1 //===- PlaceSafepoints.cpp - Place GC Safepoints --------------------------===//
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 // Place garbage collection safepoints at appropriate locations in the IR. This
10 // does not make relocation semantics or variable liveness explicit.  That's
11 // done by RewriteStatepointsForGC.
12 //
13 // Terminology:
14 // - A call is said to be "parseable" if there is a stack map generated for the
15 // return PC of the call.  A runtime can determine where values listed in the
16 // deopt arguments and (after RewriteStatepointsForGC) gc arguments are located
17 // on the stack when the code is suspended inside such a call.  Every parse
18 // point is represented by a call wrapped in an gc.statepoint intrinsic.
19 // - A "poll" is an explicit check in the generated code to determine if the
20 // runtime needs the generated code to cooperate by calling a helper routine
21 // and thus suspending its execution at a known state. The call to the helper
22 // routine will be parseable.  The (gc & runtime specific) logic of a poll is
23 // assumed to be provided in a function of the name "gc.safepoint_poll".
24 //
25 // We aim to insert polls such that running code can quickly be brought to a
26 // well defined state for inspection by the collector.  In the current
27 // implementation, this is done via the insertion of poll sites at method entry
28 // and the backedge of most loops.  We try to avoid inserting more polls than
29 // are necessary to ensure a finite period between poll sites.  This is not
30 // because the poll itself is expensive in the generated code; it's not.  Polls
31 // do tend to impact the optimizer itself in negative ways; we'd like to avoid
32 // perturbing the optimization of the method as much as we can.
33 //
34 // We also need to make most call sites parseable.  The callee might execute a
35 // poll (or otherwise be inspected by the GC).  If so, the entire stack
36 // (including the suspended frame of the current method) must be parseable.
37 //
38 // This pass will insert:
39 // - Call parse points ("call safepoints") for any call which may need to
40 // reach a safepoint during the execution of the callee function.
41 // - Backedge safepoint polls and entry safepoint polls to ensure that
42 // executing code reaches a safepoint poll in a finite amount of time.
43 //
44 // We do not currently support return statepoints, but adding them would not
45 // be hard.  They are not required for correctness - entry safepoints are an
46 // alternative - but some GCs may prefer them.  Patches welcome.
47 //
48 //===----------------------------------------------------------------------===//
49 
50 #include "llvm/InitializePasses.h"
51 #include "llvm/Pass.h"
52 
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/Analysis/CFG.h"
56 #include "llvm/Analysis/LoopInfo.h"
57 #include "llvm/Analysis/ScalarEvolution.h"
58 #include "llvm/Analysis/TargetLibraryInfo.h"
59 #include "llvm/IR/Dominators.h"
60 #include "llvm/IR/IntrinsicInst.h"
61 #include "llvm/IR/LegacyPassManager.h"
62 #include "llvm/IR/Statepoint.h"
63 #include "llvm/Support/CommandLine.h"
64 #include "llvm/Support/Debug.h"
65 #include "llvm/Transforms/Scalar.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Cloning.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 
70 #define DEBUG_TYPE "safepoint-placement"
71 
72 STATISTIC(NumEntrySafepoints, "Number of entry safepoints inserted");
73 STATISTIC(NumBackedgeSafepoints, "Number of backedge safepoints inserted");
74 
75 STATISTIC(CallInLoop,
76           "Number of loops without safepoints due to calls in loop");
77 STATISTIC(FiniteExecution,
78           "Number of loops without safepoints finite execution");
79 
80 using namespace llvm;
81 
82 // Ignore opportunities to avoid placing safepoints on backedges, useful for
83 // validation
84 static cl::opt<bool> AllBackedges("spp-all-backedges", cl::Hidden,
85                                   cl::init(false));
86 
87 /// How narrow does the trip count of a loop have to be to have to be considered
88 /// "counted"?  Counted loops do not get safepoints at backedges.
89 static cl::opt<int> CountedLoopTripWidth("spp-counted-loop-trip-width",
90                                          cl::Hidden, cl::init(32));
91 
92 // If true, split the backedge of a loop when placing the safepoint, otherwise
93 // split the latch block itself.  Both are useful to support for
94 // experimentation, but in practice, it looks like splitting the backedge
95 // optimizes better.
