1 //===- CoroFrame.cpp - Builds and manipulates coroutine frame -------------===//
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 //===----------------------------------------------------------------------===//
9 // This file contains classes used to discover if for a particular value
10 // there from sue to definition that crosses a suspend block.
12 // Using the information discovered we form a Coroutine Frame structure to
13 // contain those values. All uses of those values are replaced with appropriate
14 // GEP + load from the coroutine frame. At the point of the definition we spill
15 // the value into the coroutine frame.
17 // TODO: pack values tightly using liveness info.
18 //===----------------------------------------------------------------------===//
20 #include "CoroInternal.h"
21 #include "llvm/ADT/BitVector.h"
22 #include "llvm/IR/CFG.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/InstIterator.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/MathExtras.h"
28 #include "llvm/Support/circular_raw_ostream.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Local.h"
34 // The "coro-suspend-crossing" flag is very noisy. There is another debug type,
35 // "coro-frame", which results in leaner debug spew.
36 #define DEBUG_TYPE "coro-suspend-crossing"
38 enum { SmallVectorThreshold = 32 };
40 // Provides two way mapping between the blocks and numbers.
42 class BlockToIndexMapping {
43 SmallVector<BasicBlock *, SmallVectorThreshold> V;
46 size_t size() const { return V.size(); }
48 BlockToIndexMapping(Function &F) {
49 for (BasicBlock &BB : F)
51 std::sort(V.begin(), V.end());
54 size_t blockToIndex(BasicBlock *BB) const {
55 auto *I = std::lower_bound(V.begin(), V.end(), BB);
56 assert(I != V.end() && *I == BB && "BasicBlockNumberng: Unknown block");
60 BasicBlock *indexToBlock(unsigned Index) const { return V[Index]; }
62 } // end anonymous namespace
64 // The SuspendCrossingInfo maintains data that allows to answer a question
65 // whether given two BasicBlocks A and B there is a path from A to B that
66 // passes through a suspend point.
68 // For every basic block 'i' it maintains a BlockData that consists of:
69 // Consumes: a bit vector which contains a set of indices of blocks that can
71 // Kills: a bit vector which contains a set of indices of blocks that can
72 // reach block 'i', but one of the path will cross a suspend point
73 // Suspend: a boolean indicating whether block 'i' contains a suspend point.
74 // End: a boolean indicating whether block 'i' contains a coro.end intrinsic.
77 struct SuspendCrossingInfo {
78 BlockToIndexMapping Mapping;
86 SmallVector<BlockData, SmallVectorThreshold> Block;
88 iterator_range<succ_iterator> successors(BlockData const &BD) const {
89 BasicBlock *BB = Mapping.indexToBlock(&BD - &Block[0]);
90 return llvm::successors(BB);
93 BlockData &getBlockData(BasicBlock *BB) {
94 return Block[Mapping.blockToIndex(BB)];
98 void dump(StringRef Label, BitVector const &BV) const;
100 SuspendCrossingInfo(Function &F, coro::Shape &Shape);
102 bool hasPathCrossingSuspendPoint(BasicBlock *DefBB, BasicBlock *UseBB) const {
103 size_t const DefIndex = Mapping.blockToIndex(DefBB);
104 size_t const UseIndex = Mapping.blockToIndex(UseBB);
106 assert(Block[UseIndex].Consumes[DefIndex] && "use must consume def");
107 bool const Result = Block[UseIndex].Kills[DefIndex];
108 DEBUG(dbgs() << UseBB->getName() << " => " << DefBB->getName()
109 << " answer is " << Result << "\n");
113 bool isDefinitionAcrossSuspend(BasicBlock *DefBB, User *U) const {
114 auto *I = cast<Instruction>(U);
116 // We rewrote PHINodes, so that only the ones with exactly one incoming
117 // value need to be analyzed.
