1 | //===- CoroFrame.cpp - Builds and manipulates coroutine frame -------------===// |
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 | // This file contains classes used to discover if for a particular value |
9 | // there from sue to definition that crosses a suspend block. |
10 | // |
11 | // Using the information discovered we form a Coroutine Frame structure to |
12 | // contain those values. All uses of those values are replaced with appropriate |
13 | // GEP + load from the coroutine frame. At the point of the definition we spill |
14 | // the value into the coroutine frame. |
15 | //===----------------------------------------------------------------------===// |
16 | |
17 | #include "CoroInternal.h" |
18 | #include "llvm/ADT/BitVector.h" |
19 | #include "llvm/ADT/PostOrderIterator.h" |
20 | #include "llvm/ADT/ScopeExit.h" |
21 | #include "llvm/ADT/SmallString.h" |
22 | #include "llvm/Analysis/CFG.h" |
23 | #include "llvm/Analysis/PtrUseVisitor.h" |
24 | #include "llvm/Analysis/StackLifetime.h" |
25 | #include "llvm/Config/llvm-config.h" |
26 | #include "llvm/IR/CFG.h" |
27 | #include "llvm/IR/DIBuilder.h" |
28 | #include "llvm/IR/DebugInfo.h" |
29 | #include "llvm/IR/Dominators.h" |
30 | #include "llvm/IR/IRBuilder.h" |
31 | #include "llvm/IR/InstIterator.h" |
32 | #include "llvm/IR/IntrinsicInst.h" |
33 | #include "llvm/Support/Debug.h" |
34 | #include "llvm/Support/MathExtras.h" |
35 | #include "llvm/Support/OptimizedStructLayout.h" |
36 | #include "llvm/Support/circular_raw_ostream.h" |
37 | #include "llvm/Support/raw_ostream.h" |
38 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
39 | #include "llvm/Transforms/Utils/Local.h" |
40 | #include "llvm/Transforms/Utils/PromoteMemToReg.h" |
41 | #include <algorithm> |
42 | #include <deque> |
43 | #include <optional> |
44 | |
45 | using namespace llvm; |
46 | |
47 | extern cl::opt<bool> UseNewDbgInfoFormat; |
48 | |
49 | // The "coro-suspend-crossing" flag is very noisy. There is another debug type, |
50 | // "coro-frame", which results in leaner debug spew. |
51 | #define DEBUG_TYPE "coro-suspend-crossing" |
52 | |
53 | enum { SmallVectorThreshold = 32 }; |
54 | |
55 | // Provides two way mapping between the blocks and numbers. |
56 | namespace { |
57 | class BlockToIndexMapping { |
58 | SmallVector<BasicBlock *, SmallVectorThreshold> V; |
59 | |
60 | public: |
61 | size_t size() const { return V.size(); } |
62 | |
63 | BlockToIndexMapping(Function &F) { |
64 | for (BasicBlock &BB : F) |
65 | V.push_back(Elt: &BB); |
66 | llvm::sort(C&: V); |
67 | } |
68 | |
69 | size_t blockToIndex(BasicBlock const *BB) const { |
70 | auto *I = llvm::lower_bound(Range: V, Value&: BB); |
71 | assert(I != V.end() && *I == BB && "BasicBlockNumberng: Unknown block" ); |
72 | return I - V.begin(); |
73 | } |
74 | |
75 | BasicBlock *indexToBlock(unsigned Index) const { return V[Index]; } |
76 | }; |
77 | } // end anonymous namespace |
78 | |
79 | // The SuspendCrossingInfo maintains data that allows to answer a question |
80 | // whether given two BasicBlocks A and B there is a path from A to B that |
81 | // passes through a suspend point. |
82 | // |
83 | // For every basic block 'i' it maintains a BlockData that consists of: |
84 | // Consumes: a bit vector which contains a set of indices of blocks that can |
85 | // reach block 'i'. A block can trivially reach itself. |
86 | // Kills: a bit vector which contains a set of indices of blocks that can |
87 | // reach block 'i' but there is a path crossing a suspend point |
88 | // not repeating 'i' (path to 'i' without cycles containing 'i'). |
89 | // Suspend: a boolean indicating whether block 'i' contains a suspend point. |
90 | // End: a boolean indicating whether block 'i' contains a coro.end intrinsic. |
91 | // KillLoop: There is a path from 'i' to 'i' not otherwise repeating 'i' that |
92 | // crosses a suspend point. |
93 | // |
94 | namespace { |
95 | class SuspendCrossingInfo { |
96 | BlockToIndexMapping Mapping; |
97 | |
98 | struct BlockData { |
99 | BitVector Consumes; |
100 | BitVector Kills; |
101 | bool Suspend = false; |
102 | bool End = false; |
103 | bool KillLoop = false; |
104 | bool Changed = false; |
105 | }; |
106 | SmallVector<BlockData, SmallVectorThreshold> Block; |
107 | |
108 | iterator_range<pred_iterator> predecessors(BlockData const &BD) const { |
109 | BasicBlock *BB = Mapping.indexToBlock(Index: &BD - &Block[0]); |
110 | return llvm::predecessors(BB); |
111 | } |
112 | |
113 | BlockData &getBlockData(BasicBlock *BB) { |
114 | return Block[Mapping.blockToIndex(BB)]; |
115 | } |
116 | |
117 | /// Compute the BlockData for the current function in one iteration. |
118 | /// Initialize - Whether this is the first iteration, we can optimize |
119 | /// the initial case a little bit by manual loop switch. |
120 | /// Returns whether the BlockData changes in this iteration. |
121 | template <bool Initialize = false> |
122 | bool computeBlockData(const ReversePostOrderTraversal<Function *> &RPOT); |
123 | |
124 | public: |
125 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
126 | void dump() const; |
127 | void dump(StringRef Label, BitVector const &BV) const; |
128 | #endif |
129 | |
130 | SuspendCrossingInfo(Function &F, coro::Shape &Shape); |
131 | |
132 | /// Returns true if there is a path from \p From to \p To crossing a suspend |
133 | /// point without crossing \p From a 2nd time. |
134 | bool hasPathCrossingSuspendPoint(BasicBlock *From, BasicBlock *To) const { |
135 | size_t const FromIndex = Mapping.blockToIndex(BB: From); |
136 | size_t const ToIndex = Mapping.blockToIndex(BB: To); |
137 | bool const Result = Block[ToIndex].Kills[FromIndex]; |
138 | LLVM_DEBUG(dbgs() << From->getName() << " => " << To->getName() |
139 | << " answer is " << Result << "\n" ); |
140 | return Result; |
141 | } |
142 | |
143 | /// Returns true if there is a path from \p From to \p To crossing a suspend |
144 | /// point without crossing \p From a 2nd time. If \p From is the same as \p To |
145 | /// this will also check if there is a looping path crossing a suspend point. |
146 | bool hasPathOrLoopCrossingSuspendPoint(BasicBlock *From, |
147 | BasicBlock *To) const { |
148 | size_t const FromIndex = Mapping.blockToIndex(BB: From); |
149 | size_t const ToIndex = Mapping.blockToIndex(BB: To); |
150 | bool Result = Block[ToIndex].Kills[FromIndex] || |
151 | (From == To && Block[ToIndex].KillLoop); |
152 | LLVM_DEBUG(dbgs() << From->getName() << " => " << To->getName() |
153 | << " answer is " << Result << " (path or loop)\n" ); |
154 | return Result; |
155 | } |
156 | |
157 | bool isDefinitionAcrossSuspend(BasicBlock *DefBB, User *U) const { |
158 | auto *I = cast<Instruction>(Val: U); |
159 | |
160 | // We rewrote PHINodes, so that only the ones with exactly one incoming |
161 | // value need to be analyzed. |
162 | if (auto *PN = dyn_cast<PHINode>(Val: I)) |
163 | if (PN->getNumIncomingValues() > 1) |
164 | return false; |
165 | |
166 | BasicBlock *UseBB = I->getParent(); |
167 | |
168 | // As a special case, treat uses by an llvm.coro.suspend.retcon or an |
169 | // llvm.coro.suspend.async as if they were uses in the suspend's single |
170 | // predecessor: the uses conceptually occur before the suspend. |
171 | if (isa<CoroSuspendRetconInst>(Val: I) || isa<CoroSuspendAsyncInst>(Val: I)) { |
172 | UseBB = UseBB->getSinglePredecessor(); |
173 | assert(UseBB && "should have split coro.suspend into its own block" ); |
174 | } |
175 | |
176 | return hasPathCrossingSuspendPoint(From: DefBB, To: UseBB); |
177 | } |
178 | |
179 | bool isDefinitionAcrossSuspend(Argument &A, User *U) const { |
180 | return isDefinitionAcrossSuspend(DefBB: &A.getParent()->getEntryBlock(), U); |
181 | } |
182 | |
183 | bool isDefinitionAcrossSuspend(Instruction &I, User *U) const { |
184 | auto *DefBB = I.getParent(); |
185 | |
186 | // As a special case, treat values produced by an llvm.coro.suspend.* |
187 | // as if they were defined in the single successor: the uses |
188 | // conceptually occur after the suspend. |
189 | if (isa<AnyCoroSuspendInst>(Val: I)) { |
190 | DefBB = DefBB->getSingleSuccessor(); |
191 | assert(DefBB && "should have split coro.suspend into its own block" ); |
192 | } |
193 | |
194 | return isDefinitionAcrossSuspend(DefBB, U); |
195 | } |
196 | |
197 | bool isDefinitionAcrossSuspend(Value &V, User *U) const { |
198 | if (auto *Arg = dyn_cast<Argument>(Val: &V)) |
199 | return isDefinitionAcrossSuspend(A&: *Arg, U); |
200 | if (auto *Inst = dyn_cast<Instruction>(Val: &V)) |
