1 //===- PlaceSafepoints.cpp - Place GC Safepoints --------------------------===//
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 // Place garbage collection safepoints at appropriate locations in the IR. This
11 // does not make relocation semantics or variable liveness explicit. That's
12 // done by RewriteStatepointsForGC.
15 // - A call is said to be "parseable" if there is a stack map generated for the
16 // return PC of the call. A runtime can determine where values listed in the
17 // deopt arguments and (after RewriteStatepointsForGC) gc arguments are located
18 // on the stack when the code is suspended inside such a call. Every parse
19 // point is represented by a call wrapped in an gc.statepoint intrinsic.
20 // - A "poll" is an explicit check in the generated code to determine if the
21 // runtime needs the generated code to cooperate by calling a helper routine
22 // and thus suspending its execution at a known state. The call to the helper
23 // routine will be parseable. The (gc & runtime specific) logic of a poll is
24 // assumed to be provided in a function of the name "gc.safepoint_poll".
26 // We aim to insert polls such that running code can quickly be brought to a
27 // well defined state for inspection by the collector. In the current
28 // implementation, this is done via the insertion of poll sites at method entry
29 // and the backedge of most loops. We try to avoid inserting more polls than
30 // are neccessary to ensure a finite period between poll sites. This is not
31 // because the poll itself is expensive in the generated code; it's not. Polls
32 // do tend to impact the optimizer itself in negative ways; we'd like to avoid
33 // perturbing the optimization of the method as much as we can.
35 // We also need to make most call sites parseable. The callee might execute a
36 // poll (or otherwise be inspected by the GC). If so, the entire stack
37 // (including the suspended frame of the current method) must be parseable.
39 // This pass will insert:
40 // - Call parse points ("call safepoints") for any call which may need to
41 // reach a safepoint during the execution of the callee function.
42 // - Backedge safepoint polls and entry safepoint polls to ensure that
43 // executing code reaches a safepoint poll in a finite amount of time.
45 // We do not currently support return statepoints, but adding them would not
46 // be hard. They are not required for correctness - entry safepoints are an
47 // alternative - but some GCs may prefer them. Patches welcome.
49 //===----------------------------------------------------------------------===//
51 #include "llvm/Pass.h"
52 #include "llvm/PassManager.h"
53 #include "llvm/ADT/SetOperations.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/Analysis/LoopPass.h"
56 #include "llvm/Analysis/LoopInfo.h"
57 #include "llvm/Analysis/ScalarEvolution.h"
58 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
59 #include "llvm/Analysis/CFG.h"
60 #include "llvm/Analysis/InstructionSimplify.h"
61 #include "llvm/IR/BasicBlock.h"
62 #include "llvm/IR/CallSite.h"
63 #include "llvm/IR/Dominators.h"
64 #include "llvm/IR/Function.h"
65 #include "llvm/IR/IRBuilder.h"
66 #include "llvm/IR/InstIterator.h"
67 #include "llvm/IR/Instructions.h"
68 #include "llvm/IR/Intrinsics.h"
69 #include "llvm/IR/IntrinsicInst.h"
70 #include "llvm/IR/Module.h"
71 #include "llvm/IR/Statepoint.h"
72 #include "llvm/IR/Value.h"
73 #include "llvm/IR/Verifier.h"
74 #include "llvm/Support/Debug.h"
75 #include "llvm/Support/CommandLine.h"
76 #include "llvm/Support/raw_ostream.h"
77 #include "llvm/Transforms/Scalar.h"
78 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
79 #include "llvm/Transforms/Utils/Cloning.h"
80 #include "llvm/Transforms/Utils/Local.h"
82 #define DEBUG_TYPE "safepoint-placement"
83 STATISTIC(NumEntrySafepoints, "Number of entry safepoints inserted");
84 STATISTIC(NumCallSafepoints, "Number of call safepoints inserted");
85 STATISTIC(NumBackedgeSafepoints, "Number of backedge safepoints inserted");
87 STATISTIC(CallInLoop, "Number of loops w/o safepoints due to calls in loop");
88 STATISTIC(FiniteExecution, "Number of loops w/o safepoints finite execution");
92 // Ignore oppurtunities to avoid placing safepoints on backedges, useful for
94 static cl::opt<bool> AllBackedges("spp-all-backedges", cl::init(false));
96 /// If true, do not place backedge safepoints in counted loops.