118 if (auto *PN = dyn_cast<PHINode>(I))
119 if (PN->getNumIncomingValues() > 1)
122 BasicBlock *UseBB = I->getParent();
123 return hasPathCrossingSuspendPoint(DefBB, UseBB);
126 bool isDefinitionAcrossSuspend(Argument &A, User *U) const {
127 return isDefinitionAcrossSuspend(&A.getParent()->getEntryBlock(), U);
130 bool isDefinitionAcrossSuspend(Instruction &I, User *U) const {
131 return isDefinitionAcrossSuspend(I.getParent(), U);
134 } // end anonymous namespace
136 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
137 LLVM_DUMP_METHOD void SuspendCrossingInfo::dump(StringRef Label,
138 BitVector const &BV) const {
139 dbgs() << Label << ":";
140 for (size_t I = 0, N = BV.size(); I < N; ++I)
142 dbgs() << " " << Mapping.indexToBlock(I)->getName();
146 LLVM_DUMP_METHOD void SuspendCrossingInfo::dump() const {
147 for (size_t I = 0, N = Block.size(); I < N; ++I) {
148 BasicBlock *const B = Mapping.indexToBlock(I);
149 dbgs() << B->getName() << ":\n";
150 dump(" Consumes", Block[I].Consumes);
151 dump(" Kills", Block[I].Kills);
157 SuspendCrossingInfo::SuspendCrossingInfo(Function &F, coro::Shape &Shape)
159 const size_t N = Mapping.size();
162 // Initialize every block so that it consumes itself
163 for (size_t I = 0; I < N; ++I) {
165 B.Consumes.resize(N);
170 // Mark all CoroEnd Blocks. We do not propagate Kills beyond coro.ends as
171 // the code beyond coro.end is reachable during initial invocation of the
173 for (auto *CE : Shape.CoroEnds)
174 getBlockData(CE->getParent()).End = true;
176 // Mark all suspend blocks and indicate that they kill everything they
177 // consume. Note, that crossing coro.save also requires a spill, as any code
178 // between coro.save and coro.suspend may resume the coroutine and all of the
179 // state needs to be saved by that time.
180 auto markSuspendBlock = [&](IntrinsicInst *BarrierInst) {
181 BasicBlock *SuspendBlock = BarrierInst->getParent();
182 auto &B = getBlockData(SuspendBlock);
184 B.Kills |= B.Consumes;
186 for (CoroSuspendInst *CSI : Shape.CoroSuspends) {
187 markSuspendBlock(CSI);
188 markSuspendBlock(CSI->getCoroSave());
191 // Iterate propagating consumes and kills until they stop changing.
197 DEBUG(dbgs() << "iteration " << ++Iteration);
198 DEBUG(dbgs() << "==============\n");
201 for (size_t I = 0; I < N; ++I) {
203 for (BasicBlock *SI : successors(B)) {
205 auto SuccNo = Mapping.blockToIndex(SI);
207 // Saved Consumes and Kills bitsets so that it is easy to see
208 // if anything changed after propagation.
209 auto &S = Block[SuccNo];
210 auto SavedConsumes = S.Consumes;
211 auto SavedKills = S.Kills;
213 // Propagate Kills and Consumes from block B into its successor S.
214 S.Consumes |= B.Consumes;
217 // If block B is a suspend block, it should propagate kills into the
218 // its successor for every block B consumes.
220 S.Kills |= B.Consumes;
223 // If block S is a suspend block, it should kill all of the blocks it
225 S.Kills |= S.Consumes;
227 // If block S is an end block, it should not propagate kills as the
228 // blocks following coro.end() are reached during initial invocation
229 // of the coroutine while all the data are still available on the
230 // stack or in the registers.
233 // This is reached when S block it not Suspend nor coro.end and it
234 // need to make sure that it is not in the kill set.
235 S.Kills.reset(SuccNo);
238 // See if anything changed.