201 | return isDefinitionAcrossSuspend(I&: *Inst, U); |
202 | |
203 | llvm_unreachable( |
204 | "Coroutine could only collect Argument and Instruction now." ); |
205 | } |
206 | }; |
207 | } // end anonymous namespace |
208 | |
209 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
210 | LLVM_DUMP_METHOD void SuspendCrossingInfo::dump(StringRef Label, |
211 | BitVector const &BV) const { |
212 | dbgs() << Label << ":" ; |
213 | for (size_t I = 0, N = BV.size(); I < N; ++I) |
214 | if (BV[I]) |
215 | dbgs() << " " << Mapping.indexToBlock(I)->getName(); |
216 | dbgs() << "\n" ; |
217 | } |
218 | |
219 | LLVM_DUMP_METHOD void SuspendCrossingInfo::dump() const { |
220 | for (size_t I = 0, N = Block.size(); I < N; ++I) { |
221 | BasicBlock *const B = Mapping.indexToBlock(I); |
222 | dbgs() << B->getName() << ":\n" ; |
223 | dump(" Consumes" , Block[I].Consumes); |
224 | dump(" Kills" , Block[I].Kills); |
225 | } |
226 | dbgs() << "\n" ; |
227 | } |
228 | #endif |
229 | |
230 | template <bool Initialize> |
231 | bool SuspendCrossingInfo::computeBlockData( |
232 | const ReversePostOrderTraversal<Function *> &RPOT) { |
233 | bool Changed = false; |
234 | |
235 | for (const BasicBlock *BB : RPOT) { |
236 | auto BBNo = Mapping.blockToIndex(BB); |
237 | auto &B = Block[BBNo]; |
238 | |
239 | // We don't need to count the predecessors when initialization. |
240 | if constexpr (!Initialize) |
241 | // If all the predecessors of the current Block don't change, |
242 | // the BlockData for the current block must not change too. |
243 | if (all_of(predecessors(BD: B), [this](BasicBlock *BB) { |
244 | return !Block[Mapping.blockToIndex(BB)].Changed; |
245 | })) { |
246 | B.Changed = false; |
247 | continue; |
248 | } |
249 | |
250 | // Saved Consumes and Kills bitsets so that it is easy to see |
251 | // if anything changed after propagation. |
252 | auto SavedConsumes = B.Consumes; |
253 | auto SavedKills = B.Kills; |
254 | |
255 | for (BasicBlock *PI : predecessors(BD: B)) { |
256 | auto PrevNo = Mapping.blockToIndex(BB: PI); |
257 | auto &P = Block[PrevNo]; |
258 | |
259 | // Propagate Kills and Consumes from predecessors into B. |
260 | B.Consumes |= P.Consumes; |
261 | B.Kills |= P.Kills; |
262 | |
263 | // If block P is a suspend block, it should propagate kills into block |
264 | // B for every block P consumes. |
265 | if (P.Suspend) |
266 | B.Kills |= P.Consumes; |
267 | } |
268 | |
269 | if (B.Suspend) { |
270 | // If block B is a suspend block, it should kill all of the blocks it |
271 | // consumes. |
272 | B.Kills |= B.Consumes; |
273 | } else if (B.End) { |
274 | // If block B is an end block, it should not propagate kills as the |
275 | // blocks following coro.end() are reached during initial invocation |
276 | // of the coroutine while all the data are still available on the |
277 | // stack or in the registers. |
278 | B.Kills.reset(); |
279 | } else { |
280 | // This is reached when B block it not Suspend nor coro.end and it |
281 | // need to make sure that it is not in the kill set. |
282 | B.KillLoop |= B.Kills[BBNo]; |
283 | B.Kills.reset(Idx: BBNo); |
284 | } |
285 | |
286 | if constexpr (!Initialize) { |
287 | B.Changed = (B.Kills != SavedKills) || (B.Consumes != SavedConsumes); |
288 | Changed |= B.Changed; |
289 | } |
290 | } |
291 | |
292 | return Changed; |
293 | } |
294 | |
295 | SuspendCrossingInfo::SuspendCrossingInfo(Function &F, coro::Shape &Shape) |
296 | : Mapping(F) { |
297 | const size_t N = Mapping.size(); |
298 | Block.resize(N); |
299 | |
300 | // Initialize every block so that it consumes itself |
301 | for (size_t I = 0; I < N; ++I) { |
302 | auto &B = Block[I]; |
303 | B.Consumes.resize(N); |
304 | B.Kills.resize(N); |
305 | B.Consumes.set(I); |
306 | B.Changed = true; |
307 | } |
308 | |
309 | // Mark all CoroEnd Blocks. We do not propagate Kills beyond coro.ends as |
310 | // the code beyond coro.end is reachable during initial invocation of the |
311 | // coroutine. |
312 | for (auto *CE : Shape.CoroEnds) |
313 | getBlockData(BB: CE->getParent()).End = true; |
314 | |
315 | // Mark all suspend blocks and indicate that they kill everything they |
316 | // consume. Note, that crossing coro.save also requires a spill, as any code |
317 | // between coro.save and coro.suspend may resume the coroutine and all of the |
318 | // state needs to be saved by that time. |
319 | auto markSuspendBlock = [&](IntrinsicInst *BarrierInst) { |
320 | BasicBlock *SuspendBlock = BarrierInst->getParent(); |
321 | auto &B = getBlockData(BB: SuspendBlock); |
322 | B.Suspend = true; |
323 | B.Kills |= B.Consumes; |
324 | }; |
325 | for (auto *CSI : Shape.CoroSuspends) { |
326 | markSuspendBlock(CSI); |
327 | if (auto *Save = CSI->getCoroSave()) |
328 | markSuspendBlock(Save); |
329 | } |
330 | |
331 | // It is considered to be faster to use RPO traversal for forward-edges |
332 | // dataflow analysis. |
333 | ReversePostOrderTraversal<Function *> RPOT(&F); |
334 | computeBlockData</*Initialize=*/true>(RPOT); |
335 | while (computeBlockData</*Initialize*/ false>(RPOT)) |
336 | ; |
337 | |
338 | LLVM_DEBUG(dump()); |
339 | } |
340 | |
341 | namespace { |
342 | |
343 | // RematGraph is used to construct a DAG for rematerializable instructions |
344 | // When the constructor is invoked with a candidate instruction (which is |
345 | // materializable) it builds a DAG of materializable instructions from that |
346 | // point. |
347 | // Typically, for each instruction identified as re-materializable across a |
348 | // suspend point, a RematGraph will be created. |
349 | struct RematGraph { |
350 | // Each RematNode in the graph contains the edges to instructions providing |
351 | // operands in the current node. |
352 | struct RematNode { |
353 | Instruction *Node; |
354 | SmallVector<RematNode *> Operands; |
355 | RematNode() = default; |
356 | RematNode(Instruction *V) : Node(V) {} |
357 | }; |
358 | |
359 | RematNode *EntryNode; |
360 | using RematNodeMap = |
361 | SmallMapVector<Instruction *, std::unique_ptr<RematNode>, 8>; |
362 | RematNodeMap Remats; |
363 | const std::function<bool(Instruction &)> &MaterializableCallback; |
364 | SuspendCrossingInfo &Checker; |
365 | |
366 | RematGraph(const std::function<bool(Instruction &)> &MaterializableCallback, |
367 | Instruction *I, SuspendCrossingInfo &Checker) |
368 | : MaterializableCallback(MaterializableCallback), Checker(Checker) { |
369 | std::unique_ptr<RematNode> FirstNode = std::make_unique<RematNode>(args&: I); |
370 | EntryNode = FirstNode.get(); |
371 | std::deque<std::unique_ptr<RematNode>> WorkList; |
372 | addNode(NUPtr: std::move(FirstNode), WorkList, FirstUse: cast<User>(Val: I)); |
373 | while (WorkList.size()) { |
374 | std::unique_ptr<RematNode> N = std::move(WorkList.front()); |
375 | WorkList.pop_front(); |
376 | addNode(NUPtr: std::move(N), WorkList, FirstUse: cast<User>(Val: I)); |
377 | } |
378 | } |
379 | |
380 | void addNode(std::unique_ptr<RematNode> NUPtr, |
381 | std::deque<std::unique_ptr<RematNode>> &WorkList, |
382 | User *FirstUse) { |
383 | RematNode *N = NUPtr.get(); |
384 | if (Remats.count(Key: N->Node)) |
385 | return; |
386 | |
387 | // We haven't see this node yet - add to the list |
388 | Remats[N->Node] = std::move(NUPtr); |
389 | for (auto &Def : N->Node->operands()) { |
390 | Instruction *D = dyn_cast<Instruction>(Val: Def.get()); |
391 | if (!D || !MaterializableCallback(*D) || |
392 | !Checker.isDefinitionAcrossSuspend(I&: *D, U: FirstUse)) |
393 | continue; |
394 | |
395 | if (Remats.count(Key: D)) { |
396 | // Already have this in the graph |
397 | N->Operands.push_back(Elt: Remats[D].get()); |
398 | continue; |
399 | } |
400 | |
401 | bool NoMatch = true; |
402 | for (auto &I : WorkList) { |
403 | if (I->Node == D) { |
404 | NoMatch = false; |
405 | N->Operands.push_back(Elt: I.get()); |
406 | break; |
407 | } |
408 | } |
409 | if (NoMatch) { |
410 | // Create a new node |
411 | std::unique_ptr<RematNode> ChildNode = std::make_unique<RematNode>(args&: D); |
412 | N->Operands.push_back(Elt: ChildNode.get()); |
413 | WorkList.push_back(x: std::move(ChildNode)); |
414 | } |
415 | } |
416 | } |
417 | |
418 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
419 | void dump() const { |
420 | dbgs() << "Entry (" ; |
421 | if (EntryNode->Node->getParent()->hasName()) |
422 | dbgs() << EntryNode->Node->getParent()->getName(); |
423 | else |
424 | EntryNode->Node->getParent()->printAsOperand(dbgs(), false); |
425 | dbgs() << ") : " << *EntryNode->Node << "\n" ; |
