SelectionDAGBuilder.h revision 353358
1//===- SelectionDAGBuilder.h - Selection-DAG building -----------*- C++ -*-===//
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 implements routines for translating from LLVM IR into SelectionDAG IR.
10//
11//===----------------------------------------------------------------------===//
12
13#ifndef LLVM_LIB_CODEGEN_SELECTIONDAG_SELECTIONDAGBUILDER_H
14#define LLVM_LIB_CODEGEN_SELECTIONDAG_SELECTIONDAGBUILDER_H
15
16#include "StatepointLowering.h"
17#include "llvm/ADT/APInt.h"
18#include "llvm/ADT/ArrayRef.h"
19#include "llvm/ADT/DenseMap.h"
20#include "llvm/ADT/MapVector.h"
21#include "llvm/ADT/SmallVector.h"
22#include "llvm/Analysis/AliasAnalysis.h"
23#include "llvm/CodeGen/ISDOpcodes.h"
24#include "llvm/CodeGen/SelectionDAG.h"
25#include "llvm/CodeGen/SelectionDAGNodes.h"
26#include "llvm/CodeGen/SwitchLoweringUtils.h"
27#include "llvm/CodeGen/TargetLowering.h"
28#include "llvm/CodeGen/ValueTypes.h"
29#include "llvm/IR/CallSite.h"
30#include "llvm/IR/DebugLoc.h"
31#include "llvm/IR/Instruction.h"
32#include "llvm/IR/Statepoint.h"
33#include "llvm/Support/BranchProbability.h"
34#include "llvm/Support/CodeGen.h"
35#include "llvm/Support/ErrorHandling.h"
36#include "llvm/Support/MachineValueType.h"
37#include <algorithm>
38#include <cassert>
39#include <cstdint>
40#include <utility>
41#include <vector>
42
43namespace llvm {
44
45class AllocaInst;
46class AtomicCmpXchgInst;
47class AtomicRMWInst;
48class BasicBlock;
49class BranchInst;
50class CallInst;
51class CallBrInst;
52class CatchPadInst;
53class CatchReturnInst;
54class CatchSwitchInst;
55class CleanupPadInst;
56class CleanupReturnInst;
57class Constant;
58class ConstantInt;
59class ConstrainedFPIntrinsic;
60class DbgValueInst;
61class DataLayout;
62class DIExpression;
63class DILocalVariable;
64class DILocation;
65class FenceInst;
66class FunctionLoweringInfo;
67class GCFunctionInfo;
68class GCRelocateInst;
69class GCResultInst;
70class IndirectBrInst;
71class InvokeInst;
72class LandingPadInst;
73class LLVMContext;
74class LoadInst;
75class MachineBasicBlock;
76class PHINode;
77class ResumeInst;
78class ReturnInst;
79class SDDbgValue;
80class StoreInst;
81class SwiftErrorValueTracking;
82class SwitchInst;
83class TargetLibraryInfo;
84class TargetMachine;
85class Type;
86class VAArgInst;
87class UnreachableInst;
88class Use;
89class User;
90class Value;
91
92//===----------------------------------------------------------------------===//
93/// SelectionDAGBuilder - This is the common target-independent lowering
94/// implementation that is parameterized by a TargetLowering object.
95///
96class SelectionDAGBuilder {
97  /// The current instruction being visited.
98  const Instruction *CurInst = nullptr;
99
100  DenseMap<const Value*, SDValue> NodeMap;
101
102  /// Maps argument value for unused arguments. This is used
103  /// to preserve debug information for incoming arguments.
104  DenseMap<const Value*, SDValue> UnusedArgNodeMap;
105
106  /// Helper type for DanglingDebugInfoMap.
107  class DanglingDebugInfo {
108    const DbgValueInst* DI = nullptr;
109    DebugLoc dl;
110    unsigned SDNodeOrder = 0;
111
112  public:
113    DanglingDebugInfo() = default;
114    DanglingDebugInfo(const DbgValueInst *di, DebugLoc DL, unsigned SDNO)
115        : DI(di), dl(std::move(DL)), SDNodeOrder(SDNO) {}
116
117    const DbgValueInst* getDI() { return DI; }
118    DebugLoc getdl() { return dl; }
119    unsigned getSDNodeOrder() { return SDNodeOrder; }
120  };
121
122  /// Helper type for DanglingDebugInfoMap.
123  typedef std::vector<DanglingDebugInfo> DanglingDebugInfoVector;
124
125  /// Keeps track of dbg_values for which we have not yet seen the referent.
126  /// We defer handling these until we do see it.
127  MapVector<const Value*, DanglingDebugInfoVector> DanglingDebugInfoMap;
128
129public:
130  /// Loads are not emitted to the program immediately.  We bunch them up and
131  /// then emit token factor nodes when possible.  This allows us to get simple
132  /// disambiguation between loads without worrying about alias analysis.
133  SmallVector<SDValue, 8> PendingLoads;
134
135  /// State used while lowering a statepoint sequence (gc_statepoint,
136  /// gc_relocate, and gc_result).  See StatepointLowering.hpp/cpp for details.