239 Changed |= (S.Kills != SavedKills) || (S.Consumes != SavedConsumes);
241 if (S.Kills != SavedKills) {
242 DEBUG(dbgs() << "\nblock " << I << " follower " << SI->getName()
244 DEBUG(dump("S.Kills", S.Kills));
245 DEBUG(dump("SavedKills", SavedKills));
247 if (S.Consumes != SavedConsumes) {
248 DEBUG(dbgs() << "\nblock " << I << " follower " << SI << "\n");
249 DEBUG(dump("S.Consume", S.Consumes));
250 DEBUG(dump("SavedCons", SavedConsumes));
258 #undef DEBUG_TYPE // "coro-suspend-crossing"
259 #define DEBUG_TYPE "coro-frame"
261 // We build up the list of spills for every case where a use is separated
262 // from the definition by a suspend point.
264 struct Spill : std::pair<Value *, Instruction *> {
265 using base = std::pair<Value *, Instruction *>;
267 Spill(Value *Def, User *U) : base(Def, cast<Instruction>(U)) {}
269 Value *def() const { return first; }
270 Instruction *user() const { return second; }
271 BasicBlock *userBlock() const { return second->getParent(); }
273 std::pair<Value *, BasicBlock *> getKey() const {
274 return {def(), userBlock()};
277 bool operator<(Spill const &rhs) const { return getKey() < rhs.getKey(); }
280 // Note that there may be more than one record with the same value of Def in
281 // the SpillInfo vector.
282 using SpillInfo = SmallVector<Spill, 8>;
285 static void dump(StringRef Title, SpillInfo const &Spills) {
286 dbgs() << "------------- " << Title << "--------------\n";
287 Value *CurrentValue = nullptr;
288 for (auto const &E : Spills) {
289 if (CurrentValue != E.def()) {
290 CurrentValue = E.def();
291 CurrentValue->dump();
299 // Build a struct that will keep state for an active coroutine.
301 // ResumeFnTy ResumeFnAddr;
302 // ResumeFnTy DestroyFnAddr;
304 // ... promise (if present) ...
307 static StructType *buildFrameType(Function &F, coro::Shape &Shape,
309 LLVMContext &C = F.getContext();
310 SmallString<32> Name(F.getName());
311 Name.append(".Frame");
312 StructType *FrameTy = StructType::create(C, Name);
313 auto *FramePtrTy = FrameTy->getPointerTo();
314 auto *FnTy = FunctionType::get(Type::getVoidTy(C), FramePtrTy,
315 /*IsVarArgs=*/false);
316 auto *FnPtrTy = FnTy->getPointerTo();
318 // Figure out how wide should be an integer type storing the suspend index.
319 unsigned IndexBits = std::max(1U, Log2_64_Ceil(Shape.CoroSuspends.size()));
320 Type *PromiseType = Shape.PromiseAlloca
321 ? Shape.PromiseAlloca->getType()->getElementType()
322 : Type::getInt1Ty(C);
323 SmallVector<Type *, 8> Types{FnPtrTy, FnPtrTy, PromiseType,
324 Type::getIntNTy(C, IndexBits)};
325 Value *CurrentDef = nullptr;
327 // Create an entry for every spilled value.
328 for (auto const &S : Spills) {
329 if (CurrentDef == S.def())
332 CurrentDef = S.def();
333 // PromiseAlloca was already added to Types array earlier.
334 if (CurrentDef == Shape.PromiseAlloca)
338 if (auto *AI = dyn_cast<AllocaInst>(CurrentDef))
339 Ty = AI->getAllocatedType();
341 Ty = CurrentDef->getType();
345 FrameTy->setBody(Types);
350 // We need to make room to insert a spill after initial PHIs, but before
351 // catchswitch instruction. Placing it before violates the requirement that
352 // catchswitch, like all other EHPads must be the first nonPHI in a block.
354 // Split away catchswitch into a separate block and insert in its place:
356 // cleanuppad <InsertPt> cleanupret.
358 // cleanupret instruction will act as an insert point for the spill.