426 | for (auto &E : Remats) { |
427 | dbgs() << *(E.first) << "\n" ; |
428 | for (RematNode *U : E.second->Operands) |
429 | dbgs() << " " << *U->Node << "\n" ; |
430 | } |
431 | } |
432 | #endif |
433 | }; |
434 | } // end anonymous namespace |
435 | |
436 | namespace llvm { |
437 | |
438 | template <> struct GraphTraits<RematGraph *> { |
439 | using NodeRef = RematGraph::RematNode *; |
440 | using ChildIteratorType = RematGraph::RematNode **; |
441 | |
442 | static NodeRef getEntryNode(RematGraph *G) { return G->EntryNode; } |
443 | static ChildIteratorType child_begin(NodeRef N) { |
444 | return N->Operands.begin(); |
445 | } |
446 | static ChildIteratorType child_end(NodeRef N) { return N->Operands.end(); } |
447 | }; |
448 | |
449 | } // end namespace llvm |
450 | |
451 | #undef DEBUG_TYPE // "coro-suspend-crossing" |
452 | #define DEBUG_TYPE "coro-frame" |
453 | |
454 | namespace { |
455 | class FrameTypeBuilder; |
456 | // Mapping from the to-be-spilled value to all the users that need reload. |
457 | using SpillInfo = SmallMapVector<Value *, SmallVector<Instruction *, 2>, 8>; |
458 | struct AllocaInfo { |
459 | AllocaInst *Alloca; |
460 | DenseMap<Instruction *, std::optional<APInt>> Aliases; |
461 | bool MayWriteBeforeCoroBegin; |
462 | AllocaInfo(AllocaInst *Alloca, |
463 | DenseMap<Instruction *, std::optional<APInt>> Aliases, |
464 | bool MayWriteBeforeCoroBegin) |
465 | : Alloca(Alloca), Aliases(std::move(Aliases)), |
466 | MayWriteBeforeCoroBegin(MayWriteBeforeCoroBegin) {} |
467 | }; |
468 | struct FrameDataInfo { |
469 | // All the values (that are not allocas) that needs to be spilled to the |
470 | // frame. |
471 | SpillInfo Spills; |
472 | // Allocas contains all values defined as allocas that need to live in the |
473 | // frame. |
474 | SmallVector<AllocaInfo, 8> Allocas; |
475 | |
476 | SmallVector<Value *, 8> getAllDefs() const { |
477 | SmallVector<Value *, 8> Defs; |
478 | for (const auto &P : Spills) |
479 | Defs.push_back(Elt: P.first); |
480 | for (const auto &A : Allocas) |
481 | Defs.push_back(Elt: A.Alloca); |
482 | return Defs; |
483 | } |
484 | |
485 | uint32_t getFieldIndex(Value *V) const { |
486 | auto Itr = FieldIndexMap.find(Val: V); |
487 | assert(Itr != FieldIndexMap.end() && |
488 | "Value does not have a frame field index" ); |
489 | return Itr->second; |
490 | } |
491 | |
492 | void setFieldIndex(Value *V, uint32_t Index) { |
493 | assert((LayoutIndexUpdateStarted || FieldIndexMap.count(V) == 0) && |
494 | "Cannot set the index for the same field twice." ); |
495 | FieldIndexMap[V] = Index; |
496 | } |
497 | |
498 | Align getAlign(Value *V) const { |
499 | auto Iter = FieldAlignMap.find(Val: V); |
500 | assert(Iter != FieldAlignMap.end()); |
501 | return Iter->second; |
502 | } |
503 | |
504 | void setAlign(Value *V, Align AL) { |
505 | assert(FieldAlignMap.count(V) == 0); |
506 | FieldAlignMap.insert(KV: {V, AL}); |
507 | } |
508 | |
509 | uint64_t getDynamicAlign(Value *V) const { |
510 | auto Iter = FieldDynamicAlignMap.find(Val: V); |
511 | assert(Iter != FieldDynamicAlignMap.end()); |
512 | return Iter->second; |
513 | } |
514 | |
515 | void setDynamicAlign(Value *V, uint64_t Align) { |
516 | assert(FieldDynamicAlignMap.count(V) == 0); |
517 | FieldDynamicAlignMap.insert(KV: {V, Align}); |
518 | } |
519 | |
520 | uint64_t getOffset(Value *V) const { |
521 | auto Iter = FieldOffsetMap.find(Val: V); |
522 | assert(Iter != FieldOffsetMap.end()); |
523 | return Iter->second; |
524 | } |
525 | |
526 | void setOffset(Value *V, uint64_t Offset) { |
527 | assert(FieldOffsetMap.count(V) == 0); |
528 | FieldOffsetMap.insert(KV: {V, Offset}); |
529 | } |
530 | |
531 | // Remap the index of every field in the frame, using the final layout index. |
532 | void updateLayoutIndex(FrameTypeBuilder &B); |
533 | |
534 | private: |
535 | // LayoutIndexUpdateStarted is used to avoid updating the index of any field |
536 | // twice by mistake. |
537 | bool LayoutIndexUpdateStarted = false; |
538 | // Map from values to their slot indexes on the frame. They will be first set |
539 | // with their original insertion field index. After the frame is built, their |
540 | // indexes will be updated into the final layout index. |
541 | DenseMap<Value *, uint32_t> FieldIndexMap; |
542 | // Map from values to their alignment on the frame. They would be set after |
543 | // the frame is built. |
544 | DenseMap<Value *, Align> FieldAlignMap; |
545 | DenseMap<Value *, uint64_t> FieldDynamicAlignMap; |
546 | // Map from values to their offset on the frame. They would be set after |
547 | // the frame is built. |
548 | DenseMap<Value *, uint64_t> FieldOffsetMap; |
549 | }; |
550 | } // namespace |
551 | |
552 | #ifndef NDEBUG |
553 | static void dumpSpills(StringRef Title, const SpillInfo &Spills) { |
554 | dbgs() << "------------- " << Title << "--------------\n" ; |
555 | for (const auto &E : Spills) { |
556 | E.first->dump(); |
557 | dbgs() << " user: " ; |
558 | for (auto *I : E.second) |
559 | I->dump(); |
560 | } |
561 | } |
562 | static void dumpRemats( |
563 | StringRef Title, |
564 | const SmallMapVector<Instruction *, std::unique_ptr<RematGraph>, 8> &RM) { |
565 | dbgs() << "------------- " << Title << "--------------\n" ; |
566 | for (const auto &E : RM) { |
567 | E.second->dump(); |
568 | dbgs() << "--\n" ; |
569 | } |
570 | } |
571 | |
572 | static void dumpAllocas(const SmallVectorImpl<AllocaInfo> &Allocas) { |
573 | dbgs() << "------------- Allocas --------------\n" ; |
574 | for (const auto &A : Allocas) { |
575 | A.Alloca->dump(); |
576 | } |
577 | } |
578 | #endif |
579 | |
580 | namespace { |
581 | using FieldIDType = size_t; |
582 | // We cannot rely solely on natural alignment of a type when building a |
583 | // coroutine frame and if the alignment specified on the Alloca instruction |
584 | // differs from the natural alignment of the alloca type we will need to insert |
585 | // padding. |
586 | class FrameTypeBuilder { |
587 | private: |
588 | struct Field { |
589 | uint64_t Size; |
590 | uint64_t Offset; |
591 | Type *Ty; |
592 | FieldIDType LayoutFieldIndex; |
593 | Align Alignment; |
594 | Align TyAlignment; |
595 | uint64_t DynamicAlignBuffer; |
596 | }; |
597 | |
598 | const DataLayout &DL; |
599 | LLVMContext &Context; |
600 | uint64_t StructSize = 0; |
601 | Align StructAlign; |
602 | bool IsFinished = false; |
603 | |
604 | std::optional<Align> MaxFrameAlignment; |
605 | |
606 | SmallVector<Field, 8> Fields; |
607 | DenseMap<Value*, unsigned> FieldIndexByKey; |
608 | |
609 | public: |
610 | FrameTypeBuilder(LLVMContext &Context, const DataLayout &DL, |
611 | std::optional<Align> MaxFrameAlignment) |
612 | : DL(DL), Context(Context), MaxFrameAlignment(MaxFrameAlignment) {} |
613 | |
614 | /// Add a field to this structure for the storage of an `alloca` |
615 | /// instruction. |
616 | [[nodiscard]] FieldIDType addFieldForAlloca(AllocaInst *AI, |
617 | bool = false) { |
618 | Type *Ty = AI->getAllocatedType(); |
619 | |
620 | // Make an array type if this is a static array allocation. |
621 | if (AI->isArrayAllocation()) { |
622 | if (auto *CI = dyn_cast<ConstantInt>(Val: AI->getArraySize())) |
623 | Ty = ArrayType::get(ElementType: Ty, NumElements: CI->getValue().getZExtValue()); |
624 | else |
625 | report_fatal_error(reason: "Coroutines cannot handle non static allocas yet" ); |
626 | } |
627 | |
628 | return addField(Ty, MaybeFieldAlignment: AI->getAlign(), IsHeader); |
629 | } |
630 | |
631 | /// We want to put the allocas whose lifetime-ranges are not overlapped |
632 | /// into one slot of coroutine frame. |
633 | /// Consider the example at:https://bugs.llvm.org/show_bug.cgi?id=45566 |
634 | /// |
635 | /// cppcoro::task<void> alternative_paths(bool cond) { |
636 | /// if (cond) { |
637 | /// big_structure a; |
638 | /// process(a); |
639 | /// co_await something(); |
640 | /// } else { |
641 | /// big_structure b; |
642 | /// process2(b); |
643 | /// co_await something(); |
644 | /// } |
645 | /// } |
646 | /// |
647 | /// We want to put variable a and variable b in the same slot to |
648 | /// reduce the size of coroutine frame. |
649 | /// |
650 | /// This function use StackLifetime algorithm to partition the AllocaInsts in |
651 | /// Spills to non-overlapped sets in order to put Alloca in the same |
652 | /// non-overlapped set into the same slot in the Coroutine Frame. Then add |
653 | /// field for the allocas in the same non-overlapped set by using the largest |
654 | /// type as the field type. |