97 static cl::opt<bool> SkipCounted("spp-counted", cl::init(true));
99 // If true, split the backedge of a loop when placing the safepoint, otherwise
100 // split the latch block itself. Both are useful to support for
101 // experimentation, but in practice, it looks like splitting the backedge
103 static cl::opt<bool> SplitBackedge("spp-split-backedge", cl::init(false));
105 // Print tracing output
106 static cl::opt<bool> TraceLSP("spp-trace", cl::init(false));
110 /** An analysis pass whose purpose is to identify each of the backedges in
111 the function which require a safepoint poll to be inserted. */
112 struct PlaceBackedgeSafepointsImpl : public LoopPass {
115 /// The output of the pass - gives a list of each backedge (described by
116 /// pointing at the branch) which need a poll inserted.
117 std::vector<TerminatorInst *> PollLocations;
119 /// True unless we're running spp-no-calls in which case we need to disable
120 /// the call dependend placement opts.
121 bool CallSafepointsEnabled;
122 PlaceBackedgeSafepointsImpl(bool CallSafepoints = false)
123 : LoopPass(ID), CallSafepointsEnabled(CallSafepoints) {
124 initializePlaceBackedgeSafepointsImplPass(*PassRegistry::getPassRegistry());
127 bool runOnLoop(Loop *, LPPassManager &LPM) override;
129 void getAnalysisUsage(AnalysisUsage &AU) const override {
130 // needed for determining if the loop is finite
131 AU.addRequired<ScalarEvolution>();
132 // to ensure each edge has a single backedge
133 // TODO: is this still required?
134 AU.addRequiredID(LoopSimplifyID);
136 // We no longer modify the IR at all in this pass. Thus all
137 // analysis are preserved.
138 AU.setPreservesAll();
143 static cl::opt<bool> NoEntry("spp-no-entry", cl::init(false));
144 static cl::opt<bool> NoCall("spp-no-call", cl::init(false));
145 static cl::opt<bool> NoBackedge("spp-no-backedge", cl::init(false));
148 struct PlaceSafepoints : public ModulePass {
149 static char ID; // Pass identification, replacement for typeid
151 bool EnableEntrySafepoints;
152 bool EnableBackedgeSafepoints;
153 bool EnableCallSafepoints;
155 PlaceSafepoints() : ModulePass(ID) {
156 initializePlaceSafepointsPass(*PassRegistry::getPassRegistry());
157 EnableEntrySafepoints = !NoEntry;
158 EnableBackedgeSafepoints = !NoBackedge;
159 EnableCallSafepoints = !NoCall;
161 bool runOnModule(Module &M) override {
162 bool modified = false;
163 for (Function &F : M) {
164 modified |= runOnFunction(F);
168 bool runOnFunction(Function &F);
170 void getAnalysisUsage(AnalysisUsage &AU) const override {
171 // We modify the graph wholesale (inlining, block insertion, etc). We
172 // preserve nothing at the moment. We could potentially preserve dom tree
173 // if that was worth doing
178 // Insert a safepoint poll immediately before the given instruction. Does
179 // not handle the parsability of state at the runtime call, that's the
182 InsertSafepointPoll(DominatorTree &DT, Instruction *after,
183 std::vector<CallSite> &ParsePointsNeeded /*rval*/);
185 static bool isGCLeafFunction(const CallSite &CS);
187 static bool needsStatepoint(const CallSite &CS) {
188 if (isGCLeafFunction(CS))
191 CallInst *call = cast<CallInst>(CS.getInstruction());
192 if (call->isInlineAsm())
195 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS)) {
201 static Value *ReplaceWithStatepoint(const CallSite &CS, Pass *P);
203 /// Returns true if this loop is known to contain a call safepoint which
204 /// must unconditionally execute on any iteration of the loop which returns
205 /// to the loop header via an edge from Pred. Returns a conservative correct
206 /// answer; i.e. false is always valid.
207 static bool containsUnconditionalCallSafepoint(Loop *L, BasicBlock *Header,
210 // In general, we're looking for any cut of the graph which ensures
211 // there's a call safepoint along every edge between Header and Pred.
212 // For the moment, we look only for the 'cuts' that consist of a single call
213 // instruction in a block which is dominated by the Header and dominates the
214 // loop latch (Pred) block. Somewhat surprisingly, walking the entire chain
215 // of such dominating blocks gets substaintially more occurences than just
216 // checking the Pred and Header blocks themselves. This may be due to the
217 // density of loop exit conditions caused by range and null checks.