359 static Instruction *splitBeforeCatchSwitch(CatchSwitchInst *CatchSwitch) {
360 BasicBlock *CurrentBlock = CatchSwitch->getParent();
361 BasicBlock *NewBlock = CurrentBlock->splitBasicBlock(CatchSwitch);
362 CurrentBlock->getTerminator()->eraseFromParent();
365 CleanupPadInst::Create(CatchSwitch->getParentPad(), {}, "", CurrentBlock);
367 CleanupReturnInst::Create(CleanupPad, NewBlock, CurrentBlock);
371 // Replace all alloca and SSA values that are accessed across suspend points
372 // with GetElementPointer from coroutine frame + loads and stores. Create an
373 // AllocaSpillBB that will become the new entry block for the resume parts of
376 // %hdl = coro.begin(...)
381 // %hdl = coro.begin(...)
382 // %FramePtr = bitcast i8* hdl to %f.frame*
383 // br label %AllocaSpillBB
386 // ; geps corresponding to allocas that were moved to coroutine frame
387 // br label PostSpill
393 static Instruction *insertSpills(SpillInfo &Spills, coro::Shape &Shape) {
394 auto *CB = Shape.CoroBegin;
395 IRBuilder<> Builder(CB->getNextNode());
396 PointerType *FramePtrTy = Shape.FrameTy->getPointerTo();
398 cast<Instruction>(Builder.CreateBitCast(CB, FramePtrTy, "FramePtr"));
399 Type *FrameTy = FramePtrTy->getElementType();
401 Value *CurrentValue = nullptr;
402 BasicBlock *CurrentBlock = nullptr;
403 Value *CurrentReload = nullptr;
404 unsigned Index = coro::Shape::LastKnownField;
406 // We need to keep track of any allocas that need "spilling"
407 // since they will live in the coroutine frame now, all access to them
408 // need to be changed, not just the access across suspend points
409 // we remember allocas and their indices to be handled once we processed
411 SmallVector<std::pair<AllocaInst *, unsigned>, 4> Allocas;
412 // Promise alloca (if present) has a fixed field number (Shape::PromiseField)
413 if (Shape.PromiseAlloca)
414 Allocas.emplace_back(Shape.PromiseAlloca, coro::Shape::PromiseField);
416 // Create a load instruction to reload the spilled value from the coroutine
418 auto CreateReload = [&](Instruction *InsertBefore) {
419 Builder.SetInsertPoint(InsertBefore);
420 auto *G = Builder.CreateConstInBoundsGEP2_32(FrameTy, FramePtr, 0, Index,
421 CurrentValue->getName() +
422 Twine(".reload.addr"));
423 return isa<AllocaInst>(CurrentValue)
425 : Builder.CreateLoad(G,
426 CurrentValue->getName() + Twine(".reload"));
429 for (auto const &E : Spills) {
430 // If we have not seen the value, generate a spill.
431 if (CurrentValue != E.def()) {
432 CurrentValue = E.def();
433 CurrentBlock = nullptr;
434 CurrentReload = nullptr;
438 if (auto *AI = dyn_cast<AllocaInst>(CurrentValue)) {
439 // Spilled AllocaInst will be replaced with GEP from the coroutine frame
440 // there is no spill required.
441 Allocas.emplace_back(AI, Index);
442 if (!AI->isStaticAlloca())
443 report_fatal_error("Coroutines cannot handle non static allocas yet");
445 // Otherwise, create a store instruction storing the value into the
448 Instruction *InsertPt = nullptr;
449 if (isa<Argument>(CurrentValue)) {
450 // For arguments, we will place the store instruction right after
451 // the coroutine frame pointer instruction, i.e. bitcast of
452 // coro.begin from i8* to %f.frame*.
453 InsertPt = FramePtr->getNextNode();
454 } else if (auto *II = dyn_cast<InvokeInst>(CurrentValue)) {
455 // If we are spilling the result of the invoke instruction, split the
456 // normal edge and insert the spill in the new block.
457 auto NewBB = SplitEdge(II->getParent(), II->getNormalDest());
458 InsertPt = NewBB->getTerminator();
459 } else if (dyn_cast<PHINode>(CurrentValue)) {
460 // Skip the PHINodes and EH pads instructions.