655 | /// |
656 | /// Side Effects: Because We sort the allocas, the order of allocas in the |
657 | /// frame may be different with the order in the source code. |
658 | void addFieldForAllocas(const Function &F, FrameDataInfo &FrameData, |
659 | coro::Shape &Shape); |
660 | |
661 | /// Add a field to this structure. |
662 | [[nodiscard]] FieldIDType addField(Type *Ty, MaybeAlign MaybeFieldAlignment, |
663 | bool = false, |
664 | bool IsSpillOfValue = false) { |
665 | assert(!IsFinished && "adding fields to a finished builder" ); |
666 | assert(Ty && "must provide a type for a field" ); |
667 | |
668 | // The field size is always the alloc size of the type. |
669 | uint64_t FieldSize = DL.getTypeAllocSize(Ty); |
670 | |
671 | // For an alloca with size=0, we don't need to add a field and they |
672 | // can just point to any index in the frame. Use index 0. |
673 | if (FieldSize == 0) { |
674 | return 0; |
675 | } |
676 | |
677 | // The field alignment might not be the type alignment, but we need |
678 | // to remember the type alignment anyway to build the type. |
679 | // If we are spilling values we don't need to worry about ABI alignment |
680 | // concerns. |
681 | Align ABIAlign = DL.getABITypeAlign(Ty); |
682 | Align TyAlignment = ABIAlign; |
683 | if (IsSpillOfValue && MaxFrameAlignment && *MaxFrameAlignment < ABIAlign) |
684 | TyAlignment = *MaxFrameAlignment; |
685 | Align FieldAlignment = MaybeFieldAlignment.value_or(u&: TyAlignment); |
686 | |
687 | // The field alignment could be bigger than the max frame case, in that case |
688 | // we request additional storage to be able to dynamically align the |
689 | // pointer. |
690 | uint64_t DynamicAlignBuffer = 0; |
691 | if (MaxFrameAlignment && (FieldAlignment > *MaxFrameAlignment)) { |
692 | DynamicAlignBuffer = |
693 | offsetToAlignment(Value: MaxFrameAlignment->value(), Alignment: FieldAlignment); |
694 | FieldAlignment = *MaxFrameAlignment; |
695 | FieldSize = FieldSize + DynamicAlignBuffer; |
696 | } |
697 | |
698 | // Lay out header fields immediately. |
699 | uint64_t Offset; |
700 | if (IsHeader) { |
701 | Offset = alignTo(Size: StructSize, A: FieldAlignment); |
702 | StructSize = Offset + FieldSize; |
703 | |
704 | // Everything else has a flexible offset. |
705 | } else { |
706 | Offset = OptimizedStructLayoutField::FlexibleOffset; |
707 | } |
708 | |
709 | Fields.push_back(Elt: {.Size: FieldSize, .Offset: Offset, .Ty: Ty, .LayoutFieldIndex: 0, .Alignment: FieldAlignment, .TyAlignment: TyAlignment, |
710 | .DynamicAlignBuffer: DynamicAlignBuffer}); |
711 | return Fields.size() - 1; |
712 | } |
713 | |
714 | /// Finish the layout and set the body on the given type. |
715 | void finish(StructType *Ty); |
716 | |
717 | uint64_t getStructSize() const { |
718 | assert(IsFinished && "not yet finished!" ); |
719 | return StructSize; |
720 | } |
721 | |
722 | Align getStructAlign() const { |
723 | assert(IsFinished && "not yet finished!" ); |
724 | return StructAlign; |
725 | } |
726 | |
727 | FieldIDType getLayoutFieldIndex(FieldIDType Id) const { |
728 | assert(IsFinished && "not yet finished!" ); |
729 | return Fields[Id].LayoutFieldIndex; |
730 | } |
731 | |
732 | Field getLayoutField(FieldIDType Id) const { |
733 | assert(IsFinished && "not yet finished!" ); |
734 | return Fields[Id]; |
735 | } |
736 | }; |
737 | } // namespace |
738 | |
739 | void FrameDataInfo::updateLayoutIndex(FrameTypeBuilder &B) { |
740 | auto Updater = [&](Value *I) { |
741 | auto Field = B.getLayoutField(Id: getFieldIndex(V: I)); |
742 | setFieldIndex(V: I, Index: Field.LayoutFieldIndex); |
743 | setAlign(V: I, AL: Field.Alignment); |
744 | uint64_t dynamicAlign = |
745 | Field.DynamicAlignBuffer |
746 | ? Field.DynamicAlignBuffer + Field.Alignment.value() |
747 | : 0; |
748 | setDynamicAlign(V: I, Align: dynamicAlign); |
749 | setOffset(V: I, Offset: Field.Offset); |
750 | }; |
751 | LayoutIndexUpdateStarted = true; |
752 | for (auto &S : Spills) |
753 | Updater(S.first); |
754 | for (const auto &A : Allocas) |
755 | Updater(A.Alloca); |
756 | LayoutIndexUpdateStarted = false; |
757 | } |
758 | |
759 | void FrameTypeBuilder::addFieldForAllocas(const Function &F, |
760 | FrameDataInfo &FrameData, |
761 | coro::Shape &Shape) { |
762 | using AllocaSetType = SmallVector<AllocaInst *, 4>; |
763 | SmallVector<AllocaSetType, 4> NonOverlapedAllocas; |
764 | |
765 | // We need to add field for allocas at the end of this function. |
766 | auto AddFieldForAllocasAtExit = make_scope_exit(F: [&]() { |
767 | for (auto AllocaList : NonOverlapedAllocas) { |
768 | auto *LargestAI = *AllocaList.begin(); |
769 | FieldIDType Id = addFieldForAlloca(AI: LargestAI); |
770 | for (auto *Alloca : AllocaList) |
771 | FrameData.setFieldIndex(V: Alloca, Index: Id); |
772 | } |
773 | }); |
774 | |
775 | if (!Shape.OptimizeFrame) { |
776 | for (const auto &A : FrameData.Allocas) { |
777 | AllocaInst *Alloca = A.Alloca; |
778 | NonOverlapedAllocas.emplace_back(Args: AllocaSetType(1, Alloca)); |
779 | } |
780 | return; |
781 | } |
782 | |
783 | // Because there are paths from the lifetime.start to coro.end |
784 | // for each alloca, the liferanges for every alloca is overlaped |
785 | // in the blocks who contain coro.end and the successor blocks. |
786 | // So we choose to skip there blocks when we calculate the liferange |
787 | // for each alloca. It should be reasonable since there shouldn't be uses |
788 | // in these blocks and the coroutine frame shouldn't be used outside the |
789 | // coroutine body. |
790 | // |
791 | // Note that the user of coro.suspend may not be SwitchInst. However, this |
792 | // case seems too complex to handle. And it is harmless to skip these |
793 | // patterns since it just prevend putting the allocas to live in the same |
794 | // slot. |
795 | DenseMap<SwitchInst *, BasicBlock *> DefaultSuspendDest; |
796 | for (auto *CoroSuspendInst : Shape.CoroSuspends) { |
797 | for (auto *U : CoroSuspendInst->users()) { |
798 | if (auto *ConstSWI = dyn_cast<SwitchInst>(Val: U)) { |
799 | auto *SWI = const_cast<SwitchInst *>(ConstSWI); |
800 | DefaultSuspendDest[SWI] = SWI->getDefaultDest(); |
801 | SWI->setDefaultDest(SWI->getSuccessor(idx: 1)); |
802 | } |
803 | } |
804 | } |
805 | |
806 | auto = [&]() { |
807 | AllocaSetType Allocas; |
808 | Allocas.reserve(N: FrameData.Allocas.size()); |
809 | for (const auto &A : FrameData.Allocas) |
810 | Allocas.push_back(Elt: A.Alloca); |
811 | return Allocas; |
812 | }; |
813 | StackLifetime StackLifetimeAnalyzer(F, ExtractAllocas(), |
814 | StackLifetime::LivenessType::May); |
815 | StackLifetimeAnalyzer.run(); |
816 | auto IsAllocaInferenre = [&](const AllocaInst *AI1, const AllocaInst *AI2) { |
817 | return StackLifetimeAnalyzer.getLiveRange(AI: AI1).overlaps( |
818 | Other: StackLifetimeAnalyzer.getLiveRange(AI: AI2)); |
819 | }; |
820 | auto GetAllocaSize = [&](const AllocaInfo &A) { |
821 | std::optional<TypeSize> RetSize = A.Alloca->getAllocationSize(DL); |
822 | assert(RetSize && "Variable Length Arrays (VLA) are not supported.\n" ); |
823 | assert(!RetSize->isScalable() && "Scalable vectors are not yet supported" ); |
824 | return RetSize->getFixedValue(); |
825 | }; |
826 | // Put larger allocas in the front. So the larger allocas have higher |
827 | // priority to merge, which can save more space potentially. Also each |
828 | // AllocaSet would be ordered. So we can get the largest Alloca in one |
829 | // AllocaSet easily. |
830 | sort(C&: FrameData.Allocas, Comp: [&](const auto &Iter1, const auto &Iter2) { |
831 | return GetAllocaSize(Iter1) > GetAllocaSize(Iter2); |
832 | }); |
833 | for (const auto &A : FrameData.Allocas) { |
834 | AllocaInst *Alloca = A.Alloca; |
835 | bool Merged = false; |
836 | // Try to find if the Alloca is not inferenced with any existing |
837 | // NonOverlappedAllocaSet. If it is true, insert the alloca to that |
838 | // NonOverlappedAllocaSet. |
839 | for (auto &AllocaSet : NonOverlapedAllocas) { |
840 | assert(!AllocaSet.empty() && "Processing Alloca Set is not empty.\n" ); |
841 | bool NoInference = none_of(Range&: AllocaSet, P: [&](auto Iter) { |
842 | return IsAllocaInferenre(Alloca, Iter); |
843 | }); |
844 | // If the alignment of A is multiple of the alignment of B, the address |
845 | // of A should satisfy the requirement for aligning for B. |
846 | // |
847 | // There may be other more fine-grained strategies to handle the alignment |
848 | // infomation during the merging process. But it seems hard to handle |