96 static cl::opt<bool> SplitBackedge("spp-split-backedge", cl::Hidden,
97                                    cl::init(false));
98 
99 namespace {
100 
101 /// An analysis pass whose purpose is to identify each of the backedges in
102 /// the function which require a safepoint poll to be inserted.
103 struct PlaceBackedgeSafepointsImpl : public FunctionPass {
104   static char ID;
105 
106   /// The output of the pass - gives a list of each backedge (described by
107   /// pointing at the branch) which need a poll inserted.
108   std::vector<Instruction *> PollLocations;
109 
110   /// True unless we're running spp-no-calls in which case we need to disable
111   /// the call-dependent placement opts.
112   bool CallSafepointsEnabled;
113 
114   ScalarEvolution *SE = nullptr;
115   DominatorTree *DT = nullptr;
116   LoopInfo *LI = nullptr;
117   TargetLibraryInfo *TLI = nullptr;
118 
119   PlaceBackedgeSafepointsImpl(bool CallSafepoints = false)
120       : FunctionPass(ID), CallSafepointsEnabled(CallSafepoints) {
121     initializePlaceBackedgeSafepointsImplPass(*PassRegistry::getPassRegistry());
122   }
123 
124   bool runOnLoop(Loop *);
125   void runOnLoopAndSubLoops(Loop *L) {
126     // Visit all the subloops
127     for (Loop *I : *L)
128       runOnLoopAndSubLoops(I);
129     runOnLoop(L);
130   }
131 
132   bool runOnFunction(Function &F) override {
133     SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
134     DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
135     LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
136     TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
137     for (Loop *I : *LI) {
138       runOnLoopAndSubLoops(I);
139     }
140     return false;
141   }
142 
143   void getAnalysisUsage(AnalysisUsage &AU) const override {
144     AU.addRequired<DominatorTreeWrapperPass>();
145     AU.addRequired<ScalarEvolutionWrapperPass>();
146     AU.addRequired<LoopInfoWrapperPass>();
147     AU.addRequired<TargetLibraryInfoWrapperPass>();
148     // We no longer modify the IR at all in this pass.  Thus all
149     // analysis are preserved.
150     AU.setPreservesAll();
151   }
152 };
153 }
154 
155 static cl::opt<bool> NoEntry("spp-no-entry", cl::Hidden, cl::init(false));
156 static cl::opt<bool> NoCall("spp-no-call", cl::Hidden, cl::init(false));
157 static cl::opt<bool> NoBackedge("spp-no-backedge", cl::Hidden, cl::init(false));
158 
159 namespace {
160 struct PlaceSafepoints : public FunctionPass {
161   static char ID; // Pass identification, replacement for typeid
162 
163   PlaceSafepoints() : FunctionPass(ID) {
164     initializePlaceSafepointsPass(*PassRegistry::getPassRegistry());
165   }
166   bool runOnFunction(Function &F) override;
167 
168   void getAnalysisUsage(AnalysisUsage &AU) const override {
169     // We modify the graph wholesale (inlining, block insertion, etc).  We
170     // preserve nothing at the moment.  We could potentially preserve dom tree
171     // if that was worth doing
172     AU.addRequired<TargetLibraryInfoWrapperPass>();
173   }
174 };
175 }
176 
177 // Insert a safepoint poll immediately before the given instruction.  Does
178 // not handle the parsability of state at the runtime call, that's the
179 // callers job.
180 static void
181 InsertSafepointPoll(Instruction *InsertBefore,
182                     std::vector<CallBase *> &ParsePointsNeeded /*rval*/,
183                     const TargetLibraryInfo &TLI);
184 
185 static bool needsStatepoint(CallBase *Call, const TargetLibraryInfo &TLI) {
186   if (callsGCLeafFunction(Call, TLI))
187     return false;
188   if (auto *CI = dyn_cast<CallInst>(Call)) {
189     if (CI->isInlineAsm())
190       return false;
191   }
192 
193   return !(isa<GCStatepointInst>(Call) || isa<GCRelocateInst>(Call) ||
194            isa<GCResultInst>(Call));
195 }
196 
197 /// Returns true if this loop is known to contain a call safepoint which
198 /// must unconditionally execute on any iteration of the loop which returns
199 /// to the loop header via an edge from Pred.  Returns a conservative correct
200 /// answer; i.e. false is always valid.
