1//===- llvm/IR/Statepoint.h - gc.statepoint utilities -----------*- C++ -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file contains utility functions and a wrapper class analogous to
11// CallSite for accessing the fields of gc.statepoint, gc.relocate,
12// gc.result intrinsics; and some general utilities helpful when dealing with
13// gc.statepoint.
14//
15//===----------------------------------------------------------------------===//
16
17#ifndef LLVM_IR_STATEPOINT_H
18#define LLVM_IR_STATEPOINT_H
19
20#include "llvm/ADT/Optional.h"
21#include "llvm/ADT/iterator_range.h"
22#include "llvm/IR/Attributes.h"
23#include "llvm/IR/BasicBlock.h"
24#include "llvm/IR/CallSite.h"
25#include "llvm/IR/Constants.h"
26#include "llvm/IR/Function.h"
27#include "llvm/IR/Instruction.h"
28#include "llvm/IR/Instructions.h"
29#include "llvm/IR/IntrinsicInst.h"
30#include "llvm/IR/Intrinsics.h"
31#include "llvm/Support/Casting.h"
32#include "llvm/Support/MathExtras.h"
33#include <cassert>
34#include <cstddef>
35#include <cstdint>
36#include <vector>
37
38namespace llvm {
39
40/// The statepoint intrinsic accepts a set of flags as its third argument.
41/// Valid values come out of this set.
42enum class StatepointFlags {
43  None = 0,
44  GCTransition = 1, ///< Indicates that this statepoint is a transition from
45                    ///< GC-aware code to code that is not GC-aware.
46  /// Mark the deopt arguments associated with the statepoint as only being
47  /// "live-in". By default, deopt arguments are "live-through".  "live-through"
48  /// requires that they the value be live on entry, on exit, and at any point
49  /// during the call.  "live-in" only requires the value be available at the
50  /// start of the call.  In particular, "live-in" values can be placed in
51  /// unused argument registers or other non-callee saved registers.
52  DeoptLiveIn = 2,
53
54  MaskAll = 3 ///< A bitmask that includes all valid flags.
55};
56
57class GCRelocateInst;
58class GCResultInst;
59
60bool isStatepoint(ImmutableCallSite CS);
61bool isStatepoint(const Value *V);
62bool isStatepoint(const Value &V);
63
64bool isGCRelocate(ImmutableCallSite CS);
65bool isGCResult(ImmutableCallSite CS);
66
67/// Analogous to CallSiteBase, this provides most of the actual
68/// functionality for Statepoint and ImmutableStatepoint.  It is
69/// templatized to allow easily specializing of const and non-const
70/// concrete subtypes.  This is structured analogous to CallSite
71/// rather than the IntrinsicInst.h helpers since we need to support
72/// invokable statepoints.
73template <typename FunTy, typename InstructionTy, typename ValueTy,
74          typename CallSiteTy>
75class StatepointBase {
76  CallSiteTy StatepointCS;
77
78protected:
79  explicit StatepointBase(InstructionTy *I) {
80    if (isStatepoint(I)) {
81      StatepointCS = CallSiteTy(I);
82      assert(StatepointCS && "isStatepoint implies CallSite");
83    }
84  }
85
86  explicit StatepointBase(CallSiteTy CS) {
87    if (isStatepoint(CS))
88      StatepointCS = CS;
89  }
90
91public:
92  using arg_iterator = typename CallSiteTy::arg_iterator;
93
94  enum {
95    IDPos = 0,
96    NumPatchBytesPos = 1,
97    CalledFunctionPos = 2,
98    NumCallArgsPos = 3,
99    FlagsPos = 4,
100    CallArgsBeginPos = 5,
101  };
102
103  void *operator new(size_t, unsigned) = delete;
104  void *operator new(size_t s) = delete;
105
106  explicit operator bool() const {
107    // We do not assign non-statepoint CallSites to StatepointCS.
108    return (bool)StatepointCS;
109  }
110
111  /// Return the underlying CallSite.
112  CallSiteTy getCallSite() const {
113    assert(*this && "check validity first!");
114    return StatepointCS;
115  }
116
117  uint64_t getFlags() const {
118    return cast<ConstantInt>(getCallSite().getArgument(FlagsPos))
119        ->getZExtValue();
120  }
121
122  /// Return the ID associated with this statepoint.
123  uint64_t getID() const {
124    const Value *IDVal = getCallSite().getArgument(IDPos);
125    return cast<ConstantInt>(IDVal)->getZExtValue();
126  }
127
128  /// Return the number of patchable bytes associated with this statepoint.