461 BasicBlock *DefBlock = cast<Instruction>(E.def())->getParent();
462 if (auto *CSI = dyn_cast<CatchSwitchInst>(DefBlock->getTerminator()))
463 InsertPt = splitBeforeCatchSwitch(CSI);
465 InsertPt = &*DefBlock->getFirstInsertionPt();
467 // For all other values, the spill is placed immediately after
469 assert(!isa<TerminatorInst>(E.def()) && "unexpected terminator");
470 InsertPt = cast<Instruction>(E.def())->getNextNode();
473 Builder.SetInsertPoint(InsertPt);
474 auto *G = Builder.CreateConstInBoundsGEP2_32(
475 FrameTy, FramePtr, 0, Index,
476 CurrentValue->getName() + Twine(".spill.addr"));
477 Builder.CreateStore(CurrentValue, G);
481 // If we have not seen the use block, generate a reload in it.
482 if (CurrentBlock != E.userBlock()) {
483 CurrentBlock = E.userBlock();
484 CurrentReload = CreateReload(&*CurrentBlock->getFirstInsertionPt());
487 // If we have a single edge PHINode, remove it and replace it with a reload
488 // from the coroutine frame. (We already took care of multi edge PHINodes
489 // by rewriting them in the rewritePHIs function).
490 if (auto *PN = dyn_cast<PHINode>(E.user())) {
491 assert(PN->getNumIncomingValues() == 1 && "unexpected number of incoming "
492 "values in the PHINode");
493 PN->replaceAllUsesWith(CurrentReload);
494 PN->eraseFromParent();
498 // Replace all uses of CurrentValue in the current instruction with reload.
499 E.user()->replaceUsesOfWith(CurrentValue, CurrentReload);
502 BasicBlock *FramePtrBB = FramePtr->getParent();
503 Shape.AllocaSpillBlock =
504 FramePtrBB->splitBasicBlock(FramePtr->getNextNode(), "AllocaSpillBB");
505 Shape.AllocaSpillBlock->splitBasicBlock(&Shape.AllocaSpillBlock->front(),
508 Builder.SetInsertPoint(&Shape.AllocaSpillBlock->front());
509 // If we found any allocas, replace all of their remaining uses with Geps.
510 for (auto &P : Allocas) {
512 Builder.CreateConstInBoundsGEP2_32(FrameTy, FramePtr, 0, P.second);
513 // We are not using ReplaceInstWithInst(P.first, cast<Instruction>(G)) here,
514 // as we are changing location of the instruction.
515 G->takeName(P.first);
516 P.first->replaceAllUsesWith(G);
517 P.first->eraseFromParent();
522 // Sets the unwind edge of an instruction to a particular successor.
523 static void setUnwindEdgeTo(TerminatorInst *TI, BasicBlock *Succ) {
524 if (auto *II = dyn_cast<InvokeInst>(TI))
525 II->setUnwindDest(Succ);
526 else if (auto *CS = dyn_cast<CatchSwitchInst>(TI))
527 CS->setUnwindDest(Succ);
528 else if (auto *CR = dyn_cast<CleanupReturnInst>(TI))
529 CR->setUnwindDest(Succ);
531 llvm_unreachable("unexpected terminator instruction");
534 // Replaces all uses of OldPred with the NewPred block in all PHINodes in a
536 static void updatePhiNodes(BasicBlock *DestBB, BasicBlock *OldPred,
538 PHINode *LandingPadReplacement) {
540 for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) {
541 PHINode *PN = cast<PHINode>(I);
543 // We manually update the LandingPadReplacement PHINode and it is the last
544 // PHI Node. So, if we find it, we are done.