849 | // these strategies and benefit little. |
850 | bool Alignable = [&]() -> bool { |
851 | auto *LargestAlloca = *AllocaSet.begin(); |
852 | return LargestAlloca->getAlign().value() % Alloca->getAlign().value() == |
853 | 0; |
854 | }(); |
855 | bool CouldMerge = NoInference && Alignable; |
856 | if (!CouldMerge) |
857 | continue; |
858 | AllocaSet.push_back(Elt: Alloca); |
859 | Merged = true; |
860 | break; |
861 | } |
862 | if (!Merged) { |
863 | NonOverlapedAllocas.emplace_back(Args: AllocaSetType(1, Alloca)); |
864 | } |
865 | } |
866 | // Recover the default target destination for each Switch statement |
867 | // reserved. |
868 | for (auto SwitchAndDefaultDest : DefaultSuspendDest) { |
869 | SwitchInst *SWI = SwitchAndDefaultDest.first; |
870 | BasicBlock *DestBB = SwitchAndDefaultDest.second; |
871 | SWI->setDefaultDest(DestBB); |
872 | } |
873 | // This Debug Info could tell us which allocas are merged into one slot. |
874 | LLVM_DEBUG(for (auto &AllocaSet |
875 | : NonOverlapedAllocas) { |
876 | if (AllocaSet.size() > 1) { |
877 | dbgs() << "In Function:" << F.getName() << "\n" ; |
878 | dbgs() << "Find Union Set " |
879 | << "\n" ; |
880 | dbgs() << "\tAllocas are \n" ; |
881 | for (auto Alloca : AllocaSet) |
882 | dbgs() << "\t\t" << *Alloca << "\n" ; |
883 | } |
884 | }); |
885 | } |
886 | |
887 | void FrameTypeBuilder::finish(StructType *Ty) { |
888 | assert(!IsFinished && "already finished!" ); |
889 | |
890 | // Prepare the optimal-layout field array. |
891 | // The Id in the layout field is a pointer to our Field for it. |
892 | SmallVector<OptimizedStructLayoutField, 8> LayoutFields; |
893 | LayoutFields.reserve(N: Fields.size()); |
894 | for (auto &Field : Fields) { |
895 | LayoutFields.emplace_back(Args: &Field, Args&: Field.Size, Args&: Field.Alignment, |
896 | Args&: Field.Offset); |
897 | } |
898 | |
899 | // Perform layout. |
900 | auto SizeAndAlign = performOptimizedStructLayout(Fields: LayoutFields); |
901 | StructSize = SizeAndAlign.first; |
902 | StructAlign = SizeAndAlign.second; |
903 | |
904 | auto getField = [](const OptimizedStructLayoutField &LayoutField) -> Field & { |
905 | return *static_cast<Field *>(const_cast<void*>(LayoutField.Id)); |
906 | }; |
907 | |
908 | // We need to produce a packed struct type if there's a field whose |
909 | // assigned offset isn't a multiple of its natural type alignment. |
910 | bool Packed = [&] { |
911 | for (auto &LayoutField : LayoutFields) { |
912 | auto &F = getField(LayoutField); |
913 | if (!isAligned(Lhs: F.TyAlignment, SizeInBytes: LayoutField.Offset)) |
914 | return true; |
915 | } |
916 | return false; |
917 | }(); |
918 | |
919 | // Build the struct body. |
920 | SmallVector<Type*, 16> FieldTypes; |
921 | FieldTypes.reserve(N: LayoutFields.size() * 3 / 2); |
922 | uint64_t LastOffset = 0; |
923 | for (auto &LayoutField : LayoutFields) { |
924 | auto &F = getField(LayoutField); |
925 | |
926 | auto Offset = LayoutField.Offset; |
927 | |
928 | // Add a padding field if there's a padding gap and we're either |
929 | // building a packed struct or the padding gap is more than we'd |
930 | // get from aligning to the field type's natural alignment. |
931 | assert(Offset >= LastOffset); |
932 | if (Offset != LastOffset) { |
933 | if (Packed || alignTo(Size: LastOffset, A: F.TyAlignment) != Offset) |
934 | FieldTypes.push_back(Elt: ArrayType::get(ElementType: Type::getInt8Ty(C&: Context), |
935 | NumElements: Offset - LastOffset)); |
936 | } |
937 | |
938 | F.Offset = Offset; |
939 | F.LayoutFieldIndex = FieldTypes.size(); |
940 | |
941 | FieldTypes.push_back(Elt: F.Ty); |
942 | if (F.DynamicAlignBuffer) { |
943 | FieldTypes.push_back( |
944 | Elt: ArrayType::get(ElementType: Type::getInt8Ty(C&: Context), NumElements: F.DynamicAlignBuffer)); |
945 | } |
946 | LastOffset = Offset + F.Size; |
947 | } |
948 | |
949 | Ty->setBody(Elements: FieldTypes, isPacked: Packed); |
950 | |
951 | #ifndef NDEBUG |
952 | // Check that the IR layout matches the offsets we expect. |
953 | auto Layout = DL.getStructLayout(Ty); |
954 | for (auto &F : Fields) { |
955 | assert(Ty->getElementType(F.LayoutFieldIndex) == F.Ty); |
956 | assert(Layout->getElementOffset(F.LayoutFieldIndex) == F.Offset); |
957 | } |
958 | #endif |
959 | |
960 | IsFinished = true; |
961 | } |
962 | |
963 | static void cacheDIVar(FrameDataInfo &FrameData, |
964 | DenseMap<Value *, DILocalVariable *> &DIVarCache) { |
965 | for (auto *V : FrameData.getAllDefs()) { |
966 | if (DIVarCache.contains(Val: V)) |
967 | continue; |
968 | |
969 | auto CacheIt = [&DIVarCache, V](const auto &Container) { |
970 | auto *I = llvm::find_if(Container, [](auto *DDI) { |
971 | return DDI->getExpression()->getNumElements() == 0; |
972 | }); |
973 | if (I != Container.end()) |
974 | DIVarCache.insert({V, (*I)->getVariable()}); |
975 | }; |
976 | CacheIt(findDbgDeclares(V)); |
977 | CacheIt(findDVRDeclares(V)); |
978 | } |
979 | } |
980 | |
981 | /// Create name for Type. It uses MDString to store new created string to |
982 | /// avoid memory leak. |
983 | static StringRef solveTypeName(Type *Ty) { |
984 | if (Ty->isIntegerTy()) { |
985 | // The longest name in common may be '__int_128', which has 9 bits. |
986 | SmallString<16> Buffer; |
987 | raw_svector_ostream OS(Buffer); |
988 | OS << "__int_" << cast<IntegerType>(Val: Ty)->getBitWidth(); |
989 | auto *MDName = MDString::get(Context&: Ty->getContext(), Str: OS.str()); |
990 | return MDName->getString(); |
991 | } |
992 | |
993 | if (Ty->isFloatingPointTy()) { |
994 | if (Ty->isFloatTy()) |
995 | return "__float_" ; |
996 | if (Ty->isDoubleTy()) |
997 | return "__double_" ; |
998 | return "__floating_type_" ; |
999 | } |
1000 | |
1001 | if (Ty->isPointerTy()) |
1002 | return "PointerType" ; |
1003 | |
1004 | if (Ty->isStructTy()) { |
1005 | if (!cast<StructType>(Val: Ty)->hasName()) |
1006 | return "__LiteralStructType_" ; |
1007 | |
1008 | auto Name = Ty->getStructName(); |
1009 | |
1010 | SmallString<16> Buffer(Name); |
1011 | for (auto &Iter : Buffer) |
1012 | if (Iter == '.' || Iter == ':') |
1013 | Iter = '_'; |
1014 | auto *MDName = MDString::get(Context&: Ty->getContext(), Str: Buffer.str()); |
1015 | return MDName->getString(); |
1016 | } |
1017 | |
1018 | return "UnknownType" ; |
1019 | } |
1020 | |
1021 | static DIType *solveDIType(DIBuilder &Builder, Type *Ty, |
1022 | const DataLayout &Layout, DIScope *Scope, |
1023 | unsigned LineNum, |
1024 | DenseMap<Type *, DIType *> &DITypeCache) { |
1025 | if (DIType *DT = DITypeCache.lookup(Val: Ty)) |
1026 | return DT; |
1027 | |
1028 | StringRef Name = solveTypeName(Ty); |
1029 | |
1030 | DIType *RetType = nullptr; |
1031 | |
1032 | if (Ty->isIntegerTy()) { |
1033 | auto BitWidth = cast<IntegerType>(Val: Ty)->getBitWidth(); |
1034 | RetType = Builder.createBasicType(Name, SizeInBits: BitWidth, Encoding: dwarf::DW_ATE_signed, |
1035 | Flags: llvm::DINode::FlagArtificial); |
1036 | } else if (Ty->isFloatingPointTy()) { |
1037 | RetType = Builder.createBasicType(Name, SizeInBits: Layout.getTypeSizeInBits(Ty), |
1038 | Encoding: dwarf::DW_ATE_float, |
1039 | Flags: llvm::DINode::FlagArtificial); |
1040 | } else if (Ty->isPointerTy()) { |
1041 | // Construct PointerType points to null (aka void *) instead of exploring |
1042 | // pointee type to avoid infinite search problem. For example, we would be |
1043 | // in trouble if we traverse recursively: |
1044 | // |
1045 | // struct Node { |
1046 | // Node* ptr; |
1047 | // }; |
1048 | RetType = |
1049 | Builder.createPointerType(PointeeTy: nullptr, SizeInBits: Layout.getTypeSizeInBits(Ty), |
1050 | AlignInBits: Layout.getABITypeAlign(Ty).value() * CHAR_BIT, |
1051 | /*DWARFAddressSpace=*/std::nullopt, Name); |
1052 | } else if (Ty->isStructTy()) { |
1053 | auto *DIStruct = Builder.createStructType( |
1054 | Scope, Name, File: Scope->getFile(), LineNumber: LineNum, SizeInBits: Layout.getTypeSizeInBits(Ty), |
1055 | AlignInBits: Layout.getPrefTypeAlign(Ty).value() * CHAR_BIT, |
1056 | Flags: llvm::DINode::FlagArtificial, DerivedFrom: nullptr, Elements: llvm::DINodeArray()); |
1057 | |
1058 | auto *StructTy = cast<StructType>(Val: Ty); |
1059 | SmallVector<Metadata *, 16> Elements; |
1060 | for (unsigned I = 0; I < StructTy->getNumElements(); I++) { |
1061 | DIType *DITy = solveDIType(Builder, Ty: StructTy->getElementType(N: I), Layout, |
1062 | Scope, LineNum, DITypeCache); |
1063 | assert(DITy); |
1064 | Elements.push_back(Elt: Builder.createMemberType( |
1065 | Scope, Name: DITy->getName(), File: Scope->getFile(), LineNo: LineNum, |