129  uint32_t getNumPatchBytes() const {
130    const Value *NumPatchBytesVal = getCallSite().getArgument(NumPatchBytesPos);
131    uint64_t NumPatchBytes =
132      cast<ConstantInt>(NumPatchBytesVal)->getZExtValue();
133    assert(isInt<32>(NumPatchBytes) && "should fit in 32 bits!");
134    return NumPatchBytes;
135  }
136
137  /// Return the value actually being called or invoked.
138  ValueTy *getCalledValue() const {
139    return getCallSite().getArgument(CalledFunctionPos);
140  }
141
142  InstructionTy *getInstruction() const {
143    return getCallSite().getInstruction();
144  }
145
146  /// Return the function being called if this is a direct call, otherwise
147  /// return null (if it's an indirect call).
148  FunTy *getCalledFunction() const {
149    return dyn_cast<Function>(getCalledValue());
150  }
151
152  /// Return the caller function for this statepoint.
153  FunTy *getCaller() const { return getCallSite().getCaller(); }
154
155  /// Determine if the statepoint cannot unwind.
156  bool doesNotThrow() const {
157    Function *F = getCalledFunction();
158    return getCallSite().doesNotThrow() || (F ? F->doesNotThrow() : false);
159  }
160
161  /// Return the type of the value returned by the call underlying the
162  /// statepoint.
163  Type *getActualReturnType() const {
164    auto *FTy = cast<FunctionType>(
165        cast<PointerType>(getCalledValue()->getType())->getElementType());
166    return FTy->getReturnType();
167  }
168
169  /// Number of arguments to be passed to the actual callee.
170  int getNumCallArgs() const {
171    const Value *NumCallArgsVal = getCallSite().getArgument(NumCallArgsPos);
172    return cast<ConstantInt>(NumCallArgsVal)->getZExtValue();
173  }
174
175  size_t arg_size() const { return getNumCallArgs(); }
176  typename CallSiteTy::arg_iterator arg_begin() const {
177    assert(CallArgsBeginPos <= (int)getCallSite().arg_size());
178    return getCallSite().arg_begin() + CallArgsBeginPos;
179  }
180  typename CallSiteTy::arg_iterator arg_end() const {
181    auto I = arg_begin() + arg_size();
182    assert((getCallSite().arg_end() - I) >= 0);
183    return I;
184  }
185
186  ValueTy *getArgument(unsigned Index) {
187    assert(Index < arg_size() && "out of bounds!");
188    return *(arg_begin() + Index);
189  }
190
191  /// range adapter for call arguments
192  iterator_range<arg_iterator> call_args() const {
193    return make_range(arg_begin(), arg_end());
194  }
195
196  /// \brief Return true if the call or the callee has the given attribute.
197  bool paramHasAttr(unsigned i, Attribute::AttrKind A) const {
198    Function *F = getCalledFunction();
199    return getCallSite().paramHasAttr(i + CallArgsBeginPos, A) ||
200          (F ? F->getAttributes().hasAttribute(i, A) : false);
201  }
202
203  /// Number of GC transition args.
204  int getNumTotalGCTransitionArgs() const {
205    const Value *NumGCTransitionArgs = *arg_end();
206    return cast<ConstantInt>(NumGCTransitionArgs)->getZExtValue();
207  }
208  typename CallSiteTy::arg_iterator gc_transition_args_begin() const {
209    auto I = arg_end() + 1;
210    assert((getCallSite().arg_end() - I) >= 0);
211    return I;
212  }
213  typename CallSiteTy::arg_iterator gc_transition_args_end() const {
214    auto I = gc_transition_args_begin() + getNumTotalGCTransitionArgs();
215    assert((getCallSite().arg_end() - I) >= 0);
216    return I;
217  }
218
219  /// range adapter for GC transition arguments
220  iterator_range<arg_iterator> gc_transition_args() const {
221    return make_range(gc_transition_args_begin(), gc_transition_args_end());
222  }
223
224  /// Number of additional arguments excluding those intended
225  /// for garbage collection.
