
426 SelectionDAGBuilder &Builder) {
428 // Lower the deopt and gc arguments for this statepoint. Layout will
429 // be: deopt argument length, deopt arguments.., gc arguments...
431 SmallVector<const Value *, 64> Bases, Ptrs, Relocations;
432 getIncomingStatepointGCValues(Bases, Ptrs, Relocations, StatepointSite,
436 // Check that each of the gc pointer and bases we've gotten out of the
437 // safepoint is something the strategy thinks might be a pointer into the GC
438 // heap. This is basically just here to help catch errors during statepoint
439 // insertion. TODO: This should actually be in the Verifier, but we can't get
440 // to the GCStrategy from there (yet).
441 GCStrategy &S = Builder.GFI->getStrategy();
442 for (const Value *V : Bases) {
443 auto Opt = S.isGCManagedPointer(V);
444 if (Opt.hasValue()) {
445 assert(Opt.getValue() &&
446 "non gc managed base pointer found in statepoint");
449 for (const Value *V : Ptrs) {
450 auto Opt = S.isGCManagedPointer(V);
451 if (Opt.hasValue()) {
452 assert(Opt.getValue() &&
453 "non gc managed derived pointer found in statepoint");
456 for (const Value *V : Relocations) {
457 auto Opt = S.isGCManagedPointer(V);
458 if (Opt.hasValue()) {
459 assert(Opt.getValue() && "non gc managed pointer relocated");
464 // Before we actually start lowering (and allocating spill slots for values),
465 // reserve any stack slots which we judge to be profitable to reuse for a
466 // particular value. This is purely an optimization over the code below and
467 // doesn't change semantics at all. It is important for performance that we
468 // reserve slots for both deopt and gc values before lowering either.
469 for (auto I = StatepointSite.vm_state_begin() + 1,
470 E = StatepointSite.vm_state_end();
473 SDValue Incoming = Builder.getValue(V);
474 reservePreviousStackSlotForValue(Incoming, Builder);
476 for (unsigned i = 0; i < Bases.size() * 2; ++i) {
477 // Even elements will contain base, odd elements - derived ptr
478 const Value *V = i % 2 ? Bases[i / 2] : Ptrs[i / 2];
479 SDValue Incoming = Builder.getValue(V);
480 reservePreviousStackSlotForValue(Incoming, Builder);
483 // First, prefix the list with the number of unique values to be
484 // lowered. Note that this is the number of *Values* not the
485 // number of SDValues required to lower them.
486 const int NumVMSArgs = StatepointSite.getNumTotalVMSArgs();
487 Ops.push_back( Builder.DAG.getTargetConstant(StackMaps::ConstantOp,
488 Builder.getCurSDLoc(),
490 Ops.push_back(Builder.DAG.getTargetConstant(NumVMSArgs, Builder.getCurSDLoc(),
493 assert(NumVMSArgs + 1 == std::distance(StatepointSite.vm_state_begin(),
494 StatepointSite.vm_state_end()));
496 // The vm state arguments are lowered in an opaque manner. We do
497 // not know what type of values are contained within. We skip the
498 // first one since that happens to be the total number we lowered
499 // explicitly just above. We could have left it in the loop and
500 // not done it explicitly, but it's far easier to understand this
502 for (auto I = StatepointSite.vm_state_begin() + 1,
503 E = StatepointSite.vm_state_end();
506 SDValue Incoming = Builder.getValue(V);
507 lowerIncomingStatepointValue(Incoming, Ops, Builder);
510 // Finally, go ahead and lower all the gc arguments. There's no prefixed
511 // length for this one. After lowering, we'll have the base and pointer
512 // arrays interwoven with each (lowered) base pointer immediately followed by
513 // it's (lowered) derived pointer. i.e
514 // (base[0], ptr[0], base[1], ptr[1], ...)
515 for (unsigned i = 0; i < Bases.size() * 2; ++i) {
516 // Even elements will contain base, odd elements - derived ptr
517 const Value *V = i % 2 ? Bases[i / 2] : Ptrs[i / 2];
518 SDValue Incoming = Builder.getValue(V);
519 lowerIncomingStatepointValue(Incoming, Ops, Builder);
522 // If there are any explicit spill slots passed to the statepoint, record
523 // them, but otherwise do not do anything special. These are user provided
524 // allocas and give control over placement to the consumer. In this case,
525 // it is the contents of the slot which may get updated, not the pointer to
527 for (Value *V : StatepointSite.gc_args()) {
528 SDValue Incoming = Builder.getValue(V);
529 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Incoming)) {
530 // This handles allocas as arguments to the statepoint
531 Ops.push_back(Builder.DAG.getTargetFrameIndex(FI->getIndex(),
532 Incoming.getValueType()));
537 void SelectionDAGBuilder::visitStatepoint(const CallInst &CI) {
538 // Check some preconditions for sanity
539 assert(isStatepoint(&CI) &&
540 "function called must be the statepoint function");
542 LowerStatepoint(ImmutableStatepoint(&CI));
545 void SelectionDAGBuilder::LowerStatepoint(
546 ImmutableStatepoint ISP, MachineBasicBlock *LandingPad /*=nullptr*/) {
547 // The basic scheme here is that information about both the original call and
548 // the safepoint is encoded in the CallInst. We create a temporary call and
549 // lower it, then reverse engineer the calling sequence.