201 static bool containsUnconditionalCallSafepoint(Loop *L, BasicBlock *Header,
202                                                BasicBlock *Pred,
203                                                DominatorTree &DT,
204                                                const TargetLibraryInfo &TLI) {
205   // In general, we're looking for any cut of the graph which ensures
206   // there's a call safepoint along every edge between Header and Pred.
207   // For the moment, we look only for the 'cuts' that consist of a single call
208   // instruction in a block which is dominated by the Header and dominates the
209   // loop latch (Pred) block.  Somewhat surprisingly, walking the entire chain
210   // of such dominating blocks gets substantially more occurrences than just
211   // checking the Pred and Header blocks themselves.  This may be due to the
212   // density of loop exit conditions caused by range and null checks.
213   // TODO: structure this as an analysis pass, cache the result for subloops,
214   // avoid dom tree recalculations
215   assert(DT.dominates(Header, Pred) && "loop latch not dominated by header?");
216 
217   BasicBlock *Current = Pred;
218   while (true) {
219     for (Instruction &I : *Current) {
220       if (auto *Call = dyn_cast<CallBase>(&I))
221         // Note: Technically, needing a safepoint isn't quite the right
222         // condition here.  We should instead be checking if the target method
223         // has an
224         // unconditional poll. In practice, this is only a theoretical concern
225         // since we don't have any methods with conditional-only safepoint
226         // polls.
227         if (needsStatepoint(Call, TLI))
228           return true;
229     }
230 
231     if (Current == Header)
232       break;
233     Current = DT.getNode(Current)->getIDom()->getBlock();
234   }
235 
236   return false;
237 }
238 
239 /// Returns true if this loop is known to terminate in a finite number of
240 /// iterations.  Note that this function may return false for a loop which
241 /// does actual terminate in a finite constant number of iterations due to
242 /// conservatism in the analysis.
243 static bool mustBeFiniteCountedLoop(Loop *L, ScalarEvolution *SE,
244                                     BasicBlock *Pred) {
245   // A conservative bound on the loop as a whole.
246   const SCEV *MaxTrips = SE->getConstantMaxBackedgeTakenCount(L);
247   if (!isa<SCEVCouldNotCompute>(MaxTrips) &&
248       SE->getUnsignedRange(MaxTrips).getUnsignedMax().isIntN(
249           CountedLoopTripWidth))
250     return true;
251 
252   // If this is a conditional branch to the header with the alternate path
253   // being outside the loop, we can ask questions about the execution frequency
254   // of the exit block.
255   if (L->isLoopExiting(Pred)) {
256     // This returns an exact expression only.  TODO: We really only need an
257     // upper bound here, but SE doesn't expose that.
258     const SCEV *MaxExec = SE->getExitCount(L, Pred);
259     if (!isa<SCEVCouldNotCompute>(MaxExec) &&
260         SE->getUnsignedRange(MaxExec).getUnsignedMax().isIntN(
261             CountedLoopTripWidth))
262         return true;
263   }
264 
265   return /* not finite */ false;
266 }
267 
268 static void scanOneBB(Instruction *Start, Instruction *End,
269                       std::vector<CallInst *> &Calls,
270                       DenseSet<BasicBlock *> &Seen,
271                       std::vector<BasicBlock *> &Worklist) {
272   for (BasicBlock::iterator BBI(Start), BBE0 = Start->getParent()->end(),
273                                         BBE1 = BasicBlock::iterator(End);
274        BBI != BBE0 && BBI != BBE1; BBI++) {
275     if (CallInst *CI = dyn_cast<CallInst>(&*BBI))
276       Calls.push_back(CI);
277 
278     // FIXME: This code does not handle invokes
279     assert(!isa<InvokeInst>(&*BBI) &&
280            "support for invokes in poll code needed");
281 
282     // Only add the successor blocks if we reach the terminator instruction
283     // without encountering end first
284     if (BBI->isTerminator()) {
285       BasicBlock *BB = BBI->getParent();
286       for (BasicBlock *Succ : successors(BB)) {
287         if (Seen.insert(Succ).second) {
288           Worklist.push_back(Succ);
289         }
290       }
291     }
292   }
293 }
294 
295 static void scanInlinedCode(Instruction *Start, Instruction *End,
296                             std::vector<CallInst *> &Calls,
297                             DenseSet<BasicBlock *> &Seen) {
298   Calls.clear();
299   std::vector<BasicBlock *> Worklist;
300   Seen.insert(Start->getParent());
301   scanOneBB(Start, End, Calls, Seen, Worklist);
302   while (!Worklist.empty()) {
303     BasicBlock *BB = Worklist.back();
304     Worklist.pop_back();
305     scanOneBB(&*BB->begin(), End, Calls, Seen, Worklist);
306   }
307 }
308 
309 bool PlaceBackedgeSafepointsImpl::runOnLoop(Loop *L) {
310   // Loop through all loop latches (branches controlling backedges).  We need
311   // to place a safepoint on every backedge (potentially).