226  int getNumTotalVMSArgs() const {
227    const Value *NumVMSArgs = *gc_transition_args_end();
228    return cast<ConstantInt>(NumVMSArgs)->getZExtValue();
229  }
230
231  typename CallSiteTy::arg_iterator deopt_begin() const {
232    auto I = gc_transition_args_end() + 1;
233    assert((getCallSite().arg_end() - I) >= 0);
234    return I;
235  }
236  typename CallSiteTy::arg_iterator deopt_end() const {
237    auto I = deopt_begin() + getNumTotalVMSArgs();
238    assert((getCallSite().arg_end() - I) >= 0);
239    return I;
240  }
241
242  /// range adapter for vm state arguments
243  iterator_range<arg_iterator> deopt_operands() const {
244    return make_range(deopt_begin(), deopt_end());
245  }
246
247  typename CallSiteTy::arg_iterator gc_args_begin() const {
248    return deopt_end();
249  }
250  typename CallSiteTy::arg_iterator gc_args_end() const {
251    return getCallSite().arg_end();
252  }
253
254  unsigned gcArgsStartIdx() const {
255    return gc_args_begin() - getInstruction()->op_begin();
256  }
257
258  /// range adapter for gc arguments
259  iterator_range<arg_iterator> gc_args() const {
260    return make_range(gc_args_begin(), gc_args_end());
261  }
262
263  /// Get list of all gc reloactes linked to this statepoint
264  /// May contain several relocations for the same base/derived pair.
265  /// For example this could happen due to relocations on unwinding
266  /// path of invoke.
267  std::vector<const GCRelocateInst *> getRelocates() const;
268
269  /// Get the experimental_gc_result call tied to this statepoint.  Can be
270  /// nullptr if there isn't a gc_result tied to this statepoint.  Guaranteed to
271  /// be a CallInst if non-null.
272  const GCResultInst *getGCResult() const {
273    for (auto *U : getInstruction()->users())
274      if (auto *GRI = dyn_cast<GCResultInst>(U))
275        return GRI;
276    return nullptr;
277  }
278
279#ifndef NDEBUG
280  /// Asserts if this statepoint is malformed.  Common cases for failure
281  /// include incorrect length prefixes for variable length sections or
282  /// illegal values for parameters.
283  void verify() {
284    assert(getNumCallArgs() >= 0 &&
285           "number of arguments to actually callee can't be negative");
286
287    // The internal asserts in the iterator accessors do the rest.
288    (void)arg_begin();
289    (void)arg_end();
290    (void)gc_transition_args_begin();
291    (void)gc_transition_args_end();
292    (void)deopt_begin();
293    (void)deopt_end();
294    (void)gc_args_begin();
295    (void)gc_args_end();
296  }
297#endif
298};
299
300/// A specialization of it's base class for read only access
301/// to a gc.statepoint.
302class ImmutableStatepoint
303    : public StatepointBase<const Function, const Instruction, const Value,
304                            ImmutableCallSite> {
305  using Base =
306      StatepointBase<const Function, const Instruction, const Value,
307                     ImmutableCallSite>;
308
309public:
310  explicit ImmutableStatepoint(const Instruction *I) : Base(I) {}
311  explicit ImmutableStatepoint(ImmutableCallSite CS) : Base(CS) {}
312};
313
314/// A specialization of it's base class for read-write access
315/// to a gc.statepoint.
316class Statepoint
317    : public StatepointBase<Function, Instruction, Value, CallSite> {
318  using Base = StatepointBase<Function, Instruction, Value, CallSite>;
319
320public:
321  explicit Statepoint(Instruction *I) : Base(I) {}
322  explicit Statepoint(CallSite CS) : Base(CS) {}
323};
324
325/// Common base class for representing values projected from a statepoint.
326/// Currently, the only projections available are gc.result and gc.relocate.
327class GCProjectionInst : public IntrinsicInst {
328public:
329  static inline bool classof(const IntrinsicInst *I) {
330    return I->getIntrinsicID() == Intrinsic::experimental_gc_relocate ||
331      I->getIntrinsicID() == Intrinsic::experimental_gc_result;
332  }
333
334  static inline bool classof(const Value *V) {
335    return isa<IntrinsicInst>(V) && classof(cast<IntrinsicInst>(V));
336  }
337
338  /// Return true if this relocate is tied to the invoke statepoint.
339  /// This includes relocates which are on the unwinding path.
340  bool isTiedToInvoke() const {
341    const Value *Token = getArgOperand(0);
342
343    return isa<LandingPadInst>(Token) || isa<InvokeInst>(Token);
344  }
345
346  /// The statepoint with which this gc.relocate is associated.
347  const Instruction *getStatepoint() const {
348    const Value *Token = getArgOperand(0);
349
350    // This takes care both of relocates for call statepoints and relocates
351    // on normal path of invoke statepoint.