553 StatepointLowering.startNewStatepoint(*this);
555 ImmutableCallSite CS(ISP.getCallSite());
559 for (const User *U : CS->users()) {
560 const CallInst *Call = cast<CallInst>(U);
561 if (isGCRelocate(Call))
562 StatepointLowering.scheduleRelocCall(*Call);
567 // If this is a malformed statepoint, report it early to simplify debugging.
568 // This should catch any IR level mistake that's made when constructing or
569 // transforming statepoints.
572 // Check that the associated GCStrategy expects to encounter statepoints.
573 // TODO: This if should become an assert. For now, we allow the GCStrategy
574 // to be optional for backwards compatibility. This will only last a short
575 // period (i.e. a couple of weeks).
576 assert(GFI->getStrategy().useStatepoints() &&
577 "GCStrategy does not expect to encounter statepoints");
580 // Lower statepoint vmstate and gcstate arguments
581 SmallVector<SDValue, 10> LoweredMetaArgs;
582 lowerStatepointMetaArgs(LoweredMetaArgs, ISP, *this);
584 // Get call node, we will replace it later with statepoint
586 lowerCallFromStatepoint(ISP, LandingPad, *this, PendingExports);
588 // Construct the actual GC_TRANSITION_START, STATEPOINT, and GC_TRANSITION_END
589 // nodes with all the appropriate arguments and return values.
591 // TODO: Currently, all of these operands are being marked as read/write in
592 // PrologEpilougeInserter.cpp, we should special case the VMState arguments
593 // and flags to be read-only.
594 SmallVector<SDValue, 40> Ops;
596 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
597 SDValue Chain = CallNode->getOperand(0);
600 bool CallHasIncomingGlue = CallNode->getGluedNode();
601 if (CallHasIncomingGlue) {
602 // Glue is always last operand
603 Glue = CallNode->getOperand(CallNode->getNumOperands() - 1);
606 // Build the GC_TRANSITION_START node if necessary.
608 // The operands to the GC_TRANSITION_{START,END} nodes are laid out in the
609 // order in which they appear in the call to the statepoint intrinsic. If
610 // any of the operands is a pointer-typed, that operand is immediately
611 // followed by a SRCVALUE for the pointer that may be used during lowering
612 // (e.g. to form MachinePointerInfo values for loads/stores).
613 const bool IsGCTransition =
614 (ISP.getFlags() & (uint64_t)StatepointFlags::GCTransition) ==
615 (uint64_t)StatepointFlags::GCTransition;
616 if (IsGCTransition) {
617 SmallVector<SDValue, 8> TSOps;
620 TSOps.push_back(Chain);
622 // Add GC transition arguments
623 for (auto I = ISP.gc_transition_args_begin() + 1,
624 E = ISP.gc_transition_args_end();
626 TSOps.push_back(getValue(*I));
627 if ((*I)->getType()->isPointerTy())
628 TSOps.push_back(DAG.getSrcValue(*I));
631 // Add glue if necessary
632 if (CallHasIncomingGlue)
633 TSOps.push_back(Glue);
635 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
637 SDValue GCTransitionStart =
638 DAG.getNode(ISD::GC_TRANSITION_START, getCurSDLoc(), NodeTys, TSOps);
640 Chain = GCTransitionStart.getValue(0);
641 Glue = GCTransitionStart.getValue(1);
644 // Calculate and push starting position of vmstate arguments
645 // Get number of arguments incoming directly into call node
646 unsigned NumCallRegArgs =
647 CallNode->getNumOperands() - (CallHasIncomingGlue ? 4 : 3);
648 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, getCurSDLoc(), MVT::i32));
651 SDValue CallTarget = SDValue(CallNode->getOperand(1).getNode(), 0);
652 Ops.push_back(CallTarget);
654 // Add call arguments
655 // Get position of register mask in the call
656 SDNode::op_iterator RegMaskIt;
657 if (CallHasIncomingGlue)
658 RegMaskIt = CallNode->op_end() - 2;
660 RegMaskIt = CallNode->op_end() - 1;
661 Ops.insert(Ops.end(), CallNode->op_begin() + 2, RegMaskIt);
663 // Add a leading constant argument with the Flags and the calling convention
665 CallingConv::ID CallConv = CS.getCallingConv();
666 uint64_t Flags = cast<ConstantInt>(CS.getArgument(2))->getZExtValue();
668 ((Flags & ~(uint64_t)StatepointFlags::MaskAll) == 0)
669 && "unknown flag used");
672 ((~(uint64_t)0 << Shift) & (uint64_t)StatepointFlags::MaskAll) == 0,
673 "shift width too small");
674 Ops.push_back(DAG.getTargetConstant(StackMaps::ConstantOp, getCurSDLoc(),
676 Ops.push_back(DAG.getTargetConstant(Flags | ((unsigned)CallConv << Shift),
677 getCurSDLoc(), MVT::i64));
679 // Insert all vmstate and gcstate arguments
680 Ops.insert(Ops.end(), LoweredMetaArgs.begin(), LoweredMetaArgs.end());
682 // Add register mask from call node
683 Ops.push_back(*RegMaskIt);
686 Ops.push_back(Chain);
688 // Same for the glue, but we add it only if original call had it
692 // Compute return values. Provide a glue output since we consume one as
693 // input. This allows someone else to chain off us as needed.