218 // TODO: structure this as an analysis pass, cache the result for subloops,
219 // avoid dom tree recalculations
220 assert(DT.dominates(Header, Pred) && "loop latch not dominated by header?");
222 BasicBlock *Current = Pred;
224 for (Instruction &I : *Current) {
225 if (CallSite CS = &I)
226 // Note: Technically, needing a safepoint isn't quite the right
227 // condition here. We should instead be checking if the target method
229 // unconditional poll. In practice, this is only a theoretical concern
230 // since we don't have any methods with conditional-only safepoint
232 if (needsStatepoint(CS))
236 if (Current == Header)
238 Current = DT.getNode(Current)->getIDom()->getBlock();
244 /// Returns true if this loop is known to terminate in a finite number of
245 /// iterations. Note that this function may return false for a loop which
246 /// does actual terminate in a finite constant number of iterations due to
247 /// conservatism in the analysis.
248 static bool mustBeFiniteCountedLoop(Loop *L, ScalarEvolution *SE,
250 // Only used when SkipCounted is off
251 const unsigned upperTripBound = 8192;
253 // A conservative bound on the loop as a whole.
254 const SCEV *MaxTrips = SE->getMaxBackedgeTakenCount(L);
255 if (MaxTrips != SE->getCouldNotCompute()) {
256 if (SE->getUnsignedRange(MaxTrips).getUnsignedMax().ult(upperTripBound))
259 SE->getUnsignedRange(MaxTrips).getUnsignedMax().isIntN(32))
263 // If this is a conditional branch to the header with the alternate path
264 // being outside the loop, we can ask questions about the execution frequency
265 // of the exit block.
266 if (L->isLoopExiting(Pred)) {
267 // This returns an exact expression only. TODO: We really only need an
268 // upper bound here, but SE doesn't expose that.
269 const SCEV *MaxExec = SE->getExitCount(L, Pred);
270 if (MaxExec != SE->getCouldNotCompute()) {
271 if (SE->getUnsignedRange(MaxExec).getUnsignedMax().ult(upperTripBound))
274 SE->getUnsignedRange(MaxExec).getUnsignedMax().isIntN(32))
279 return /* not finite */ false;
282 static void scanOneBB(Instruction *start, Instruction *end,
283 std::vector<CallInst *> &calls,
284 std::set<BasicBlock *> &seen,
285 std::vector<BasicBlock *> &worklist) {
286 for (BasicBlock::iterator itr(start);
287 itr != start->getParent()->end() && itr != BasicBlock::iterator(end);
289 if (CallInst *CI = dyn_cast<CallInst>(&*itr)) {
292 // FIXME: This code does not handle invokes
293 assert(!dyn_cast<InvokeInst>(&*itr) &&
294 "support for invokes in poll code needed");
295 // Only add the successor blocks if we reach the terminator instruction
296 // without encountering end first
297 if (itr->isTerminator()) {
298 BasicBlock *BB = itr->getParent();
299 for (BasicBlock *Succ : successors(BB)) {
300 if (seen.count(Succ) == 0) {
301 worklist.push_back(Succ);
308 static void scanInlinedCode(Instruction *start, Instruction *end,
309 std::vector<CallInst *> &calls,
310 std::set<BasicBlock *> &seen) {
312 std::vector<BasicBlock *> worklist;
313 seen.insert(start->getParent());
314 scanOneBB(start, end, calls, seen, worklist);
315 while (!worklist.empty()) {
316 BasicBlock *BB = worklist.back();
318 scanOneBB(&*BB->begin(), end, calls, seen, worklist);
322 bool PlaceBackedgeSafepointsImpl::runOnLoop(Loop *L, LPPassManager &LPM) {
323 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
325 // Loop through all predecessors of the loop header and identify all
326 // backedges. We need to place a safepoint on every backedge (potentially).
327 // Note: Due to LoopSimplify there should only be one. Assert? Or can we
329 BasicBlock *header = L->getHeader();
331 // TODO: Use the analysis pass infrastructure for this. There is no reason
332 // to recalculate this here.
334 DT.recalculate(*header->getParent());
336 bool modified = false;
337 for (BasicBlock *pred : predecessors(header)) {
338 if (!L->contains(pred)) {
339 // This is not a backedge, it's coming from outside the loop
343 // Make a policy decision about whether this loop needs a safepoint or
344 // not. Note that this is about unburdening the optimizer in loops, not
345 // avoiding the runtime cost of the actual safepoint.