352    if (!isa<LandingPadInst>(Token)) {
353      assert(isStatepoint(Token));
354      return cast<Instruction>(Token);
355    }
356
357    // This relocate is on exceptional path of an invoke statepoint
358    const BasicBlock *InvokeBB =
359        cast<Instruction>(Token)->getParent()->getUniquePredecessor();
360
361    assert(InvokeBB && "safepoints should have unique landingpads");
362    assert(InvokeBB->getTerminator() &&
363           "safepoint block should be well formed");
364    assert(isStatepoint(InvokeBB->getTerminator()));
365
366    return InvokeBB->getTerminator();
367  }
368};
369
370/// Represents calls to the gc.relocate intrinsic.
371class GCRelocateInst : public GCProjectionInst {
372public:
373  static inline bool classof(const IntrinsicInst *I) {
374    return I->getIntrinsicID() == Intrinsic::experimental_gc_relocate;
375  }
376
377  static inline bool classof(const Value *V) {
378    return isa<IntrinsicInst>(V) && classof(cast<IntrinsicInst>(V));
379  }
380
381  /// The index into the associate statepoint's argument list
382  /// which contains the base pointer of the pointer whose
383  /// relocation this gc.relocate describes.
384  unsigned getBasePtrIndex() const {
385    return cast<ConstantInt>(getArgOperand(1))->getZExtValue();
386  }
387
388  /// The index into the associate statepoint's argument list which
389  /// contains the pointer whose relocation this gc.relocate describes.
390  unsigned getDerivedPtrIndex() const {
391    return cast<ConstantInt>(getArgOperand(2))->getZExtValue();
392  }
393
394  Value *getBasePtr() const {
395    ImmutableCallSite CS(getStatepoint());
396    return *(CS.arg_begin() + getBasePtrIndex());
397  }
398
399  Value *getDerivedPtr() const {
400    ImmutableCallSite CS(getStatepoint());
401    return *(CS.arg_begin() + getDerivedPtrIndex());
402  }
403};
404
405/// Represents calls to the gc.result intrinsic.
406class GCResultInst : public GCProjectionInst {
407public:
408  static inline bool classof(const IntrinsicInst *I) {
409    return I->getIntrinsicID() == Intrinsic::experimental_gc_result;
410  }
411
412  static inline bool classof(const Value *V) {
413    return isa<IntrinsicInst>(V) && classof(cast<IntrinsicInst>(V));
414  }
415};
416
417template <typename FunTy, typename InstructionTy, typename ValueTy,
418          typename CallSiteTy>
419std::vector<const GCRelocateInst *>
420StatepointBase<FunTy, InstructionTy, ValueTy, CallSiteTy>::getRelocates()
421    const {
422
423  std::vector<const GCRelocateInst *> Result;
424
425  CallSiteTy StatepointCS = getCallSite();
426
427  // Search for relocated pointers.  Note that working backwards from the
428  // gc_relocates ensures that we only get pairs which are actually relocated
429  // and used after the statepoint.
430  for (const User *U : getInstruction()->users())
431    if (auto *Relocate = dyn_cast<GCRelocateInst>(U))
432      Result.push_back(Relocate);
433
434  if (!StatepointCS.isInvoke())
435    return Result;
436
437  // We need to scan thorough exceptional relocations if it is invoke statepoint
438  LandingPadInst *LandingPad =
439      cast<InvokeInst>(getInstruction())->getLandingPadInst();
440
441  // Search for gc relocates that are attached to this landingpad.
442  for (const User *LandingPadUser : LandingPad->users()) {
443    if (auto *Relocate = dyn_cast<GCRelocateInst>(LandingPadUser))
444      Result.push_back(Relocate);
445  }
446  return Result;
447}
448
449/// Call sites that get wrapped by a gc.statepoint (currently only in
450/// RewriteStatepointsForGC and potentially in other passes in the future) can
451/// have attributes that describe properties of gc.statepoint call they will be
452/// eventually be wrapped in.  This struct is used represent such directives.
453struct StatepointDirectives {
454  Optional<uint32_t> NumPatchBytes;
455  Optional<uint64_t> StatepointID;
456
457  static const uint64_t DefaultStatepointID = 0xABCDEF00;
458  static const uint64_t DeoptBundleStatepointID = 0xABCDEF0F;
459};
460
461/// Parse out statepoint directives from the function attributes present in \p
462/// AS.
463StatepointDirectives parseStatepointDirectivesFromAttrs(AttributeList AS);
464
465/// Return \c true if the the \p Attr is an attribute that is a statepoint
466/// directive.
467bool isStatepointDirectiveAttr(Attribute Attr);
468
469} // end namespace llvm
470
471#endif // LLVM_IR_STATEPOINT_H
472