137  StatepointLoweringState StatepointLowering;
138
139private:
140  /// CopyToReg nodes that copy values to virtual registers for export to other
141  /// blocks need to be emitted before any terminator instruction, but they have
142  /// no other ordering requirements. We bunch them up and the emit a single
143  /// tokenfactor for them just before terminator instructions.
144  SmallVector<SDValue, 8> PendingExports;
145
146  /// A unique monotonically increasing number used to order the SDNodes we
147  /// create.
148  unsigned SDNodeOrder;
149
150  /// Determine the rank by weight of CC in [First,Last]. If CC has more weight
151  /// than each cluster in the range, its rank is 0.
152  unsigned caseClusterRank(const SwitchCG::CaseCluster &CC,
153                           SwitchCG::CaseClusterIt First,
154                           SwitchCG::CaseClusterIt Last);
155
156  /// Emit comparison and split W into two subtrees.
157  void splitWorkItem(SwitchCG::SwitchWorkList &WorkList,
158                     const SwitchCG::SwitchWorkListItem &W, Value *Cond,
159                     MachineBasicBlock *SwitchMBB);
160
161  /// Lower W.
162  void lowerWorkItem(SwitchCG::SwitchWorkListItem W, Value *Cond,
163                     MachineBasicBlock *SwitchMBB,
164                     MachineBasicBlock *DefaultMBB);
165
166  /// Peel the top probability case if it exceeds the threshold
167  MachineBasicBlock *
168  peelDominantCaseCluster(const SwitchInst &SI,
169                          SwitchCG::CaseClusterVector &Clusters,
170                          BranchProbability &PeeledCaseProb);
171
172  /// A class which encapsulates all of the information needed to generate a
173  /// stack protector check and signals to isel via its state being initialized
174  /// that a stack protector needs to be generated.
175  ///
176  /// *NOTE* The following is a high level documentation of SelectionDAG Stack
177  /// Protector Generation. The reason that it is placed here is for a lack of
178  /// other good places to stick it.
179  ///
180  /// High Level Overview of SelectionDAG Stack Protector Generation:
181  ///
182  /// Previously, generation of stack protectors was done exclusively in the
183  /// pre-SelectionDAG Codegen LLVM IR Pass "Stack Protector". This necessitated
184  /// splitting basic blocks at the IR level to create the success/failure basic
185  /// blocks in the tail of the basic block in question. As a result of this,
186  /// calls that would have qualified for the sibling call optimization were no
187  /// longer eligible for optimization since said calls were no longer right in
188  /// the "tail position" (i.e. the immediate predecessor of a ReturnInst
189  /// instruction).
190  ///
191  /// Then it was noticed that since the sibling call optimization causes the
192  /// callee to reuse the caller's stack, if we could delay the generation of
193  /// the stack protector check until later in CodeGen after the sibling call
194  /// decision was made, we get both the tail call optimization and the stack
195  /// protector check!
196  ///
197  /// A few goals in solving this problem were:
198  ///
199  ///   1. Preserve the architecture independence of stack protector generation.
200  ///
201  ///   2. Preserve the normal IR level stack protector check for platforms like
202  ///      OpenBSD for which we support platform-specific stack protector
203  ///      generation.
204  ///
205  /// The main problem that guided the present solution is that one can not
206  /// solve this problem in an architecture independent manner at the IR level
207  /// only. This is because:
208  ///
209  ///   1. The decision on whether or not to perform a sibling call on certain
210  ///      platforms (for instance i386) requires lower level information
211  ///      related to available registers that can not be known at the IR level.
212  ///
213  ///   2. Even if the previous point were not true, the decision on whether to
214  ///      perform a tail call is done in LowerCallTo in SelectionDAG which
215  ///      occurs after the Stack Protector Pass. As a result, one would need to
216  ///      put the relevant callinst into the stack protector check success
217  ///      basic block (where the return inst is placed) and then move it back
218  ///      later at SelectionDAG/MI time before the stack protector check if the
219  ///      tail call optimization failed. The MI level option was nixed
220  ///      immediately since it would require platform-specific pattern
221  ///      matching. The SelectionDAG level option was nixed because
222  ///      SelectionDAG only processes one IR level basic block at a time
223  ///      implying one could not create a DAG Combine to move the callinst.
224  ///
225  /// To get around this problem a few things were realized:
226  ///
227  ///   1. While one can not handle multiple IR level basic blocks at the
228  ///      SelectionDAG Level, one can generate multiple machine basic blocks
229  ///      for one IR level basic block. This is how we handle bit tests and
230  ///      switches.
231  ///
232  ///   2. At the MI level, tail calls are represented via a special return
233  ///      MIInst called "tcreturn". Thus if we know the basic block in which we
234  ///      wish to insert the stack protector check, we get the correct behavior
235  ///      by always inserting the stack protector check right before the return
236  ///      statement. This is a "magical transformation" since no matter where
237  ///      the stack protector check intrinsic is, we always insert the stack
238  ///      protector check code at the end of the BB.
239  ///
240  /// Given the aforementioned constraints, the following solution was devised:
241  ///
242  ///   1. On platforms that do not support SelectionDAG stack protector check
243  ///      generation, allow for the normal IR level stack protector check
244  ///      generation to continue.