347 if (mustBeFiniteCountedLoop(L, SE, pred)) {
349 errs() << "skipping safepoint placement in finite loop\n";
353 if (CallSafepointsEnabled &&
354 containsUnconditionalCallSafepoint(L, header, pred, DT)) {
355 // Note: This is only semantically legal since we won't do any further
356 // IPO or inlining before the actual call insertion.. If we hadn't, we
357 // might latter loose this call safepoint.
359 errs() << "skipping safepoint placement due to unconditional call\n";
365 // TODO: We can create an inner loop which runs a finite number of
366 // iterations with an outer loop which contains a safepoint. This would
367 // not help runtime performance that much, but it might help our ability to
368 // optimize the inner loop.
370 // We're unconditionally going to modify this loop.
373 // Safepoint insertion would involve creating a new basic block (as the
374 // target of the current backedge) which does the safepoint (of all live
375 // variables) and branches to the true header
376 TerminatorInst *term = pred->getTerminator();
379 errs() << "[LSP] terminator instruction: ";
383 PollLocations.push_back(term);
389 static Instruction *findLocationForEntrySafepoint(Function &F,
392 // Conceptually, this poll needs to be on method entry, but in
393 // practice, we place it as late in the entry block as possible. We
394 // can place it as late as we want as long as it dominates all calls
395 // that can grow the stack. This, combined with backedge polls,
396 // give us all the progress guarantees we need.
398 // Due to the way the frontend generates IR, we may have a couple of initial
399 // basic blocks before the first bytecode. These will be single-entry
400 // single-exit blocks which conceptually are just part of the first 'real
401 // basic block'. Since we don't have deopt state until the first bytecode,
402 // walk forward until we've found the first unconditional branch or merge.
404 // hasNextInstruction and nextInstruction are used to iterate
405 // through a "straight line" execution sequence.
407 auto hasNextInstruction = [](Instruction *I) {
408 if (!I->isTerminator()) {
411 BasicBlock *nextBB = I->getParent()->getUniqueSuccessor();
412 return nextBB && (nextBB->getUniquePredecessor() != nullptr);
415 auto nextInstruction = [&hasNextInstruction](Instruction *I) {
416 assert(hasNextInstruction(I) &&
417 "first check if there is a next instruction!");
418 if (I->isTerminator()) {
419 return I->getParent()->getUniqueSuccessor()->begin();
421 return std::next(BasicBlock::iterator(I));
425 Instruction *cursor = nullptr;
426 for (cursor = F.getEntryBlock().begin(); hasNextInstruction(cursor);
427 cursor = nextInstruction(cursor)) {
429 // We need to stop going forward as soon as we see a call that can
430 // grow the stack (i.e. the call target has a non-zero frame
432 if (CallSite CS = cursor) {
433 (void)CS; // Silence an unused variable warning by gcc 4.8.2
434 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(cursor)) {
435 // llvm.assume(...) are not really calls.
436 if (II->getIntrinsicID() == Intrinsic::assume) {
444 assert((hasNextInstruction(cursor) || cursor->isTerminator()) &&
445 "either we stopped because of a call, or because of terminator");
447 if (cursor->isTerminator()) {
451 BasicBlock *BB = cursor->getParent();
452 SplitBlock(BB, cursor, nullptr);
454 // Note: SplitBlock modifies the DT. Simply passing a Pass (which is a
455 // module pass) is not enough.
458 // SplitBlock updates the DT
462 return BB->getTerminator();
465 /// Identify the list of call sites which need to be have parseable state
466 static void findCallSafepoints(Function &F,
467 std::vector<CallSite> &Found /*rval*/) {
468 assert(Found.empty() && "must be empty!");
469 for (Instruction &I : inst_range(F)) {
470 Instruction *inst = &I;
471 if (isa<CallInst>(inst) || isa<InvokeInst>(inst)) {
474 // No safepoint needed or wanted
475 if (!needsStatepoint(CS)) {
484 /// Implement a unique function which doesn't require we sort the input
485 /// vector. Doing so has the effect of changing the output of a couple of
486 /// tests in ways which make them less useful in testing fused safepoints.
487 template <typename T> static void unique_unsorted(std::vector<T> &vec) {
490 vec.reserve(vec.size());
493 if (seen.insert(V).second) {
499 static std::string GCSafepointPollName("gc.safepoint_poll");
501 static bool isGCSafepointPoll(Function &F) {
502 return F.getName().equals(GCSafepointPollName);
505 bool PlaceSafepoints::runOnFunction(Function &F) {
506 if (F.isDeclaration() || F.empty()) {
507 // This is a declaration, nothing to do. Must exit early to avoid crash in
508 // dom tree calculation
512 bool modified = false;
514 // In various bits below, we rely on the fact that uses are reachable from
515 // defs. When there are basic blocks unreachable from the entry, dominance
516 // and reachablity queries return non-sensical results. Thus, we preprocess
517 // the function to ensure these properties hold.