1066 | SizeInBits: DITy->getSizeInBits(), AlignInBits: DITy->getAlignInBits(), |
1067 | OffsetInBits: Layout.getStructLayout(Ty: StructTy)->getElementOffsetInBits(Idx: I), |
1068 | Flags: llvm::DINode::FlagArtificial, Ty: DITy)); |
1069 | } |
1070 | |
1071 | Builder.replaceArrays(T&: DIStruct, Elements: Builder.getOrCreateArray(Elements)); |
1072 | |
1073 | RetType = DIStruct; |
1074 | } else { |
1075 | LLVM_DEBUG(dbgs() << "Unresolved Type: " << *Ty << "\n" ); |
1076 | TypeSize Size = Layout.getTypeSizeInBits(Ty); |
1077 | auto *CharSizeType = Builder.createBasicType( |
1078 | Name, SizeInBits: 8, Encoding: dwarf::DW_ATE_unsigned_char, Flags: llvm::DINode::FlagArtificial); |
1079 | |
1080 | if (Size <= 8) |
1081 | RetType = CharSizeType; |
1082 | else { |
1083 | if (Size % 8 != 0) |
1084 | Size = TypeSize::getFixed(ExactSize: Size + 8 - (Size % 8)); |
1085 | |
1086 | RetType = Builder.createArrayType( |
1087 | Size, AlignInBits: Layout.getPrefTypeAlign(Ty).value(), Ty: CharSizeType, |
1088 | Subscripts: Builder.getOrCreateArray(Elements: Builder.getOrCreateSubrange(Lo: 0, Count: Size / 8))); |
1089 | } |
1090 | } |
1091 | |
1092 | DITypeCache.insert(KV: {Ty, RetType}); |
1093 | return RetType; |
1094 | } |
1095 | |
1096 | /// Build artificial debug info for C++ coroutine frames to allow users to |
1097 | /// inspect the contents of the frame directly |
1098 | /// |
1099 | /// Create Debug information for coroutine frame with debug name "__coro_frame". |
1100 | /// The debug information for the fields of coroutine frame is constructed from |
1101 | /// the following way: |
1102 | /// 1. For all the value in the Frame, we search the use of dbg.declare to find |
1103 | /// the corresponding debug variables for the value. If we can find the |
1104 | /// debug variable, we can get full and accurate debug information. |
1105 | /// 2. If we can't get debug information in step 1 and 2, we could only try to |
1106 | /// build the DIType by Type. We did this in solveDIType. We only handle |
1107 | /// integer, float, double, integer type and struct type for now. |
1108 | static void buildFrameDebugInfo(Function &F, coro::Shape &Shape, |
1109 | FrameDataInfo &FrameData) { |
1110 | DISubprogram *DIS = F.getSubprogram(); |
1111 | // If there is no DISubprogram for F, it implies the Function are not compiled |
1112 | // with debug info. So we also don't need to generate debug info for the frame |
1113 | // neither. |
1114 | if (!DIS || !DIS->getUnit() || |
1115 | !dwarf::isCPlusPlus( |
1116 | S: (dwarf::SourceLanguage)DIS->getUnit()->getSourceLanguage())) |
1117 | return; |
1118 | |
1119 | assert(Shape.ABI == coro::ABI::Switch && |
1120 | "We could only build debug infomation for C++ coroutine now.\n" ); |
1121 | |
1122 | DIBuilder DBuilder(*F.getParent(), /*AllowUnresolved*/ false); |
1123 | |
1124 | AllocaInst *PromiseAlloca = Shape.getPromiseAlloca(); |
1125 | assert(PromiseAlloca && |
1126 | "Coroutine with switch ABI should own Promise alloca" ); |
1127 | |
1128 | TinyPtrVector<DbgDeclareInst *> DIs = findDbgDeclares(V: PromiseAlloca); |
1129 | TinyPtrVector<DbgVariableRecord *> DVRs = findDVRDeclares(V: PromiseAlloca); |
1130 | |
1131 | DILocalVariable *PromiseDIVariable = nullptr; |
1132 | DILocation *DILoc = nullptr; |
1133 | if (!DIs.empty()) { |
1134 | DbgDeclareInst *PromiseDDI = DIs.front(); |
1135 | PromiseDIVariable = PromiseDDI->getVariable(); |
1136 | DILoc = PromiseDDI->getDebugLoc().get(); |
1137 | } else if (!DVRs.empty()) { |
1138 | DbgVariableRecord *PromiseDVR = DVRs.front(); |
1139 | PromiseDIVariable = PromiseDVR->getVariable(); |
1140 | DILoc = PromiseDVR->getDebugLoc().get(); |
1141 | } else { |
1142 | return; |
1143 | } |
1144 | |
1145 | DILocalScope *PromiseDIScope = PromiseDIVariable->getScope(); |
1146 | DIFile *DFile = PromiseDIScope->getFile(); |
1147 | unsigned LineNum = PromiseDIVariable->getLine(); |
1148 | |
1149 | DICompositeType *FrameDITy = DBuilder.createStructType( |
1150 | Scope: DIS->getUnit(), Name: Twine(F.getName() + ".coro_frame_ty" ).str(), |
1151 | File: DFile, LineNumber: LineNum, SizeInBits: Shape.FrameSize * 8, |
1152 | AlignInBits: Shape.FrameAlign.value() * 8, Flags: llvm::DINode::FlagArtificial, DerivedFrom: nullptr, |
1153 | Elements: llvm::DINodeArray()); |
1154 | StructType *FrameTy = Shape.FrameTy; |
1155 | SmallVector<Metadata *, 16> Elements; |
1156 | DataLayout Layout = F.getDataLayout(); |
1157 | |
1158 | DenseMap<Value *, DILocalVariable *> DIVarCache; |
1159 | cacheDIVar(FrameData, DIVarCache); |
1160 | |
1161 | unsigned ResumeIndex = coro::Shape::SwitchFieldIndex::Resume; |
1162 | unsigned DestroyIndex = coro::Shape::SwitchFieldIndex::Destroy; |
1163 | unsigned IndexIndex = Shape.SwitchLowering.IndexField; |
1164 | |
1165 | DenseMap<unsigned, StringRef> NameCache; |
1166 | NameCache.insert(KV: {ResumeIndex, "__resume_fn" }); |
1167 | NameCache.insert(KV: {DestroyIndex, "__destroy_fn" }); |
1168 | NameCache.insert(KV: {IndexIndex, "__coro_index" }); |
1169 | |
1170 | Type *ResumeFnTy = FrameTy->getElementType(N: ResumeIndex), |
1171 | *DestroyFnTy = FrameTy->getElementType(N: DestroyIndex), |
1172 | *IndexTy = FrameTy->getElementType(N: IndexIndex); |
1173 | |
1174 | DenseMap<unsigned, DIType *> TyCache; |
1175 | TyCache.insert( |
1176 | KV: {ResumeIndex, DBuilder.createPointerType( |
1177 | PointeeTy: nullptr, SizeInBits: Layout.getTypeSizeInBits(Ty: ResumeFnTy))}); |
1178 | TyCache.insert( |
1179 | KV: {DestroyIndex, DBuilder.createPointerType( |
1180 | PointeeTy: nullptr, SizeInBits: Layout.getTypeSizeInBits(Ty: DestroyFnTy))}); |
1181 | |
1182 | /// FIXME: If we fill the field `SizeInBits` with the actual size of |
1183 | /// __coro_index in bits, then __coro_index wouldn't show in the debugger. |
1184 | TyCache.insert(KV: {IndexIndex, DBuilder.createBasicType( |
1185 | Name: "__coro_index" , |
1186 | SizeInBits: (Layout.getTypeSizeInBits(Ty: IndexTy) < 8) |
1187 | ? 8 |
1188 | : Layout.getTypeSizeInBits(Ty: IndexTy), |
1189 | Encoding: dwarf::DW_ATE_unsigned_char)}); |
1190 | |
1191 | for (auto *V : FrameData.getAllDefs()) { |
1192 | if (!DIVarCache.contains(Val: V)) |
1193 | continue; |
1194 | |
1195 | auto Index = FrameData.getFieldIndex(V); |
1196 | |
1197 | NameCache.insert(KV: {Index, DIVarCache[V]->getName()}); |
1198 | TyCache.insert(KV: {Index, DIVarCache[V]->getType()}); |
1199 | } |
1200 | |
1201 | // Cache from index to (Align, Offset Pair) |
1202 | DenseMap<unsigned, std::pair<unsigned, unsigned>> OffsetCache; |
1203 | // The Align and Offset of Resume function and Destroy function are fixed. |
1204 | OffsetCache.insert(KV: {ResumeIndex, {8, 0}}); |
1205 | OffsetCache.insert(KV: {DestroyIndex, {8, 8}}); |
1206 | OffsetCache.insert( |
1207 | KV: {IndexIndex, |
1208 | {Shape.SwitchLowering.IndexAlign, Shape.SwitchLowering.IndexOffset}}); |
1209 | |
1210 | for (auto *V : FrameData.getAllDefs()) { |
1211 | auto Index = FrameData.getFieldIndex(V); |
1212 | |
1213 | OffsetCache.insert( |
1214 | KV: {Index, {FrameData.getAlign(V).value(), FrameData.getOffset(V)}}); |
1215 | } |
1216 | |
1217 | DenseMap<Type *, DIType *> DITypeCache; |
1218 | // This counter is used to avoid same type names. e.g., there would be |
1219 | // many i32 and i64 types in one coroutine. And we would use i32_0 and |
1220 | // i32_1 to avoid the same type. Since it makes no sense the name of the |
1221 | // fields confilicts with each other. |
1222 | unsigned UnknownTypeNum = 0; |
1223 | for (unsigned Index = 0; Index < FrameTy->getNumElements(); Index++) { |
1224 | if (!OffsetCache.contains(Val: Index)) |
1225 | continue; |
1226 | |
1227 | std::string Name; |
1228 | uint64_t SizeInBits; |
1229 | uint32_t AlignInBits; |
1230 | uint64_t OffsetInBits; |
1231 | DIType *DITy = nullptr; |
1232 | |
1233 | Type *Ty = FrameTy->getElementType(N: Index); |
1234 | assert(Ty->isSized() && "We can't handle type which is not sized.\n" ); |
1235 | SizeInBits = Layout.getTypeSizeInBits(Ty).getFixedValue(); |
1236 | AlignInBits = OffsetCache[Index].first * 8; |
1237 | OffsetInBits = OffsetCache[Index].second * 8; |
1238 | |
1239 | if (NameCache.contains(Val: Index)) { |
1240 | Name = NameCache[Index].str(); |
1241 | DITy = TyCache[Index]; |
1242 | } else { |
1243 | DITy = solveDIType(Builder&: DBuilder, Ty, Layout, Scope: FrameDITy, LineNum, DITypeCache); |
1244 | assert(DITy && "SolveDIType shouldn't return nullptr.\n" ); |
1245 | Name = DITy->getName().str(); |
1246 | Name += "_" + std::to_string(val: UnknownTypeNum); |