245  ///
246  ///   2. On platforms that do support SelectionDAG stack protector check
247  ///      generation:
248  ///
249  ///     a. Use the IR level stack protector pass to decide if a stack
250  ///        protector is required/which BB we insert the stack protector check
251  ///        in by reusing the logic already therein. If we wish to generate a
252  ///        stack protector check in a basic block, we place a special IR
253  ///        intrinsic called llvm.stackprotectorcheck right before the BB's
254  ///        returninst or if there is a callinst that could potentially be
255  ///        sibling call optimized, before the call inst.
256  ///
257  ///     b. Then when a BB with said intrinsic is processed, we codegen the BB
258  ///        normally via SelectBasicBlock. In said process, when we visit the
259  ///        stack protector check, we do not actually emit anything into the
260  ///        BB. Instead, we just initialize the stack protector descriptor
261  ///        class (which involves stashing information/creating the success
262  ///        mbbb and the failure mbb if we have not created one for this
263  ///        function yet) and export the guard variable that we are going to
264  ///        compare.
265  ///
266  ///     c. After we finish selecting the basic block, in FinishBasicBlock if
267  ///        the StackProtectorDescriptor attached to the SelectionDAGBuilder is
268  ///        initialized, we produce the validation code with one of these
269  ///        techniques:
270  ///          1) with a call to a guard check function
271  ///          2) with inlined instrumentation
272  ///
273  ///        1) We insert a call to the check function before the terminator.
274  ///
275  ///        2) We first find a splice point in the parent basic block
276  ///        before the terminator and then splice the terminator of said basic
277  ///        block into the success basic block. Then we code-gen a new tail for
278  ///        the parent basic block consisting of the two loads, the comparison,
279  ///        and finally two branches to the success/failure basic blocks. We
280  ///        conclude by code-gening the failure basic block if we have not
281  ///        code-gened it already (all stack protector checks we generate in
282  ///        the same function, use the same failure basic block).
283  class StackProtectorDescriptor {
284  public:
285    StackProtectorDescriptor() = default;
286
287    /// Returns true if all fields of the stack protector descriptor are
288    /// initialized implying that we should/are ready to emit a stack protector.
289    bool shouldEmitStackProtector() const {
290      return ParentMBB && SuccessMBB && FailureMBB;
291    }
292
293    bool shouldEmitFunctionBasedCheckStackProtector() const {
294      return ParentMBB && !SuccessMBB && !FailureMBB;
295    }
296
297    /// Initialize the stack protector descriptor structure for a new basic
298    /// block.
299    void initialize(const BasicBlock *BB, MachineBasicBlock *MBB,
300                    bool FunctionBasedInstrumentation) {
301      // Make sure we are not initialized yet.
302      assert(!shouldEmitStackProtector() && "Stack Protector Descriptor is "
303             "already initialized!");
304      ParentMBB = MBB;
305      if (!FunctionBasedInstrumentation) {
306        SuccessMBB = AddSuccessorMBB(BB, MBB, /* IsLikely */ true);
307        FailureMBB = AddSuccessorMBB(BB, MBB, /* IsLikely */ false, FailureMBB);
308      }
309    }
310
311    /// Reset state that changes when we handle different basic blocks.
312    ///
313    /// This currently includes:
314    ///
315    /// 1. The specific basic block we are generating a
316    /// stack protector for (ParentMBB).
317    ///
318    /// 2. The successor machine basic block that will contain the tail of
319    /// parent mbb after we create the stack protector check (SuccessMBB). This
320    /// BB is visited only on stack protector check success.
321    void resetPerBBState() {
322      ParentMBB = nullptr;
323      SuccessMBB = nullptr;
324    }
325
326    /// Reset state that only changes when we switch functions.
327    ///
328    /// This currently includes:
329    ///
330    /// 1. FailureMBB since we reuse the failure code path for all stack
331    /// protector checks created in an individual function.
332    ///
333    /// 2.The guard variable since the guard variable we are checking against is
334    /// always the same.
335    void resetPerFunctionState() {
336      FailureMBB = nullptr;
337    }
338
339    MachineBasicBlock *getParentMBB() { return ParentMBB; }
340    MachineBasicBlock *getSuccessMBB() { return SuccessMBB; }
341    MachineBasicBlock *getFailureMBB() { return FailureMBB; }
342
343  private:
344    /// The basic block for which we are generating the stack protector.
345    ///
346    /// As a result of stack protector generation, we will splice the
347    /// terminators of this basic block into the successor mbb SuccessMBB and
348    /// replace it with a compare/branch to the successor mbbs
349    /// SuccessMBB/FailureMBB depending on whether or not the stack protector
350    /// was violated.
351    MachineBasicBlock *ParentMBB = nullptr;
352
353    /// A basic block visited on stack protector check success that contains the
354    /// terminators of ParentMBB.
355    MachineBasicBlock *SuccessMBB = nullptr;
356
357    /// This basic block visited on stack protector check failure that will
358    /// contain a call to __stack_chk_fail().