518 modified |= removeUnreachableBlocks(F);
520 // STEP 1 - Insert the safepoint polling locations. We do not need to
521 // actually insert parse points yet. That will be done for all polls and
522 // calls in a single pass.
524 // Note: With the migration, we need to recompute this for each 'pass'. Once
525 // we merge these, we'll do it once before the analysis
528 std::vector<CallSite> ParsePointNeeded;
530 if (EnableBackedgeSafepoints && !isGCSafepointPoll(F)) {
531 // Construct a pass manager to run the LoopPass backedge logic. We
532 // need the pass manager to handle scheduling all the loop passes
533 // appropriately. Doing this by hand is painful and just not worth messing
534 // with for the moment.
535 FunctionPassManager FPM(F.getParent());
536 bool CanAssumeCallSafepoints = EnableCallSafepoints &&
537 !isGCSafepointPoll(F);
538 PlaceBackedgeSafepointsImpl *PBS =
539 new PlaceBackedgeSafepointsImpl(CanAssumeCallSafepoints);
541 // Note: While the analysis pass itself won't modify the IR, LoopSimplify
542 // (which it depends on) may. i.e. analysis must be recalculated after run
545 // We preserve dominance information when inserting the poll, otherwise
546 // we'd have to recalculate this on every insert
549 // Insert a poll at each point the analysis pass identified
550 for (size_t i = 0; i < PBS->PollLocations.size(); i++) {
551 // We are inserting a poll, the function is modified
554 // The poll location must be the terminator of a loop latch block.
555 TerminatorInst *Term = PBS->PollLocations[i];
557 std::vector<CallSite> ParsePoints;
559 // Split the backedge of the loop and insert the poll within that new
560 // basic block. This creates a loop with two latches per original
561 // latch (which is non-ideal), but this appears to be easier to
562 // optimize in practice than inserting the poll immediately before the
565 // Since this is a latch, at least one of the successors must dominate
566 // it. Its possible that we have a) duplicate edges to the same header
567 // and b) edges to distinct loop headers. We need to insert pools on
568 // each. (Note: This still relies on LoopSimplify.)
569 DenseSet<BasicBlock *> Headers;
570 for (unsigned i = 0; i < Term->getNumSuccessors(); i++) {
571 BasicBlock *Succ = Term->getSuccessor(i);
572 if (DT.dominates(Succ, Term->getParent())) {
573 Headers.insert(Succ);
576 assert(!Headers.empty() && "poll location is not a loop latch?");
578 // The split loop structure here is so that we only need to recalculate
579 // the dominator tree once. Alternatively, we could just keep it up to
580 // date and use a more natural merged loop.
581 DenseSet<BasicBlock *> SplitBackedges;
582 for (BasicBlock *Header : Headers) {
583 BasicBlock *NewBB = SplitEdge(Term->getParent(), Header, nullptr);
584 SplitBackedges.insert(NewBB);
587 for (BasicBlock *NewBB : SplitBackedges) {
588 InsertSafepointPoll(DT, NewBB->getTerminator(), ParsePoints);
589 NumBackedgeSafepoints++;
593 // Split the latch block itself, right before the terminator.
594 InsertSafepointPoll(DT, Term, ParsePoints);
595 NumBackedgeSafepoints++;
598 // Record the parse points for later use
599 ParsePointNeeded.insert(ParsePointNeeded.end(), ParsePoints.begin(),
604 if (EnableEntrySafepoints && !isGCSafepointPoll(F)) {
606 Instruction *term = findLocationForEntrySafepoint(F, DT);
608 // policy choice not to insert?
610 std::vector<CallSite> RuntimeCalls;
611 InsertSafepointPoll(DT, term, RuntimeCalls);
613 NumEntrySafepoints++;
614 ParsePointNeeded.insert(ParsePointNeeded.end(), RuntimeCalls.begin(),
619 if (EnableCallSafepoints && !isGCSafepointPoll(F)) {
621 std::vector<CallSite> Calls;
622 findCallSafepoints(F, Calls);
623 NumCallSafepoints += Calls.size();
624 ParsePointNeeded.insert(ParsePointNeeded.end(), Calls.begin(), Calls.end());
627 // Unique the vectors since we can end up with duplicates if we scan the call
628 // site for call safepoints after we add it for entry or backedge. The
629 // only reason we need tracking at all is that some functions might have
630 // polls but not call safepoints and thus we might miss marking the runtime
631 // calls for the polls. (This is useful in test cases!)