312   // Note: In common usage, there will be only one edge due to LoopSimplify
313   // having run sometime earlier in the pipeline, but this code must be correct
314   // w.r.t. loops with multiple backedges.
315   BasicBlock *Header = L->getHeader();
316   SmallVector<BasicBlock*, 16> LoopLatches;
317   L->getLoopLatches(LoopLatches);
318   for (BasicBlock *Pred : LoopLatches) {
319     assert(L->contains(Pred));
320 
321     // Make a policy decision about whether this loop needs a safepoint or
322     // not.  Note that this is about unburdening the optimizer in loops, not
323     // avoiding the runtime cost of the actual safepoint.
324     if (!AllBackedges) {
325       if (mustBeFiniteCountedLoop(L, SE, Pred)) {
326         LLVM_DEBUG(dbgs() << "skipping safepoint placement in finite loop\n");
327         FiniteExecution++;
328         continue;
329       }
330       if (CallSafepointsEnabled &&
331           containsUnconditionalCallSafepoint(L, Header, Pred, *DT, *TLI)) {
332         // Note: This is only semantically legal since we won't do any further
333         // IPO or inlining before the actual call insertion..  If we hadn't, we
334         // might latter loose this call safepoint.
335         LLVM_DEBUG(
336             dbgs()
337             << "skipping safepoint placement due to unconditional call\n");
338         CallInLoop++;
339         continue;
340       }
341     }
342 
343     // TODO: We can create an inner loop which runs a finite number of
344     // iterations with an outer loop which contains a safepoint.  This would
345     // not help runtime performance that much, but it might help our ability to
346     // optimize the inner loop.
347 
348     // Safepoint insertion would involve creating a new basic block (as the
349     // target of the current backedge) which does the safepoint (of all live
350     // variables) and branches to the true header
351     Instruction *Term = Pred->getTerminator();
352 
353     LLVM_DEBUG(dbgs() << "[LSP] terminator instruction: " << *Term);
354 
355     PollLocations.push_back(Term);
356   }
357 
358   return false;
359 }
360 
361 /// Returns true if an entry safepoint is not required before this callsite in
362 /// the caller function.
363 static bool doesNotRequireEntrySafepointBefore(CallBase *Call) {
364   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Call)) {
365     switch (II->getIntrinsicID()) {
366     case Intrinsic::experimental_gc_statepoint:
367     case Intrinsic::experimental_patchpoint_void:
368     case Intrinsic::experimental_patchpoint_i64:
369       // The can wrap an actual call which may grow the stack by an unbounded
370       // amount or run forever.
371       return false;
372     default:
373       // Most LLVM intrinsics are things which do not expand to actual calls, or
374       // at least if they do, are leaf functions that cause only finite stack
375       // growth.  In particular, the optimizer likes to form things like memsets
376       // out of stores in the original IR.  Another important example is
377       // llvm.localescape which must occur in the entry block.  Inserting a
378       // safepoint before it is not legal since it could push the localescape
379       // out of the entry block.
380       return true;
381     }
382   }
383   return false;
384 }
385 
386 static Instruction *findLocationForEntrySafepoint(Function &F,
387                                                   DominatorTree &DT) {
388 
389   // Conceptually, this poll needs to be on method entry, but in
390   // practice, we place it as late in the entry block as possible.  We
391   // can place it as late as we want as long as it dominates all calls
392   // that can grow the stack.  This, combined with backedge polls,
393   // give us all the progress guarantees we need.