359    MachineBasicBlock *FailureMBB = nullptr;
360
361    /// Add a successor machine basic block to ParentMBB. If the successor mbb
362    /// has not been created yet (i.e. if SuccMBB = 0), then the machine basic
363    /// block will be created. Assign a large weight if IsLikely is true.
364    MachineBasicBlock *AddSuccessorMBB(const BasicBlock *BB,
365                                       MachineBasicBlock *ParentMBB,
366                                       bool IsLikely,
367                                       MachineBasicBlock *SuccMBB = nullptr);
368  };
369
370private:
371  const TargetMachine &TM;
372
373public:
374  /// Lowest valid SDNodeOrder. The special case 0 is reserved for scheduling
375  /// nodes without a corresponding SDNode.
376  static const unsigned LowestSDNodeOrder = 1;
377
378  SelectionDAG &DAG;
379  const DataLayout *DL = nullptr;
380  AliasAnalysis *AA = nullptr;
381  const TargetLibraryInfo *LibInfo;
382
383  class SDAGSwitchLowering : public SwitchCG::SwitchLowering {
384  public:
385    SDAGSwitchLowering(SelectionDAGBuilder *sdb, FunctionLoweringInfo &funcinfo)
386        : SwitchCG::SwitchLowering(funcinfo), SDB(sdb) {}
387
388    virtual void addSuccessorWithProb(
389        MachineBasicBlock *Src, MachineBasicBlock *Dst,
390        BranchProbability Prob = BranchProbability::getUnknown()) override {
391      SDB->addSuccessorWithProb(Src, Dst, Prob);
392    }
393
394  private:
395    SelectionDAGBuilder *SDB;
396  };
397
398  std::unique_ptr<SDAGSwitchLowering> SL;
399
400  /// A StackProtectorDescriptor structure used to communicate stack protector
401  /// information in between SelectBasicBlock and FinishBasicBlock.
402  StackProtectorDescriptor SPDescriptor;
403
404  // Emit PHI-node-operand constants only once even if used by multiple
405  // PHI nodes.
406  DenseMap<const Constant *, unsigned> ConstantsOut;
407
408  /// Information about the function as a whole.
409  FunctionLoweringInfo &FuncInfo;
410
411  /// Information about the swifterror values used throughout the function.
412  SwiftErrorValueTracking &SwiftError;
413
414  /// Garbage collection metadata for the function.
415  GCFunctionInfo *GFI;
416
417  /// Map a landing pad to the call site indexes.
418  DenseMap<MachineBasicBlock *, SmallVector<unsigned, 4>> LPadToCallSiteMap;
419
420  /// This is set to true if a call in the current block has been translated as
421  /// a tail call. In this case, no subsequent DAG nodes should be created.
422  bool HasTailCall = false;
423
424  LLVMContext *Context;
425
426  SelectionDAGBuilder(SelectionDAG &dag, FunctionLoweringInfo &funcinfo,
427                      SwiftErrorValueTracking &swifterror, CodeGenOpt::Level ol)
428      : SDNodeOrder(LowestSDNodeOrder), TM(dag.getTarget()), DAG(dag),
429        SL(make_unique<SDAGSwitchLowering>(this, funcinfo)), FuncInfo(funcinfo),
430        SwiftError(swifterror) {}
431
432  void init(GCFunctionInfo *gfi, AliasAnalysis *AA,
433            const TargetLibraryInfo *li);
434
435  /// Clear out the current SelectionDAG and the associated state and prepare
436  /// this SelectionDAGBuilder object to be used for a new block. This doesn't
437  /// clear out information about additional blocks that are needed to complete
438  /// switch lowering or PHI node updating; that information is cleared out as
439  /// it is consumed.
440  void clear();
441
442  /// Clear the dangling debug information map. This function is separated from
443  /// the clear so that debug information that is dangling in a basic block can
444  /// be properly resolved in a different basic block. This allows the
445  /// SelectionDAG to resolve dangling debug information attached to PHI nodes.
446  void clearDanglingDebugInfo();
447
448  /// Return the current virtual root of the Selection DAG, flushing any
449  /// PendingLoad items. This must be done before emitting a store or any other
450  /// node that may need to be ordered after any prior load instructions.
451  SDValue getRoot();
452
453  /// Similar to getRoot, but instead of flushing all the PendingLoad items,
454  /// flush all the PendingExports items. It is necessary to do this before
455  /// emitting a terminator instruction.
456  SDValue getControlRoot();
457
458  SDLoc getCurSDLoc() const {
459    return SDLoc(CurInst, SDNodeOrder);
460  }
461
462  DebugLoc getCurDebugLoc() const {
463    return CurInst ? CurInst->getDebugLoc() : DebugLoc();
464  }
465
466  void CopyValueToVirtualRegister(const Value *V, unsigned Reg);
467
468  void visit(const Instruction &I);
469
470  void visit(unsigned Opcode, const User &I);
471
472  /// If there was virtual register allocated for the value V emit CopyFromReg
473  /// of the specified type Ty. Return empty SDValue() otherwise.