545 if (LandingPadReplacement == PN)
548 // Reuse the previous value of BBIdx if it lines up. In cases where we
549 // have multiple phi nodes with *lots* of predecessors, this is a speed
550 // win because we don't have to scan the PHI looking for TIBB. This
551 // happens because the BB list of PHI nodes are usually in the same
553 if (PN->getIncomingBlock(BBIdx) != OldPred)
554 BBIdx = PN->getBasicBlockIndex(OldPred);
556 assert(BBIdx != (unsigned)-1 && "Invalid PHI Index!");
557 PN->setIncomingBlock(BBIdx, NewPred);
561 // Uses SplitEdge unless the successor block is an EHPad, in which case do EH
562 // specific handling.
563 static BasicBlock *ehAwareSplitEdge(BasicBlock *BB, BasicBlock *Succ,
564 LandingPadInst *OriginalPad,
565 PHINode *LandingPadReplacement) {
566 auto *PadInst = Succ->getFirstNonPHI();
567 if (!LandingPadReplacement && !PadInst->isEHPad())
568 return SplitEdge(BB, Succ);
570 auto *NewBB = BasicBlock::Create(BB->getContext(), "", BB->getParent(), Succ);
571 setUnwindEdgeTo(BB->getTerminator(), NewBB);
572 updatePhiNodes(Succ, BB, NewBB, LandingPadReplacement);
574 if (LandingPadReplacement) {
575 auto *NewLP = OriginalPad->clone();
576 auto *Terminator = BranchInst::Create(Succ, NewBB);
577 NewLP->insertBefore(Terminator);
578 LandingPadReplacement->addIncoming(NewLP, NewBB);
581 Value *ParentPad = nullptr;
582 if (auto *FuncletPad = dyn_cast<FuncletPadInst>(PadInst))
583 ParentPad = FuncletPad->getParentPad();
584 else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(PadInst))
585 ParentPad = CatchSwitch->getParentPad();
587 llvm_unreachable("handling for other EHPads not implemented yet");
589 auto *NewCleanupPad = CleanupPadInst::Create(ParentPad, {}, "", NewBB);
590 CleanupReturnInst::Create(NewCleanupPad, Succ, NewBB);
594 static void rewritePHIs(BasicBlock &BB) {
595 // For every incoming edge we will create a block holding all
596 // incoming values in a single PHI nodes.
599 // %n.val = phi i32[%n, %entry], [%inc, %loop]
604 // %n.loop.pre = phi i32 [%n, %entry]
607 // %inc.loop.pre = phi i32 [%inc, %loop]
610 // After this rewrite, further analysis will ignore any phi nodes with more
611 // than one incoming edge.
613 // TODO: Simplify PHINodes in the basic block to remove duplicate
616 LandingPadInst *LandingPad = nullptr;
617 PHINode *ReplPHI = nullptr;
618 if ((LandingPad = dyn_cast_or_null<LandingPadInst>(BB.getFirstNonPHI()))) {
619 // ehAwareSplitEdge will clone the LandingPad in all the edge blocks.
620 // We replace the original landing pad with a PHINode that will collect the
621 // results from all of them.
622 ReplPHI = PHINode::Create(LandingPad->getType(), 1, "", LandingPad);
623 ReplPHI->takeName(LandingPad);
624 LandingPad->replaceAllUsesWith(ReplPHI);
625 // We will erase the original landing pad at the end of this function after
626 // ehAwareSplitEdge cloned it in the transition blocks.
629 SmallVector<BasicBlock *, 8> Preds(pred_begin(&BB), pred_end(&BB));
630 for (BasicBlock *Pred : Preds) {
631 auto *IncomingBB = ehAwareSplitEdge(Pred, &BB, LandingPad, ReplPHI);
632 IncomingBB->setName(BB.getName() + Twine(".from.") + Pred->getName());
633 auto *PN = cast<PHINode>(&BB.front());
635 int Index = PN->getBasicBlockIndex(IncomingBB);
636 Value *V = PN->getIncomingValue(Index);
637 PHINode *InputV = PHINode::Create(
638 V->getType(), 1, V->getName() + Twine(".") + BB.getName(),
639 &IncomingBB->front());
640 InputV->addIncoming(V, Pred);
641 PN->setIncomingValue(Index, InputV);
642 PN = dyn_cast<PHINode>(PN->getNextNode());
643 } while (PN != ReplPHI); // ReplPHI is either null or the PHI that replaced
648 // Calls to ehAwareSplitEdge function cloned the original lading pad.