394 
395   // hasNextInstruction and nextInstruction are used to iterate
396   // through a "straight line" execution sequence.
397 
398   auto HasNextInstruction = [](Instruction *I) {
399     if (!I->isTerminator())
400       return true;
401 
402     BasicBlock *nextBB = I->getParent()->getUniqueSuccessor();
403     return nextBB && (nextBB->getUniquePredecessor() != nullptr);
404   };
405 
406   auto NextInstruction = [&](Instruction *I) {
407     assert(HasNextInstruction(I) &&
408            "first check if there is a next instruction!");
409 
410     if (I->isTerminator())
411       return &I->getParent()->getUniqueSuccessor()->front();
412     return &*++I->getIterator();
413   };
414 
415   Instruction *Cursor = nullptr;
416   for (Cursor = &F.getEntryBlock().front(); HasNextInstruction(Cursor);
417        Cursor = NextInstruction(Cursor)) {
418 
419     // We need to ensure a safepoint poll occurs before any 'real' call.  The
420     // easiest way to ensure finite execution between safepoints in the face of
421     // recursive and mutually recursive functions is to enforce that each take
422     // a safepoint.  Additionally, we need to ensure a poll before any call
423     // which can grow the stack by an unbounded amount.  This isn't required
424     // for GC semantics per se, but is a common requirement for languages
425     // which detect stack overflow via guard pages and then throw exceptions.
426     if (auto *Call = dyn_cast<CallBase>(Cursor)) {
427       if (doesNotRequireEntrySafepointBefore(Call))
428         continue;
429       break;
430     }
431   }
432 
433   assert((HasNextInstruction(Cursor) || Cursor->isTerminator()) &&
434          "either we stopped because of a call, or because of terminator");
435 
436   return Cursor;
437 }
438 
439 const char GCSafepointPollName[] = "gc.safepoint_poll";
440 
441 static bool isGCSafepointPoll(Function &F) {
442   return F.getName().equals(GCSafepointPollName);
443 }
444 
445 /// Returns true if this function should be rewritten to include safepoint
446 /// polls and parseable call sites.  The main point of this function is to be
447 /// an extension point for custom logic.
448 static bool shouldRewriteFunction(Function &F) {
449   // TODO: This should check the GCStrategy
450   if (F.hasGC()) {
451     const auto &FunctionGCName = F.getGC();
452     const StringRef StatepointExampleName("statepoint-example");
453     const StringRef CoreCLRName("coreclr");
454     return (StatepointExampleName == FunctionGCName) ||
455            (CoreCLRName == FunctionGCName);
456   } else
457     return false;
458 }
459 
460 // TODO: These should become properties of the GCStrategy, possibly with
461 // command line overrides.
462 static bool enableEntrySafepoints(Function &F) { return !NoEntry; }
463 static bool enableBackedgeSafepoints(Function &F) { return !NoBackedge; }
464 static bool enableCallSafepoints(Function &F) { return !NoCall; }
465 
466 bool PlaceSafepoints::runOnFunction(Function &F) {
467   if (F.isDeclaration() || F.empty()) {
468     // This is a declaration, nothing to do.  Must exit early to avoid crash in
469     // dom tree calculation
470     return false;
471   }
472 
473   if (isGCSafepointPoll(F)) {
474     // Given we're inlining this inside of safepoint poll insertion, this
475     // doesn't make any sense.  Note that we do make any contained calls
476     // parseable after we inline a poll.
477     return false;
478   }
479 
480   if (!shouldRewriteFunction(F))
481     return false;
482 
483   const TargetLibraryInfo &TLI =
484       getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
485 
486   bool Modified = false;
487 
488   // In various bits below, we rely on the fact that uses are reachable from
489   // defs.  When there are basic blocks unreachable from the entry, dominance
490   // and reachablity queries return non-sensical results.  Thus, we preprocess
491   // the function to ensure these properties hold.
492   Modified |= removeUnreachableBlocks(F);
493 
494   // STEP 1 - Insert the safepoint polling locations.  We do not need to
495   // actually insert parse points yet.  That will be done for all polls and
496   // calls in a single pass.