474  SDValue getCopyFromRegs(const Value *V, Type *Ty);
475
476  /// If we have dangling debug info that describes \p Variable, or an
477  /// overlapping part of variable considering the \p Expr, then this method
478  /// will drop that debug info as it isn't valid any longer.
479  void dropDanglingDebugInfo(const DILocalVariable *Variable,
480                             const DIExpression *Expr);
481
482  /// If we saw an earlier dbg_value referring to V, generate the debug data
483  /// structures now that we've seen its definition.
484  void resolveDanglingDebugInfo(const Value *V, SDValue Val);
485
486  /// For the given dangling debuginfo record, perform last-ditch efforts to
487  /// resolve the debuginfo to something that is represented in this DAG. If
488  /// this cannot be done, produce an Undef debug value record.
489  void salvageUnresolvedDbgValue(DanglingDebugInfo &DDI);
490
491  /// For a given Value, attempt to create and record a SDDbgValue in the
492  /// SelectionDAG.
493  bool handleDebugValue(const Value *V, DILocalVariable *Var,
494                        DIExpression *Expr, DebugLoc CurDL,
495                        DebugLoc InstDL, unsigned Order);
496
497  /// Evict any dangling debug information, attempting to salvage it first.
498  void resolveOrClearDbgInfo();
499
500  SDValue getValue(const Value *V);
501  bool findValue(const Value *V) const;
502
503  /// Return the SDNode for the specified IR value if it exists.
504  SDNode *getNodeForIRValue(const Value *V) {
505    if (NodeMap.find(V) == NodeMap.end())
506      return nullptr;
507    return NodeMap[V].getNode();
508  }
509
510  SDValue getNonRegisterValue(const Value *V);
511  SDValue getValueImpl(const Value *V);
512
513  void setValue(const Value *V, SDValue NewN) {
514    SDValue &N = NodeMap[V];
515    assert(!N.getNode() && "Already set a value for this node!");
516    N = NewN;
517  }
518
519  void setUnusedArgValue(const Value *V, SDValue NewN) {
520    SDValue &N = UnusedArgNodeMap[V];
521    assert(!N.getNode() && "Already set a value for this node!");
522    N = NewN;
523  }
524
525  void FindMergedConditions(const Value *Cond, MachineBasicBlock *TBB,
526                            MachineBasicBlock *FBB, MachineBasicBlock *CurBB,
527                            MachineBasicBlock *SwitchBB,
528                            Instruction::BinaryOps Opc, BranchProbability TProb,
529                            BranchProbability FProb, bool InvertCond);
530  void EmitBranchForMergedCondition(const Value *Cond, MachineBasicBlock *TBB,
531                                    MachineBasicBlock *FBB,
532                                    MachineBasicBlock *CurBB,
533                                    MachineBasicBlock *SwitchBB,
534                                    BranchProbability TProb, BranchProbability FProb,
535                                    bool InvertCond);
536  bool ShouldEmitAsBranches(const std::vector<SwitchCG::CaseBlock> &Cases);
537  bool isExportableFromCurrentBlock(const Value *V, const BasicBlock *FromBB);
538  void CopyToExportRegsIfNeeded(const Value *V);
539  void ExportFromCurrentBlock(const Value *V);
540  void LowerCallTo(ImmutableCallSite CS, SDValue Callee, bool IsTailCall,
541                   const BasicBlock *EHPadBB = nullptr);
542
543  // Lower range metadata from 0 to N to assert zext to an integer of nearest
544  // floor power of two.
545  SDValue lowerRangeToAssertZExt(SelectionDAG &DAG, const Instruction &I,
546                                 SDValue Op);
547
548  void populateCallLoweringInfo(TargetLowering::CallLoweringInfo &CLI,
549                                const CallBase *Call, unsigned ArgIdx,
550                                unsigned NumArgs, SDValue Callee,
551                                Type *ReturnTy, bool IsPatchPoint);
552
553  std::pair<SDValue, SDValue>
554  lowerInvokable(TargetLowering::CallLoweringInfo &CLI,
555                 const BasicBlock *EHPadBB = nullptr);
556
557  /// When an MBB was split during scheduling, update the
558  /// references that need to refer to the last resulting block.
559  void UpdateSplitBlock(MachineBasicBlock *First, MachineBasicBlock *Last);
560
561  /// Describes a gc.statepoint or a gc.statepoint like thing for the purposes
562  /// of lowering into a STATEPOINT node.
563  struct StatepointLoweringInfo {
564    /// Bases[i] is the base pointer for Ptrs[i].  Together they denote the set
565    /// of gc pointers this STATEPOINT has to relocate.
566    SmallVector<const Value *, 16> Bases;
567    SmallVector<const Value *, 16> Ptrs;
568
569    /// The set of gc.relocate calls associated with this gc.statepoint.
570    SmallVector<const GCRelocateInst *, 16> GCRelocates;
571
572    /// The full list of gc arguments to the gc.statepoint being lowered.
573    ArrayRef<const Use> GCArgs;
574
575    /// The gc.statepoint instruction.
576    const Instruction *StatepointInstr = nullptr;
577
578    /// The list of gc transition arguments present in the gc.statepoint being
579    /// lowered.