1247 | UnknownTypeNum++; |
1248 | } |
1249 | |
1250 | Elements.push_back(Elt: DBuilder.createMemberType( |
1251 | Scope: FrameDITy, Name, File: DFile, LineNo: LineNum, SizeInBits, AlignInBits, OffsetInBits, |
1252 | Flags: llvm::DINode::FlagArtificial, Ty: DITy)); |
1253 | } |
1254 | |
1255 | DBuilder.replaceArrays(T&: FrameDITy, Elements: DBuilder.getOrCreateArray(Elements)); |
1256 | |
1257 | auto *FrameDIVar = DBuilder.createAutoVariable(Scope: PromiseDIScope, Name: "__coro_frame" , |
1258 | File: DFile, LineNo: LineNum, Ty: FrameDITy, |
1259 | AlwaysPreserve: true, Flags: DINode::FlagArtificial); |
1260 | assert(FrameDIVar->isValidLocationForIntrinsic(DILoc)); |
1261 | |
1262 | // Subprogram would have ContainedNodes field which records the debug |
1263 | // variables it contained. So we need to add __coro_frame to the |
1264 | // ContainedNodes of it. |
1265 | // |
1266 | // If we don't add __coro_frame to the RetainedNodes, user may get |
1267 | // `no symbol __coro_frame in context` rather than `__coro_frame` |
1268 | // is optimized out, which is more precise. |
1269 | if (auto *SubProgram = dyn_cast<DISubprogram>(Val: PromiseDIScope)) { |
1270 | auto RetainedNodes = SubProgram->getRetainedNodes(); |
1271 | SmallVector<Metadata *, 32> RetainedNodesVec(RetainedNodes.begin(), |
1272 | RetainedNodes.end()); |
1273 | RetainedNodesVec.push_back(Elt: FrameDIVar); |
1274 | SubProgram->replaceOperandWith( |
1275 | I: 7, New: (MDTuple::get(Context&: F.getContext(), MDs: RetainedNodesVec))); |
1276 | } |
1277 | |
1278 | if (UseNewDbgInfoFormat) { |
1279 | DbgVariableRecord *NewDVR = |
1280 | new DbgVariableRecord(ValueAsMetadata::get(V: Shape.FramePtr), FrameDIVar, |
1281 | DBuilder.createExpression(), DILoc, |
1282 | DbgVariableRecord::LocationType::Declare); |
1283 | BasicBlock::iterator It = Shape.getInsertPtAfterFramePtr(); |
1284 | It->getParent()->insertDbgRecordBefore(DR: NewDVR, Here: It); |
1285 | } else { |
1286 | DBuilder.insertDeclare(Storage: Shape.FramePtr, VarInfo: FrameDIVar, |
1287 | Expr: DBuilder.createExpression(), DL: DILoc, |
1288 | InsertBefore: &*Shape.getInsertPtAfterFramePtr()); |
1289 | } |
1290 | } |
1291 | |
1292 | // Build a struct that will keep state for an active coroutine. |
1293 | // struct f.frame { |
1294 | // ResumeFnTy ResumeFnAddr; |
1295 | // ResumeFnTy DestroyFnAddr; |
1296 | // ... promise (if present) ... |
1297 | // int ResumeIndex; |
1298 | // ... spills ... |
1299 | // }; |
1300 | static StructType *buildFrameType(Function &F, coro::Shape &Shape, |
1301 | FrameDataInfo &FrameData) { |
1302 | LLVMContext &C = F.getContext(); |
1303 | const DataLayout &DL = F.getDataLayout(); |
1304 | StructType *FrameTy = [&] { |
1305 | SmallString<32> Name(F.getName()); |
1306 | Name.append(RHS: ".Frame" ); |
1307 | return StructType::create(Context&: C, Name); |
1308 | }(); |
1309 | |
1310 | // We will use this value to cap the alignment of spilled values. |
1311 | std::optional<Align> MaxFrameAlignment; |
1312 | if (Shape.ABI == coro::ABI::Async) |
1313 | MaxFrameAlignment = Shape.AsyncLowering.getContextAlignment(); |
1314 | FrameTypeBuilder B(C, DL, MaxFrameAlignment); |
1315 | |
1316 | AllocaInst *PromiseAlloca = Shape.getPromiseAlloca(); |
1317 | std::optional<FieldIDType> SwitchIndexFieldId; |
1318 | |
1319 | if (Shape.ABI == coro::ABI::Switch) { |
1320 | auto *FnPtrTy = PointerType::getUnqual(C); |
1321 | |
1322 | // Add header fields for the resume and destroy functions. |
1323 | // We can rely on these being perfectly packed. |
1324 | (void)B.addField(Ty: FnPtrTy, MaybeFieldAlignment: std::nullopt, /*header*/ IsHeader: true); |
1325 | (void)B.addField(Ty: FnPtrTy, MaybeFieldAlignment: std::nullopt, /*header*/ IsHeader: true); |
1326 | |
1327 | // PromiseAlloca field needs to be explicitly added here because it's |
1328 | // a header field with a fixed offset based on its alignment. Hence it |
1329 | // needs special handling and cannot be added to FrameData.Allocas. |
1330 | if (PromiseAlloca) |
1331 | FrameData.setFieldIndex( |
1332 | V: PromiseAlloca, Index: B.addFieldForAlloca(AI: PromiseAlloca, /*header*/ IsHeader: true)); |
1333 | |
1334 | // Add a field to store the suspend index. This doesn't need to |
1335 | // be in the header. |
1336 | unsigned IndexBits = std::max(a: 1U, b: Log2_64_Ceil(Value: Shape.CoroSuspends.size())); |
1337 | Type *IndexType = Type::getIntNTy(C, N: IndexBits); |
1338 | |
1339 | SwitchIndexFieldId = B.addField(Ty: IndexType, MaybeFieldAlignment: std::nullopt); |
1340 | } else { |
1341 | assert(PromiseAlloca == nullptr && "lowering doesn't support promises" ); |
1342 | } |
1343 | |
1344 | // Because multiple allocas may own the same field slot, |
1345 | // we add allocas to field here. |
1346 | B.addFieldForAllocas(F, FrameData, Shape); |
1347 | // Add PromiseAlloca to Allocas list so that |
1348 | // 1. updateLayoutIndex could update its index after |
1349 | // `performOptimizedStructLayout` |
1350 | // 2. it is processed in insertSpills. |
1351 | if (Shape.ABI == coro::ABI::Switch && PromiseAlloca) |
1352 | // We assume that the promise alloca won't be modified before |
1353 | // CoroBegin and no alias will be create before CoroBegin. |
1354 | FrameData.Allocas.emplace_back( |
1355 | Args&: PromiseAlloca, Args: DenseMap<Instruction *, std::optional<APInt>>{}, Args: false); |
1356 | // Create an entry for every spilled value. |
1357 | for (auto &S : FrameData.Spills) { |
1358 | Type *FieldType = S.first->getType(); |
1359 | // For byval arguments, we need to store the pointed value in the frame, |
1360 | // instead of the pointer itself. |
1361 | if (const Argument *A = dyn_cast<Argument>(Val: S.first)) |
1362 | if (A->hasByValAttr()) |
1363 | FieldType = A->getParamByValType(); |
1364 | FieldIDType Id = B.addField(Ty: FieldType, MaybeFieldAlignment: std::nullopt, IsHeader: false /*header*/, |
1365 | IsSpillOfValue: true /*IsSpillOfValue*/); |
1366 | FrameData.setFieldIndex(V: S.first, Index: Id); |
1367 | } |
1368 | |
1369 | B.finish(Ty: FrameTy); |
1370 | FrameData.updateLayoutIndex(B); |
1371 | Shape.FrameAlign = B.getStructAlign(); |
1372 | Shape.FrameSize = B.getStructSize(); |
1373 | |
1374 | switch (Shape.ABI) { |
1375 | case coro::ABI::Switch: { |
1376 | // In the switch ABI, remember the switch-index field. |
1377 | auto IndexField = B.getLayoutField(Id: *SwitchIndexFieldId); |
1378 | Shape.SwitchLowering.IndexField = IndexField.LayoutFieldIndex; |
1379 | Shape.SwitchLowering.IndexAlign = IndexField.Alignment.value(); |
1380 | Shape.SwitchLowering.IndexOffset = IndexField.Offset; |
1381 | |
1382 | // Also round the frame size up to a multiple of its alignment, as is |
1383 | // generally expected in C/C++. |
1384 | Shape.FrameSize = alignTo(Size: Shape.FrameSize, A: Shape.FrameAlign); |
1385 | break; |
1386 | } |
1387 | |
1388 | // In the retcon ABI, remember whether the frame is inline in the storage. |
1389 | case coro::ABI::Retcon: |
1390 | case coro::ABI::RetconOnce: { |
1391 | auto Id = Shape.getRetconCoroId(); |
1392 | Shape.RetconLowering.IsFrameInlineInStorage |
1393 | = (B.getStructSize() <= Id->getStorageSize() && |
1394 | B.getStructAlign() <= Id->getStorageAlignment()); |
1395 | break; |
1396 | } |
1397 | case coro::ABI::Async: { |
1398 | Shape.AsyncLowering.FrameOffset = |
1399 | alignTo(Size: Shape.AsyncLowering.ContextHeaderSize, A: Shape.FrameAlign); |
1400 | // Also make the final context size a multiple of the context alignment to |
1401 | // make allocation easier for allocators. |
1402 | Shape.AsyncLowering.ContextSize = |
1403 | alignTo(Size: Shape.AsyncLowering.FrameOffset + Shape.FrameSize, |
1404 | A: Shape.AsyncLowering.getContextAlignment()); |
1405 | if (Shape.AsyncLowering.getContextAlignment() < Shape.FrameAlign) { |
1406 | report_fatal_error( |
1407 | reason: "The alignment requirment of frame variables cannot be higher than " |
1408 | "the alignment of the async function context" ); |
1409 | } |
1410 | break; |
1411 | } |
1412 | } |
1413 | |
1414 | return FrameTy; |
1415 | } |
1416 | |
1417 | // We use a pointer use visitor to track how an alloca is being used. |
1418 | // The goal is to be able to answer the following three questions: |
1419 | // 1. Should this alloca be allocated on the frame instead. |
1420 | // 2. Could the content of the alloca be modified prior to CoroBegn, which would |
1421 | // require copying the data from alloca to the frame after CoroBegin. |
1422 | // 3. Is there any alias created for this alloca prior to CoroBegin, but used |