497 
498   DominatorTree DT;
499   DT.recalculate(F);
500 
501   SmallVector<Instruction *, 16> PollsNeeded;
502   std::vector<CallBase *> ParsePointNeeded;
503 
504   if (enableBackedgeSafepoints(F)) {
505     // Construct a pass manager to run the LoopPass backedge logic.  We
506     // need the pass manager to handle scheduling all the loop passes
507     // appropriately.  Doing this by hand is painful and just not worth messing
508     // with for the moment.
509     legacy::FunctionPassManager FPM(F.getParent());
510     bool CanAssumeCallSafepoints = enableCallSafepoints(F);
511     auto *PBS = new PlaceBackedgeSafepointsImpl(CanAssumeCallSafepoints);
512     FPM.add(PBS);
513     FPM.run(F);
514 
515     // We preserve dominance information when inserting the poll, otherwise
516     // we'd have to recalculate this on every insert
517     DT.recalculate(F);
518 
519     auto &PollLocations = PBS->PollLocations;
520 
521     auto OrderByBBName = [](Instruction *a, Instruction *b) {
522       return a->getParent()->getName() < b->getParent()->getName();
523     };
524     // We need the order of list to be stable so that naming ends up stable
525     // when we split edges.  This makes test cases much easier to write.
526     llvm::sort(PollLocations, OrderByBBName);
527 
528     // We can sometimes end up with duplicate poll locations.  This happens if
529     // a single loop is visited more than once.   The fact this happens seems
530     // wrong, but it does happen for the split-backedge.ll test case.
531     PollLocations.erase(std::unique(PollLocations.begin(),
532                                     PollLocations.end()),
533                         PollLocations.end());
534 
535     // Insert a poll at each point the analysis pass identified
536     // The poll location must be the terminator of a loop latch block.
537     for (Instruction *Term : PollLocations) {
538       // We are inserting a poll, the function is modified
539       Modified = true;
540 
541       if (SplitBackedge) {
542         // Split the backedge of the loop and insert the poll within that new
543         // basic block.  This creates a loop with two latches per original
544         // latch (which is non-ideal), but this appears to be easier to
545         // optimize in practice than inserting the poll immediately before the
546         // latch test.
547 
548         // Since this is a latch, at least one of the successors must dominate
549         // it. Its possible that we have a) duplicate edges to the same header
550         // and b) edges to distinct loop headers.  We need to insert pools on
551         // each.
552         SetVector<BasicBlock *> Headers;
553         for (unsigned i = 0; i < Term->getNumSuccessors(); i++) {
554           BasicBlock *Succ = Term->getSuccessor(i);
555           if (DT.dominates(Succ, Term->getParent())) {
556             Headers.insert(Succ);
557           }
558         }
559         assert(!Headers.empty() && "poll location is not a loop latch?");
560 
561         // The split loop structure here is so that we only need to recalculate
562         // the dominator tree once.  Alternatively, we could just keep it up to
563         // date and use a more natural merged loop.
564         SetVector<BasicBlock *> SplitBackedges;
565         for (BasicBlock *Header : Headers) {
566           BasicBlock *NewBB = SplitEdge(Term->getParent(), Header, &DT);
567           PollsNeeded.push_back(NewBB->getTerminator());
568           NumBackedgeSafepoints++;
569         }
570       } else {
571         // Split the latch block itself, right before the terminator.
572         PollsNeeded.push_back(Term);
573         NumBackedgeSafepoints++;
574       }
575     }
576   }
577 
578   if (enableEntrySafepoints(F)) {
579     if (Instruction *Location = findLocationForEntrySafepoint(F, DT)) {
580       PollsNeeded.push_back(Location);
581       Modified = true;
582       NumEntrySafepoints++;
583     }
584     // TODO: else we should assert that there was, in fact, a policy choice to
585     // not insert a entry safepoint poll.
586   }
587 
588   // Now that we've identified all the needed safepoint poll locations, insert
589   // safepoint polls themselves.