580    ArrayRef<const Use> GCTransitionArgs;
581
582    /// The ID that the resulting STATEPOINT instruction has to report.
583    unsigned ID = -1;
584
585    /// Information regarding the underlying call instruction.
586    TargetLowering::CallLoweringInfo CLI;
587
588    /// The deoptimization state associated with this gc.statepoint call, if
589    /// any.
590    ArrayRef<const Use> DeoptState;
591
592    /// Flags associated with the meta arguments being lowered.
593    uint64_t StatepointFlags = -1;
594
595    /// The number of patchable bytes the call needs to get lowered into.
596    unsigned NumPatchBytes = -1;
597
598    /// The exception handling unwind destination, in case this represents an
599    /// invoke of gc.statepoint.
600    const BasicBlock *EHPadBB = nullptr;
601
602    explicit StatepointLoweringInfo(SelectionDAG &DAG) : CLI(DAG) {}
603  };
604
605  /// Lower \p SLI into a STATEPOINT instruction.
606  SDValue LowerAsSTATEPOINT(StatepointLoweringInfo &SI);
607
608  // This function is responsible for the whole statepoint lowering process.
609  // It uniformly handles invoke and call statepoints.
610  void LowerStatepoint(ImmutableStatepoint ISP,
611                       const BasicBlock *EHPadBB = nullptr);
612
613  void LowerCallSiteWithDeoptBundle(const CallBase *Call, SDValue Callee,
614                                    const BasicBlock *EHPadBB);
615
616  void LowerDeoptimizeCall(const CallInst *CI);
617  void LowerDeoptimizingReturn();
618
619  void LowerCallSiteWithDeoptBundleImpl(const CallBase *Call, SDValue Callee,
620                                        const BasicBlock *EHPadBB,
621                                        bool VarArgDisallowed,
622                                        bool ForceVoidReturnTy);
623
624  /// Returns the type of FrameIndex and TargetFrameIndex nodes.
625  MVT getFrameIndexTy() {
626    return DAG.getTargetLoweringInfo().getFrameIndexTy(DAG.getDataLayout());
627  }
628
629private:
630  // Terminator instructions.
631  void visitRet(const ReturnInst &I);
632  void visitBr(const BranchInst &I);
633  void visitSwitch(const SwitchInst &I);
634  void visitIndirectBr(const IndirectBrInst &I);
635  void visitUnreachable(const UnreachableInst &I);
636  void visitCleanupRet(const CleanupReturnInst &I);
637  void visitCatchSwitch(const CatchSwitchInst &I);
638  void visitCatchRet(const CatchReturnInst &I);
639  void visitCatchPad(const CatchPadInst &I);
640  void visitCleanupPad(const CleanupPadInst &CPI);
641
642  BranchProbability getEdgeProbability(const MachineBasicBlock *Src,
643                                       const MachineBasicBlock *Dst) const;
644  void addSuccessorWithProb(
645      MachineBasicBlock *Src, MachineBasicBlock *Dst,
646      BranchProbability Prob = BranchProbability::getUnknown());
647
648public:
649  void visitSwitchCase(SwitchCG::CaseBlock &CB, MachineBasicBlock *SwitchBB);
650  void visitSPDescriptorParent(StackProtectorDescriptor &SPD,
651                               MachineBasicBlock *ParentBB);
652  void visitSPDescriptorFailure(StackProtectorDescriptor &SPD);
653  void visitBitTestHeader(SwitchCG::BitTestBlock &B,
654                          MachineBasicBlock *SwitchBB);
655  void visitBitTestCase(SwitchCG::BitTestBlock &BB, MachineBasicBlock *NextMBB,
656                        BranchProbability BranchProbToNext, unsigned Reg,
657                        SwitchCG::BitTestCase &B, MachineBasicBlock *SwitchBB);
658  void visitJumpTable(SwitchCG::JumpTable &JT);
659  void visitJumpTableHeader(SwitchCG::JumpTable &JT,
660                            SwitchCG::JumpTableHeader &JTH,
661                            MachineBasicBlock *SwitchBB);
662
663private:
664  // These all get lowered before this pass.