632 unique_unsorted(ParsePointNeeded);
634 // Any parse point (no matter what source) will be handled here
635 DT.recalculate(F); // Needed?
637 // We're about to start modifying the function
638 if (!ParsePointNeeded.empty())
641 // Now run through and insert the safepoints, but do _NOT_ update or remove
642 // any existing uses. We have references to live variables that need to
643 // survive to the last iteration of this loop.
644 std::vector<Value *> Results;
645 Results.reserve(ParsePointNeeded.size());
646 for (size_t i = 0; i < ParsePointNeeded.size(); i++) {
647 CallSite &CS = ParsePointNeeded[i];
648 Value *GCResult = ReplaceWithStatepoint(CS, nullptr);
649 Results.push_back(GCResult);
651 assert(Results.size() == ParsePointNeeded.size());
653 // Adjust all users of the old call sites to use the new ones instead
654 for (size_t i = 0; i < ParsePointNeeded.size(); i++) {
655 CallSite &CS = ParsePointNeeded[i];
656 Value *GCResult = Results[i];
658 // In case if we inserted result in a different basic block than the
659 // original safepoint (this can happen for invokes). We need to be sure
661 // original result value was not used in any of the phi nodes at the
662 // beginning of basic block with gc result. Because we know that all such
663 // blocks will have single predecessor we can safely assume that all phi
664 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
665 // Just remove them all here.
667 FoldSingleEntryPHINodes(cast<Instruction>(GCResult)->getParent(),
670 !isa<PHINode>(cast<Instruction>(GCResult)->getParent()->begin()));
673 // Replace all uses with the new call
674 CS.getInstruction()->replaceAllUsesWith(GCResult);
677 // Now that we've handled all uses, remove the original call itself
678 // Note: The insert point can't be the deleted instruction!
679 CS.getInstruction()->eraseFromParent();
684 char PlaceBackedgeSafepointsImpl::ID = 0;
685 char PlaceSafepoints::ID = 0;
687 ModulePass *llvm::createPlaceSafepointsPass() { return new PlaceSafepoints(); }
689 INITIALIZE_PASS_BEGIN(PlaceBackedgeSafepointsImpl,
690 "place-backedge-safepoints-impl",
691 "Place Backedge Safepoints", false, false)
692 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
693 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
694 INITIALIZE_PASS_END(PlaceBackedgeSafepointsImpl,
695 "place-backedge-safepoints-impl",
696 "Place Backedge Safepoints", false, false)
698 INITIALIZE_PASS_BEGIN(PlaceSafepoints, "place-safepoints", "Place Safepoints",
700 INITIALIZE_PASS_END(PlaceSafepoints, "place-safepoints", "Place Safepoints",
703 static bool isGCLeafFunction(const CallSite &CS) {
704 Instruction *inst = CS.getInstruction();
705 if (isa<IntrinsicInst>(inst)) {
706 // Most LLVM intrinsics are things which can never take a safepoint.
707 // As a result, we don't need to have the stack parsable at the
708 // callsite. This is a highly useful optimization since intrinsic
709 // calls are fairly prevelent, particularly in debug builds.
713 // If this function is marked explicitly as a leaf call, we don't need to
714 // place a safepoint of it. In fact, for correctness we *can't* in many
715 // cases. Note: Indirect calls return Null for the called function,
716 // these obviously aren't runtime functions with attributes
717 // TODO: Support attributes on the call site as well.
718 const Function *F = CS.getCalledFunction();
721 F->getFnAttribute("gc-leaf-function").getValueAsString().equals("true");
729 InsertSafepointPoll(DominatorTree &DT, Instruction *term,
730 std::vector<CallSite> &ParsePointsNeeded /*rval*/) {
731 Module *M = term->getParent()->getParent()->getParent();
734 // Inline the safepoint poll implementation - this will get all the branch,
735 // control flow, etc.. Most importantly, it will introduce the actual slow
736 // path call - where we need to insert a safepoint (parsepoint).