694 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
696 SDNode *StatepointMCNode =
697 DAG.getMachineNode(TargetOpcode::STATEPOINT, getCurSDLoc(), NodeTys, Ops);
699 SDNode *SinkNode = StatepointMCNode;
701 // Build the GC_TRANSITION_END node if necessary.
703 // See the comment above regarding GC_TRANSITION_START for the layout of
704 // the operands to the GC_TRANSITION_END node.
705 if (IsGCTransition) {
706 SmallVector<SDValue, 8> TEOps;
709 TEOps.push_back(SDValue(StatepointMCNode, 0));
711 // Add GC transition arguments
712 for (auto I = ISP.gc_transition_args_begin() + 1,
713 E = ISP.gc_transition_args_end();
715 TEOps.push_back(getValue(*I));
716 if ((*I)->getType()->isPointerTy())
717 TEOps.push_back(DAG.getSrcValue(*I));
721 TEOps.push_back(SDValue(StatepointMCNode, 1));
723 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
725 SDValue GCTransitionStart =
726 DAG.getNode(ISD::GC_TRANSITION_END, getCurSDLoc(), NodeTys, TEOps);
728 SinkNode = GCTransitionStart.getNode();
731 // Replace original call
732 DAG.ReplaceAllUsesWith(CallNode, SinkNode); // This may update Root
733 // Remove originall call node
734 DAG.DeleteNode(CallNode);
736 // DON'T set the root - under the assumption that it's already set past the
737 // inserted node we created.
739 // TODO: A better future implementation would be to emit a single variable
740 // argument, variable return value STATEPOINT node here and then hookup the
741 // return value of each gc.relocate to the respective output of the
742 // previously emitted STATEPOINT value. Unfortunately, this doesn't appear
743 // to actually be possible today.
746 void SelectionDAGBuilder::visitGCResult(const CallInst &CI) {
747 // The result value of the gc_result is simply the result of the actual
748 // call. We've already emitted this, so just grab the value.
749 Instruction *I = cast<Instruction>(CI.getArgOperand(0));
750 assert(isStatepoint(I) && "first argument must be a statepoint token");
752 if (isa<InvokeInst>(I)) {
753 // For invokes we should have stored call result in a virtual register.
754 // We can not use default getValue() functionality to copy value from this
755 // register because statepoint and actuall call return types can be
756 // different, and getValue() will use CopyFromReg of the wrong type,
757 // which is always i32 in our case.
758 PointerType *CalleeType =
759 cast<PointerType>(ImmutableStatepoint(I).getActualCallee()->getType());
761 cast<FunctionType>(CalleeType->getElementType())->getReturnType();
762 SDValue CopyFromReg = getCopyFromRegs(I, RetTy);
764 assert(CopyFromReg.getNode());
765 setValue(&CI, CopyFromReg);
767 setValue(&CI, getValue(I));
771 void SelectionDAGBuilder::visitGCRelocate(const CallInst &CI) {
774 StatepointLowering.relocCallVisited(CI);
777 GCRelocateOperands relocateOpers(&CI);
778 SDValue SD = getValue(relocateOpers.getDerivedPtr());
780 if (isa<ConstantSDNode>(SD) || isa<FrameIndexSDNode>(SD)) {
781 // We didn't need to spill these special cases (constants and allocas).
782 // See the handling in spillIncomingValueForStatepoint for detail.
787 SDValue Loc = StatepointLowering.getRelocLocation(SD);
788 // Emit new load if we did not emit it before
789 if (!Loc.getNode()) {
790 SDValue SpillSlot = StatepointLowering.getLocation(SD);
791 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
793 // Be conservative: flush all pending loads
794 // TODO: Probably we can be less restrictive on this,
795 // it may allow more scheduling opprtunities
796 SDValue Chain = getRoot();
798 Loc = DAG.getLoad(SpillSlot.getValueType(), getCurSDLoc(), Chain, SpillSlot,
799 MachinePointerInfo::getFixedStack(FI), false, false,
802 StatepointLowering.setRelocLocation(SD, Loc);
804 // Again, be conservative, don't emit pending loads
805 DAG.setRoot(Loc.getValue(1));
808 assert(Loc.getNode());