665  void visitInvoke(const InvokeInst &I);
666  void visitCallBr(const CallBrInst &I);
667  void visitResume(const ResumeInst &I);
668
669  void visitUnary(const User &I, unsigned Opcode);
670  void visitFNeg(const User &I) { visitUnary(I, ISD::FNEG); }
671
672  void visitBinary(const User &I, unsigned Opcode);
673  void visitShift(const User &I, unsigned Opcode);
674  void visitAdd(const User &I)  { visitBinary(I, ISD::ADD); }
675  void visitFAdd(const User &I) { visitBinary(I, ISD::FADD); }
676  void visitSub(const User &I)  { visitBinary(I, ISD::SUB); }
677  void visitFSub(const User &I);
678  void visitMul(const User &I)  { visitBinary(I, ISD::MUL); }
679  void visitFMul(const User &I) { visitBinary(I, ISD::FMUL); }
680  void visitURem(const User &I) { visitBinary(I, ISD::UREM); }
681  void visitSRem(const User &I) { visitBinary(I, ISD::SREM); }
682  void visitFRem(const User &I) { visitBinary(I, ISD::FREM); }
683  void visitUDiv(const User &I) { visitBinary(I, ISD::UDIV); }
684  void visitSDiv(const User &I);
685  void visitFDiv(const User &I) { visitBinary(I, ISD::FDIV); }
686  void visitAnd (const User &I) { visitBinary(I, ISD::AND); }
687  void visitOr  (const User &I) { visitBinary(I, ISD::OR); }
688  void visitXor (const User &I) { visitBinary(I, ISD::XOR); }
689  void visitShl (const User &I) { visitShift(I, ISD::SHL); }
690  void visitLShr(const User &I) { visitShift(I, ISD::SRL); }
691  void visitAShr(const User &I) { visitShift(I, ISD::SRA); }
692  void visitICmp(const User &I);
693  void visitFCmp(const User &I);
694  // Visit the conversion instructions
695  void visitTrunc(const User &I);
696  void visitZExt(const User &I);
697  void visitSExt(const User &I);
698  void visitFPTrunc(const User &I);
699  void visitFPExt(const User &I);
700  void visitFPToUI(const User &I);
701  void visitFPToSI(const User &I);
702  void visitUIToFP(const User &I);
703  void visitSIToFP(const User &I);
704  void visitPtrToInt(const User &I);
705  void visitIntToPtr(const User &I);
706  void visitBitCast(const User &I);
707  void visitAddrSpaceCast(const User &I);
708
709  void visitExtractElement(const User &I);
710  void visitInsertElement(const User &I);
711  void visitShuffleVector(const User &I);
712
713  void visitExtractValue(const User &I);
714  void visitInsertValue(const User &I);
715  void visitLandingPad(const LandingPadInst &LP);
716
717  void visitGetElementPtr(const User &I);
718  void visitSelect(const User &I);
719
720  void visitAlloca(const AllocaInst &I);
721  void visitLoad(const LoadInst &I);
722  void visitStore(const StoreInst &I);
723  void visitMaskedLoad(const CallInst &I, bool IsExpanding = false);
724  void visitMaskedStore(const CallInst &I, bool IsCompressing = false);
725  void visitMaskedGather(const CallInst &I);
726  void visitMaskedScatter(const CallInst &I);
727  void visitAtomicCmpXchg(const AtomicCmpXchgInst &I);
728  void visitAtomicRMW(const AtomicRMWInst &I);
729  void visitFence(const FenceInst &I);
730  void visitPHI(const PHINode &I);
731  void visitCall(const CallInst &I);
732  bool visitMemCmpCall(const CallInst &I);
733  bool visitMemPCpyCall(const CallInst &I);
734  bool visitMemChrCall(const CallInst &I);
735  bool visitStrCpyCall(const CallInst &I, bool isStpcpy);
736  bool visitStrCmpCall(const CallInst &I);
737  bool visitStrLenCall(const CallInst &I);
738  bool visitStrNLenCall(const CallInst &I);
739  bool visitUnaryFloatCall(const CallInst &I, unsigned Opcode);
740  bool visitBinaryFloatCall(const CallInst &I, unsigned Opcode);
741  void visitAtomicLoad(const LoadInst &I);
742  void visitAtomicStore(const StoreInst &I);
743  void visitLoadFromSwiftError(const LoadInst &I);
744  void visitStoreToSwiftError(const StoreInst &I);
745
746  void visitInlineAsm(ImmutableCallSite CS);
747  void visitIntrinsicCall(const CallInst &I, unsigned Intrinsic);
748  void visitTargetIntrinsic(const CallInst &I, unsigned Intrinsic);
749  void visitConstrainedFPIntrinsic(const ConstrainedFPIntrinsic &FPI);
750
751  void visitVAStart(const CallInst &I);
752  void visitVAArg(const VAArgInst &I);
753  void visitVAEnd(const CallInst &I);
754  void visitVACopy(const CallInst &I);
755  void visitStackmap(const CallInst &I);
756  void visitPatchpoint(ImmutableCallSite CS,
757                       const BasicBlock *EHPadBB = nullptr);
758
759  // These two are implemented in StatepointLowering.cpp
760  void visitGCRelocate(const GCRelocateInst &Relocate);
761  void visitGCResult(const GCResultInst &I);
762
763  void visitVectorReduce(const CallInst &I, unsigned Intrinsic);
764
765  void visitUserOp1(const Instruction &I) {
766    llvm_unreachable("UserOp1 should not exist at instruction selection time!");
767  }
768  void visitUserOp2(const Instruction &I) {
769    llvm_unreachable("UserOp2 should not exist at instruction selection time!");
770  }
771
772  void processIntegerCallValue(const Instruction &I,
773                               SDValue Value, bool IsSigned);
774
775  void HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB);
776
777  void emitInlineAsmError(ImmutableCallSite CS, const Twine &Message);
778
779  /// If V is an function argument then create corresponding DBG_VALUE machine
780  /// instruction for it now. At the end of instruction selection, they will be
781  /// inserted to the entry BB.