737 FunctionType *ftype =
738 FunctionType::get(Type::getVoidTy(M->getContext()), false);
739 assert(ftype && "null?");
740 // Note: This cast can fail if there's a function of the same name with a
741 // different type inserted previously
743 dyn_cast<Function>(M->getOrInsertFunction("gc.safepoint_poll", ftype));
744 assert(F && !F->empty() && "definition must exist");
745 CallInst *poll = CallInst::Create(F, "", term);
747 // Record some information about the call site we're replacing
748 BasicBlock *OrigBB = term->getParent();
749 BasicBlock::iterator before(poll), after(poll);
751 if (before == term->getParent()->begin()) {
757 assert(after != poll->getParent()->end() && "must have successor");
758 assert(DT.dominates(before, after) && "trivially true");
760 // do the actual inlining
761 InlineFunctionInfo IFI;
762 bool inlineStatus = InlineFunction(poll, IFI);
763 assert(inlineStatus && "inline must succeed");
764 (void)inlineStatus; // suppress warning in release-asserts
766 // Check post conditions
767 assert(IFI.StaticAllocas.empty() && "can't have allocs");
769 std::vector<CallInst *> calls; // new calls
770 std::set<BasicBlock *> BBs; // new BBs + insertee
771 // Include only the newly inserted instructions, Note: begin may not be valid
772 // if we inserted to the beginning of the basic block
773 BasicBlock::iterator start;
775 start = OrigBB->begin();
781 // If your poll function includes an unreachable at the end, that's not
782 // valid. Bugpoint likes to create this, so check for it.
783 assert(isPotentiallyReachable(&*start, &*after, nullptr, nullptr) &&
784 "malformed poll function");
786 scanInlinedCode(&*(start), &*(after), calls, BBs);
788 // Recompute since we've invalidated cached data. Conceptually we
789 // shouldn't need to do this, but implementation wise we appear to. Needed
790 // so we can insert safepoints correctly.
791 // TODO: update more cheaply
792 DT.recalculate(*after->getParent()->getParent());
794 assert(!calls.empty() && "slow path not found for safepoint poll");
796 // Record the fact we need a parsable state at the runtime call contained in
797 // the poll function. This is required so that the runtime knows how to
798 // parse the last frame when we actually take the safepoint (i.e. execute
800 assert(ParsePointsNeeded.empty());
801 for (size_t i = 0; i < calls.size(); i++) {
803 // No safepoint needed or wanted
804 if (!needsStatepoint(calls[i])) {
808 // These are likely runtime calls. Should we assert that via calling
809 // convention or something?
810 ParsePointsNeeded.push_back(CallSite(calls[i]));
812 assert(ParsePointsNeeded.size() <= calls.size());
815 // Normalize basic block to make it ready to be target of invoke statepoint.
816 // It means spliting it to have single predecessor. Return newly created BB
817 // ready to be successor of invoke statepoint.
818 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
819 BasicBlock *InvokeParent) {
820 BasicBlock *ret = BB;
822 if (!BB->getUniquePredecessor()) {
823 ret = SplitBlockPredecessors(BB, InvokeParent, "");
826 // Another requirement for such basic blocks is to not have any phi nodes.
827 // Since we just ensured that new BB will have single predecessor,
828 // all phi nodes in it will have one value. Here it would be naturall place
830 // remove them all. But we can not do this because we are risking to remove
831 // one of the values stored in liveset of another statepoint. We will do it
832 // later after placing all safepoints.
837 /// Replaces the given call site (Call or Invoke) with a gc.statepoint
838 /// intrinsic with an empty deoptimization arguments list. This does
839 /// NOT do explicit relocation for GC support.
840 static Value *ReplaceWithStatepoint(const CallSite &CS, /* to replace */
842 BasicBlock *BB = CS.getInstruction()->getParent();
843 Function *F = BB->getParent();
844 Module *M = F->getParent();
845 assert(M && "must be set");
847 // TODO: technically, a pass is not allowed to get functions from within a
848 // function pass since it might trigger a new function addition. Refactor
849 // this logic out to the initialization of the pass. Doesn't appear to
850 // matter in practice.
852 // Fill in the one generic type'd argument (the function is also vararg)
853 std::vector<Type *> argTypes;
854 argTypes.push_back(CS.getCalledValue()->getType());
856 Function *gc_statepoint_decl = Intrinsic::getDeclaration(
857 M, Intrinsic::experimental_gc_statepoint, argTypes);
859 // Then go ahead and use the builder do actually do the inserts. We insert
860 // immediately before the previous instruction under the assumption that all
861 // arguments will be available here. We can't insert afterwards since we may
862 // be replacing a terminator.