590   for (Instruction *PollLocation : PollsNeeded) {
591     std::vector<CallBase *> RuntimeCalls;
592     InsertSafepointPoll(PollLocation, RuntimeCalls, TLI);
593     llvm::append_range(ParsePointNeeded, RuntimeCalls);
594   }
595 
596   return Modified;
597 }
598 
599 char PlaceBackedgeSafepointsImpl::ID = 0;
600 char PlaceSafepoints::ID = 0;
601 
602 FunctionPass *llvm::createPlaceSafepointsPass() {
603   return new PlaceSafepoints();
604 }
605 
606 INITIALIZE_PASS_BEGIN(PlaceBackedgeSafepointsImpl,
607                       "place-backedge-safepoints-impl",
608                       "Place Backedge Safepoints", false, false)
609 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
610 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
611 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
612 INITIALIZE_PASS_END(PlaceBackedgeSafepointsImpl,
613                     "place-backedge-safepoints-impl",
614                     "Place Backedge Safepoints", false, false)
615 
616 INITIALIZE_PASS_BEGIN(PlaceSafepoints, "place-safepoints", "Place Safepoints",
617                       false, false)
618 INITIALIZE_PASS_END(PlaceSafepoints, "place-safepoints", "Place Safepoints",
619                     false, false)
620 
621 static void
622 InsertSafepointPoll(Instruction *InsertBefore,
623                     std::vector<CallBase *> &ParsePointsNeeded /*rval*/,
624                     const TargetLibraryInfo &TLI) {
625   BasicBlock *OrigBB = InsertBefore->getParent();
626   Module *M = InsertBefore->getModule();
627   assert(M && "must be part of a module");
628 
629   // Inline the safepoint poll implementation - this will get all the branch,
630   // control flow, etc..  Most importantly, it will introduce the actual slow
631   // path call - where we need to insert a safepoint (parsepoint).
632 
633   auto *F = M->getFunction(GCSafepointPollName);
634   assert(F && "gc.safepoint_poll function is missing");
635   assert(F->getValueType() ==
636          FunctionType::get(Type::getVoidTy(M->getContext()), false) &&
637          "gc.safepoint_poll declared with wrong type");
638   assert(!F->empty() && "gc.safepoint_poll must be a non-empty function");
639   CallInst *PollCall = CallInst::Create(F, "", InsertBefore);
640 
641   // Record some information about the call site we're replacing
642   BasicBlock::iterator Before(PollCall), After(PollCall);
643   bool IsBegin = false;
644   if (Before == OrigBB->begin())
645     IsBegin = true;
646   else
647     Before--;
648 
649   After++;
650   assert(After != OrigBB->end() && "must have successor");
651 
652   // Do the actual inlining
653   InlineFunctionInfo IFI;
654   bool InlineStatus = InlineFunction(*PollCall, IFI).isSuccess();
655   assert(InlineStatus && "inline must succeed");
656   (void)InlineStatus; // suppress warning in release-asserts
657 
658   // Check post-conditions
659   assert(IFI.StaticAllocas.empty() && "can't have allocs");
660 
661   std::vector<CallInst *> Calls; // new calls
662   DenseSet<BasicBlock *> BBs;    // new BBs + insertee
663 
664   // Include only the newly inserted instructions, Note: begin may not be valid
665   // if we inserted to the beginning of the basic block
666   BasicBlock::iterator Start = IsBegin ? OrigBB->begin() : std::next(Before);
667 
668   // If your poll function includes an unreachable at the end, that's not
669   // valid.  Bugpoint likes to create this, so check for it.
670   assert(isPotentiallyReachable(&*Start, &*After) &&
671          "malformed poll function");
672 
673   scanInlinedCode(&*Start, &*After, Calls, BBs);
674   assert(!Calls.empty() && "slow path not found for safepoint poll");
675 
676   // Record the fact we need a parsable state at the runtime call contained in
677   // the poll function.  This is required so that the runtime knows how to
678   // parse the last frame when we actually take  the safepoint (i.e. execute
679   // the slow path)
680   assert(ParsePointsNeeded.empty());
681   for (auto *CI : Calls) {
682     // No safepoint needed or wanted
683     if (!needsStatepoint(CI, TLI))
684       continue;
685 
686     // These are likely runtime calls.  Should we assert that via calling
687     // convention or something?
688     ParsePointsNeeded.push_back(CI);
689   }
690   assert(ParsePointsNeeded.size() <= Calls.size());
691 }
692