782  bool EmitFuncArgumentDbgValue(const Value *V, DILocalVariable *Variable,
783                                DIExpression *Expr, DILocation *DL,
784                                bool IsDbgDeclare, const SDValue &N);
785
786  /// Return the next block after MBB, or nullptr if there is none.
787  MachineBasicBlock *NextBlock(MachineBasicBlock *MBB);
788
789  /// Update the DAG and DAG builder with the relevant information after
790  /// a new root node has been created which could be a tail call.
791  void updateDAGForMaybeTailCall(SDValue MaybeTC);
792
793  /// Return the appropriate SDDbgValue based on N.
794  SDDbgValue *getDbgValue(SDValue N, DILocalVariable *Variable,
795                          DIExpression *Expr, const DebugLoc &dl,
796                          unsigned DbgSDNodeOrder);
797
798  /// Lowers CallInst to an external symbol.
799  void lowerCallToExternalSymbol(const CallInst &I, const char *FunctionName);
800};
801
802/// This struct represents the registers (physical or virtual)
803/// that a particular set of values is assigned, and the type information about
804/// the value. The most common situation is to represent one value at a time,
805/// but struct or array values are handled element-wise as multiple values.  The
806/// splitting of aggregates is performed recursively, so that we never have
807/// aggregate-typed registers. The values at this point do not necessarily have
808/// legal types, so each value may require one or more registers of some legal
809/// type.
810///
811struct RegsForValue {
812  /// The value types of the values, which may not be legal, and
813  /// may need be promoted or synthesized from one or more registers.
814  SmallVector<EVT, 4> ValueVTs;
815
816  /// The value types of the registers. This is the same size as ValueVTs and it
817  /// records, for each value, what the type of the assigned register or
818  /// registers are. (Individual values are never synthesized from more than one
819  /// type of register.)
820  ///
821  /// With virtual registers, the contents of RegVTs is redundant with TLI's
822  /// getRegisterType member function, however when with physical registers
823  /// it is necessary to have a separate record of the types.
824  SmallVector<MVT, 4> RegVTs;
825
826  /// This list holds the registers assigned to the values.
827  /// Each legal or promoted value requires one register, and each
828  /// expanded value requires multiple registers.
829  SmallVector<unsigned, 4> Regs;
830
831  /// This list holds the number of registers for each value.
832  SmallVector<unsigned, 4> RegCount;
833
834  /// Records if this value needs to be treated in an ABI dependant manner,
835  /// different to normal type legalization.
836  Optional<CallingConv::ID> CallConv;
837
838  RegsForValue() = default;
839  RegsForValue(const SmallVector<unsigned, 4> &regs, MVT regvt, EVT valuevt,
840               Optional<CallingConv::ID> CC = None);
841  RegsForValue(LLVMContext &Context, const TargetLowering &TLI,
842               const DataLayout &DL, unsigned Reg, Type *Ty,
843               Optional<CallingConv::ID> CC);
844
845  bool isABIMangled() const {
846    return CallConv.hasValue();
847  }
848
849  /// Add the specified values to this one.
850  void append(const RegsForValue &RHS) {
851    ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
852    RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
853    Regs.append(RHS.Regs.begin(), RHS.Regs.end());
854    RegCount.push_back(RHS.Regs.size());
855  }
856
857  /// Emit a series of CopyFromReg nodes that copies from this value and returns
858  /// the result as a ValueVTs value. This uses Chain/Flag as the input and
859  /// updates them for the output Chain/Flag. If the Flag pointer is NULL, no
860  /// flag is used.
861  SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
862                          const SDLoc &dl, SDValue &Chain, SDValue *Flag,
863                          const Value *V = nullptr) const;
864
865  /// Emit a series of CopyToReg nodes that copies the specified value into the
866  /// registers specified by this object. This uses Chain/Flag as the input and
867  /// updates them for the output Chain/Flag. If the Flag pointer is nullptr, no
868  /// flag is used. If V is not nullptr, then it is used in printing better
869  /// diagnostic messages on error.
870  void getCopyToRegs(SDValue Val, SelectionDAG &DAG, const SDLoc &dl,
871                     SDValue &Chain, SDValue *Flag, const Value *V = nullptr,
872                     ISD::NodeType PreferredExtendType = ISD::ANY_EXTEND) const;
873
874  /// Add this value to the specified inlineasm node operand list. This adds the
875  /// code marker, matching input operand index (if applicable), and includes
876  /// the number of values added into it.
877  void AddInlineAsmOperands(unsigned Code, bool HasMatching,
878                            unsigned MatchingIdx, const SDLoc &dl,
879                            SelectionDAG &DAG, std::vector<SDValue> &Ops) const;
880
881  /// Check if the total RegCount is greater than one.
882  bool occupiesMultipleRegs() const {
883    return std::accumulate(RegCount.begin(), RegCount.end(), 0) > 1;
884  }
885
886  /// Return a list of registers and their sizes.
887  SmallVector<std::pair<unsigned, unsigned>, 4> getRegsAndSizes() const;
888};
889
890} // end namespace llvm
891
892#endif // LLVM_LIB_CODEGEN_SELECTIONDAG_SELECTIONDAGBUILDER_H
893