863 Instruction *insertBefore = CS.getInstruction();
864 IRBuilder<> Builder(insertBefore);
865 // First, create the statepoint (with all live ptrs as arguments).
866 std::vector<llvm::Value *> args;
867 // target, #call args, unused, call args..., #deopt args, deopt args..., gc args...
868 Value *Target = CS.getCalledValue();
869 args.push_back(Target);
870 int callArgSize = CS.arg_size();
872 ConstantInt::get(Type::getInt32Ty(M->getContext()), callArgSize));
873 // TODO: add a 'Needs GC-rewrite' later flag
874 args.push_back(ConstantInt::get(Type::getInt32Ty(M->getContext()), 0));
876 // Copy all the arguments of the original call
877 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
879 // # of deopt arguments: this pass currently does not support the
880 // identification of deopt arguments. If this is interesting to you,
881 // please ask on llvm-dev.
882 args.push_back(ConstantInt::get(Type::getInt32Ty(M->getContext()), 0));
884 // Note: The gc args are not filled in at this time, that's handled by
885 // RewriteStatepointsForGC (which is currently under review).
887 // Create the statepoint given all the arguments
888 Instruction *token = nullptr;
889 AttributeSet return_attributes;
891 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
893 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
894 call->setTailCall(toReplace->isTailCall());
895 call->setCallingConv(toReplace->getCallingConv());
897 // Before we have to worry about GC semantics, all attributes are legal
898 AttributeSet new_attrs = toReplace->getAttributes();
899 // In case if we can handle this set of sttributes - set up function attrs
900 // directly on statepoint and return attrs later for gc_result intrinsic.
901 call->setAttributes(new_attrs.getFnAttributes());
902 return_attributes = new_attrs.getRetAttributes();
903 // TODO: handle param attributes
907 // Put the following gc_result and gc_relocate calls immediately after the
908 // the old call (which we're about to delete)
909 BasicBlock::iterator next(toReplace);
910 assert(BB->end() != next && "not a terminator, must have next");
912 Instruction *IP = &*(next);
913 Builder.SetInsertPoint(IP);
914 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
916 } else if (CS.isInvoke()) {
917 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
919 // Insert the new invoke into the old block. We'll remove the old one in a
920 // moment at which point this will become the new terminator for the
922 InvokeInst *invoke = InvokeInst::Create(
923 gc_statepoint_decl, toReplace->getNormalDest(),
924 toReplace->getUnwindDest(), args, "", toReplace->getParent());
925 invoke->setCallingConv(toReplace->getCallingConv());
927 // Currently we will fail on parameter attributes and on certain
928 // function attributes.
929 AttributeSet new_attrs = toReplace->getAttributes();
930 // In case if we can handle this set of sttributes - set up function attrs
931 // directly on statepoint and return attrs later for gc_result intrinsic.
932 invoke->setAttributes(new_attrs.getFnAttributes());
933 return_attributes = new_attrs.getRetAttributes();
937 // We'll insert the gc.result into the normal block
938 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
939 toReplace->getNormalDest(), invoke->getParent());
940 Instruction *IP = &*(normalDest->getFirstInsertionPt());
941 Builder.SetInsertPoint(IP);
943 llvm_unreachable("unexpect type of CallSite");
947 // Handle the return value of the original call - update all uses to use a
948 // gc_result hanging off the statepoint node we just inserted
950 // Only add the gc_result iff there is actually a used result
951 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
952 Instruction *gc_result = nullptr;
953 std::vector<Type *> types; // one per 'any' type
954 types.push_back(CS.getType()); // result type
955 auto get_gc_result_id = [&](Type &Ty) {
956 if (Ty.isIntegerTy()) {
957 return Intrinsic::experimental_gc_result_int;
958 } else if (Ty.isFloatingPointTy()) {
959 return Intrinsic::experimental_gc_result_float;
960 } else if (Ty.isPointerTy()) {
961 return Intrinsic::experimental_gc_result_ptr;
963 llvm_unreachable("non java type encountered");
966 Intrinsic::ID Id = get_gc_result_id(*CS.getType());
967 Value *gc_result_func = Intrinsic::getDeclaration(M, Id, types);
969 std::vector<Value *> args;
970 args.push_back(token);
971 gc_result = Builder.CreateCall(
972 gc_result_func, args,
973 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "");
975 cast<CallInst>(gc_result)->setAttributes(return_attributes);
978 // No return value for the call.