649 // No longer need it.
650 LandingPad->eraseFromParent();
654 static void rewritePHIs(Function &F) {
655 SmallVector<BasicBlock *, 8> WorkList;
657 for (BasicBlock &BB : F)
658 if (auto *PN = dyn_cast<PHINode>(&BB.front()))
659 if (PN->getNumIncomingValues() > 1)
660 WorkList.push_back(&BB);
662 for (BasicBlock *BB : WorkList)
666 // Check for instructions that we can recreate on resume as opposed to spill
667 // the result into a coroutine frame.
668 static bool materializable(Instruction &V) {
669 return isa<CastInst>(&V) || isa<GetElementPtrInst>(&V) ||
670 isa<BinaryOperator>(&V) || isa<CmpInst>(&V) || isa<SelectInst>(&V);
673 // Check for structural coroutine intrinsics that should not be spilled into
674 // the coroutine frame.
675 static bool isCoroutineStructureIntrinsic(Instruction &I) {
676 return isa<CoroIdInst>(&I) || isa<CoroBeginInst>(&I) ||
677 isa<CoroSaveInst>(&I) || isa<CoroSuspendInst>(&I);
680 // For every use of the value that is across suspend point, recreate that value
681 // after a suspend point.
682 static void rewriteMaterializableInstructions(IRBuilder<> &IRB,
683 SpillInfo const &Spills) {
684 BasicBlock *CurrentBlock = nullptr;
685 Instruction *CurrentMaterialization = nullptr;
686 Instruction *CurrentDef = nullptr;
688 for (auto const &E : Spills) {
689 // If it is a new definition, update CurrentXXX variables.
690 if (CurrentDef != E.def()) {
691 CurrentDef = cast<Instruction>(E.def());
692 CurrentBlock = nullptr;
693 CurrentMaterialization = nullptr;
696 // If we have not seen this block, materialize the value.
697 if (CurrentBlock != E.userBlock()) {
698 CurrentBlock = E.userBlock();
699 CurrentMaterialization = cast<Instruction>(CurrentDef)->clone();
700 CurrentMaterialization->setName(CurrentDef->getName());
701 CurrentMaterialization->insertBefore(
702 &*CurrentBlock->getFirstInsertionPt());
705 if (auto *PN = dyn_cast<PHINode>(E.user())) {
706 assert(PN->getNumIncomingValues() == 1 && "unexpected number of incoming "
707 "values in the PHINode");
708 PN->replaceAllUsesWith(CurrentMaterialization);
709 PN->eraseFromParent();
713 // Replace all uses of CurrentDef in the current instruction with the
714 // CurrentMaterialization for the block.
715 E.user()->replaceUsesOfWith(CurrentDef, CurrentMaterialization);
719 // Move early uses of spilled variable after CoroBegin.
720 // For example, if a parameter had address taken, we may end up with the code
722 // define @f(i32 %n) {
723 // %n.addr = alloca i32
727 // we need to move the store after coro.begin
728 static void moveSpillUsesAfterCoroBegin(Function &F, SpillInfo const &Spills,
729 CoroBeginInst *CoroBegin) {
731 SmallVector<Instruction *, 8> NeedsMoving;
733 Value *CurrentValue = nullptr;
735 for (auto const &E : Spills) {
736 if (CurrentValue == E.def())
739 CurrentValue = E.def();
741 for (User *U : CurrentValue->users()) {
742 Instruction *I = cast<Instruction>(U);
743 if (!DT.dominates(CoroBegin, I)) {
744 // TODO: Make this more robust. Currently if we run into a situation
745 // where simple instruction move won't work we panic and
746 // report_fatal_error.
747 for (User *UI : I->users()) {
748 if (!DT.dominates(CoroBegin, cast<Instruction>(UI)))
749 report_fatal_error("cannot move instruction since its users are not"
750 " dominated by CoroBegin");
753 DEBUG(dbgs() << "will move: " << *I << "\n");
754 NeedsMoving.push_back(I);
759 Instruction *InsertPt = CoroBegin->getNextNode();
760 for (Instruction *I : NeedsMoving)
761 I->moveBefore(InsertPt);
764 // Splits the block at a particular instruction unless it is the first
765 // instruction in the block with a single predecessor.
766 static BasicBlock *splitBlockIfNotFirst(Instruction *I, const Twine &Name) {
767 auto *BB = I->getParent();
768 if (&BB->front() == I) {
769 if (BB->getSinglePredecessor()) {
774 return BB->splitBasicBlock(I, Name);
777 // Split above and below a particular instruction so that it
778 // will be all alone by itself in a block.
779 static void splitAround(Instruction *I, const Twine &Name) {
780 splitBlockIfNotFirst(I, Name);
781 splitBlockIfNotFirst(I->getNextNode(), "After" + Name);
784 void coro::buildCoroutineFrame(Function &F, Shape &Shape) {
785 // Lower coro.dbg.declare to coro.dbg.value, since we are going to rewrite
786 // access to local variables.
789 Shape.PromiseAlloca = Shape.CoroBegin->getId()->getPromise();
790 if (Shape.PromiseAlloca) {
791 Shape.CoroBegin->getId()->clearPromise();
794 // Make sure that all coro.save, coro.suspend and the fallthrough coro.end
795 // intrinsics are in their own blocks to simplify the logic of building up
796 // SuspendCrossing data.
797 for (CoroSuspendInst *CSI : Shape.CoroSuspends) {
798 splitAround(CSI->getCoroSave(), "CoroSave");
799 splitAround(CSI, "CoroSuspend");
802 // Put fallthrough CoroEnd into its own block. Note: Shape::buildFrom places
803 // the fallthrough coro.end as the first element of CoroEnds array.
804 splitAround(Shape.CoroEnds.front(), "CoroEnd");
806 // Transforms multi-edge PHI Nodes, so that any value feeding into a PHI will
807 // never has its definition separated from the PHI by the suspend point.
810 // Build suspend crossing info.
811 SuspendCrossingInfo Checker(F, Shape);
813 IRBuilder<> Builder(F.getContext());
816 // See if there are materializable instructions across suspend points.
817 for (Instruction &I : instructions(F))
818 if (materializable(I))
819 for (User *U : I.users())
820 if (Checker.isDefinitionAcrossSuspend(I, U))
821 Spills.emplace_back(&I, U);
823 // Rewrite materializable instructions to be materialized at the use point.
824 DEBUG(dump("Materializations", Spills));
825 rewriteMaterializableInstructions(Builder, Spills);
827 // Collect the spills for arguments and other not-materializable values.
829 for (Argument &A : F.args())
830 for (User *U : A.users())
831 if (Checker.isDefinitionAcrossSuspend(A, U))
832 Spills.emplace_back(&A, U);
834 for (Instruction &I : instructions(F)) {
835 // Values returned from coroutine structure intrinsics should not be part
836 // of the Coroutine Frame.
837 if (isCoroutineStructureIntrinsic(I))
839 // The Coroutine Promise always included into coroutine frame, no need to
840 // check for suspend crossing.
841 if (Shape.PromiseAlloca == &I)
844 for (User *U : I.users())
845 if (Checker.isDefinitionAcrossSuspend(I, U)) {
846 // We cannot spill a token.
847 if (I.getType()->isTokenTy())
849 "token definition is separated from the use by a suspend point");
850 assert(!materializable(I) &&
851 "rewriteMaterializable did not do its job");
852 Spills.emplace_back(&I, U);
855 DEBUG(dump("Spills", Spills));
856 moveSpillUsesAfterCoroBegin(F, Spills, Shape.CoroBegin);
857 Shape.FrameTy = buildFrameType(F, Shape, Spills);
858 Shape.FramePtr = insertSpills(Spills, Shape);