1423 | // after CoroBegin. In that case, we will need to recreate the alias after |
1424 | // CoroBegin based off the frame. To answer question 1, we track two things: |
1425 | // a. List of all BasicBlocks that use this alloca or any of the aliases of |
1426 | // the alloca. In the end, we check if there exists any two basic blocks that |
1427 | // cross suspension points. If so, this alloca must be put on the frame. b. |
1428 | // Whether the alloca or any alias of the alloca is escaped at some point, |
1429 | // either by storing the address somewhere, or the address is used in a |
1430 | // function call that might capture. If it's ever escaped, this alloca must be |
1431 | // put on the frame conservatively. |
1432 | // To answer quetion 2, we track through the variable MayWriteBeforeCoroBegin. |
1433 | // Whenever a potential write happens, either through a store instruction, a |
1434 | // function call or any of the memory intrinsics, we check whether this |
1435 | // instruction is prior to CoroBegin. To answer question 3, we track the offsets |
1436 | // of all aliases created for the alloca prior to CoroBegin but used after |
1437 | // CoroBegin. std::optional is used to be able to represent the case when the |
1438 | // offset is unknown (e.g. when you have a PHINode that takes in different |
1439 | // offset values). We cannot handle unknown offsets and will assert. This is the |
1440 | // potential issue left out. An ideal solution would likely require a |
1441 | // significant redesign. |
1442 | namespace { |
1443 | struct AllocaUseVisitor : PtrUseVisitor<AllocaUseVisitor> { |
1444 | using Base = PtrUseVisitor<AllocaUseVisitor>; |
1445 | AllocaUseVisitor(const DataLayout &DL, const DominatorTree &DT, |
1446 | const coro::Shape &CoroShape, |
1447 | const SuspendCrossingInfo &Checker, |
1448 | bool ShouldUseLifetimeStartInfo) |
1449 | : PtrUseVisitor(DL), DT(DT), CoroShape(CoroShape), Checker(Checker), |
1450 | ShouldUseLifetimeStartInfo(ShouldUseLifetimeStartInfo) { |
1451 | for (AnyCoroSuspendInst *SuspendInst : CoroShape.CoroSuspends) |
1452 | CoroSuspendBBs.insert(Ptr: SuspendInst->getParent()); |
1453 | } |
1454 | |
1455 | void visit(Instruction &I) { |
1456 | Users.insert(Ptr: &I); |
1457 | Base::visit(I); |
1458 | // If the pointer is escaped prior to CoroBegin, we have to assume it would |
1459 | // be written into before CoroBegin as well. |
1460 | if (PI.isEscaped() && |
1461 | !DT.dominates(Def: CoroShape.CoroBegin, User: PI.getEscapingInst())) { |
1462 | MayWriteBeforeCoroBegin = true; |
1463 | } |
1464 | } |
1465 | // We need to provide this overload as PtrUseVisitor uses a pointer based |
1466 | // visiting function. |
1467 | void visit(Instruction *I) { return visit(I&: *I); } |
1468 | |
1469 | void visitPHINode(PHINode &I) { |
1470 | enqueueUsers(I); |
1471 | handleAlias(I); |
1472 | } |
1473 | |
1474 | void visitSelectInst(SelectInst &I) { |
1475 | enqueueUsers(I); |
1476 | handleAlias(I); |
1477 | } |
1478 | |
1479 | void visitStoreInst(StoreInst &SI) { |
1480 | // Regardless whether the alias of the alloca is the value operand or the |
1481 | // pointer operand, we need to assume the alloca is been written. |
1482 | handleMayWrite(I: SI); |
1483 | |
1484 | if (SI.getValueOperand() != U->get()) |
1485 | return; |
1486 | |
1487 | // We are storing the pointer into a memory location, potentially escaping. |
1488 | // As an optimization, we try to detect simple cases where it doesn't |
1489 | // actually escape, for example: |
1490 | // %ptr = alloca .. |
1491 | // %addr = alloca .. |
1492 | // store %ptr, %addr |
1493 | // %x = load %addr |
1494 | // .. |
1495 | // If %addr is only used by loading from it, we could simply treat %x as |
1496 | // another alias of %ptr, and not considering %ptr being escaped. |
1497 | auto IsSimpleStoreThenLoad = [&]() { |
1498 | auto *AI = dyn_cast<AllocaInst>(Val: SI.getPointerOperand()); |
1499 | // If the memory location we are storing to is not an alloca, it |
1500 | // could be an alias of some other memory locations, which is difficult |
1501 | // to analyze. |
1502 | if (!AI) |
1503 | return false; |
1504 | // StoreAliases contains aliases of the memory location stored into. |
1505 | SmallVector<Instruction *, 4> StoreAliases = {AI}; |
1506 | while (!StoreAliases.empty()) { |
1507 | Instruction *I = StoreAliases.pop_back_val(); |
1508 | for (User *U : I->users()) { |
1509 | // If we are loading from the memory location, we are creating an |
1510 | // alias of the original pointer. |
1511 | if (auto *LI = dyn_cast<LoadInst>(Val: U)) { |
1512 | enqueueUsers(I&: *LI); |
1513 | handleAlias(I&: *LI); |
1514 | continue; |
1515 | } |
1516 | // If we are overriding the memory location, the pointer certainly |
1517 | // won't escape. |
1518 | if (auto *S = dyn_cast<StoreInst>(Val: U)) |
1519 | if (S->getPointerOperand() == I) |
1520 | continue; |
1521 | if (auto *II = dyn_cast<IntrinsicInst>(Val: U)) |
1522 | if (II->isLifetimeStartOrEnd()) |
1523 | continue; |
1524 | // BitCastInst creats aliases of the memory location being stored |
1525 | // into. |
1526 | if (auto *BI = dyn_cast<BitCastInst>(Val: U)) { |
1527 | StoreAliases.push_back(Elt: BI); |
1528 | continue; |
1529 | } |
1530 | return false; |
1531 | } |
1532 | } |
1533 | |
1534 | return true; |
1535 | }; |
1536 | |
1537 | if (!IsSimpleStoreThenLoad()) |
1538 | PI.setEscaped(&SI); |
1539 | } |
1540 | |
1541 | // All mem intrinsics modify the data. |
1542 | void visitMemIntrinsic(MemIntrinsic &MI) { handleMayWrite(I: MI); } |
1543 | |
1544 | void visitBitCastInst(BitCastInst &BC) { |
1545 | Base::visitBitCastInst(BC); |
1546 | handleAlias(I&: BC); |
1547 | } |
1548 | |
1549 | void visitAddrSpaceCastInst(AddrSpaceCastInst &ASC) { |
1550 | Base::visitAddrSpaceCastInst(ASC); |
1551 | handleAlias(I&: ASC); |
1552 | } |
1553 | |
1554 | void visitGetElementPtrInst(GetElementPtrInst &GEPI) { |
1555 | // The base visitor will adjust Offset accordingly. |
1556 | Base::visitGetElementPtrInst(GEPI); |
1557 | handleAlias(I&: GEPI); |
1558 | } |
1559 | |
1560 | void visitIntrinsicInst(IntrinsicInst &II) { |
1561 | // When we found the lifetime markers refers to a |
1562 | // subrange of the original alloca, ignore the lifetime |
1563 | // markers to avoid misleading the analysis. |
1564 | if (!IsOffsetKnown || !Offset.isZero()) |
1565 | return Base::visitIntrinsicInst(II); |
1566 | switch (II.getIntrinsicID()) { |
1567 | default: |
1568 | return Base::visitIntrinsicInst(II); |
1569 | case Intrinsic::lifetime_start: |
1570 | LifetimeStarts.insert(Ptr: &II); |
1571 | LifetimeStartBBs.push_back(Elt: II.getParent()); |
1572 | break; |
1573 | case Intrinsic::lifetime_end: |
1574 | LifetimeEndBBs.insert(Ptr: II.getParent()); |
1575 | break; |
1576 | } |
1577 | } |
1578 | |
1579 | void visitCallBase(CallBase &CB) { |
1580 | for (unsigned Op = 0, OpCount = CB.arg_size(); Op < OpCount; ++Op) |
1581 | if (U->get() == CB.getArgOperand(i: Op) && !CB.doesNotCapture(OpNo: Op)) |
1582 | PI.setEscaped(&CB); |
1583 | handleMayWrite(I: CB); |
1584 | } |
1585 | |
1586 | bool getShouldLiveOnFrame() const { |
1587 | if (!ShouldLiveOnFrame) |
1588 | ShouldLiveOnFrame = computeShouldLiveOnFrame(); |
1589 | return *ShouldLiveOnFrame; |
1590 | } |
1591 | |
1592 | bool getMayWriteBeforeCoroBegin() const { return MayWriteBeforeCoroBegin; } |
1593 | |
1594 | DenseMap<Instruction *, std::optional<APInt>> getAliasesCopy() const { |
1595 | assert(getShouldLiveOnFrame() && "This method should only be called if the " |
1596 | "alloca needs to live on the frame." ); |
1597 | for (const auto &P : AliasOffetMap) |
1598 | if (!P.second) |
1599 | report_fatal_error(reason: "Unable to handle an alias with unknown offset " |
1600 | "created before CoroBegin." ); |
1601 | return AliasOffetMap; |
1602 | } |
1603 | |
1604 | private: |
1605 | const DominatorTree &DT; |
1606 | const coro::Shape &CoroShape; |
1607 | const SuspendCrossingInfo &Checker; |
1608 | // All alias to the original AllocaInst, created before CoroBegin and used |
1609 | // after CoroBegin. Each entry contains the instruction and the offset in the |
1610 | // original Alloca. They need to be recreated after CoroBegin off the frame. |
1611 | DenseMap<Instruction *, std::optional<APInt>> AliasOffetMap{}; |
1612 | SmallPtrSet<Instruction *, 4> Users{}; |
1613 | SmallPtrSet<IntrinsicInst *, 2> LifetimeStarts{}; |
1614 | SmallVector<BasicBlock *> LifetimeStartBBs{}; |
1615 | SmallPtrSet<BasicBlock *, 2> LifetimeEndBBs{}; |
1616 | SmallPtrSet<const BasicBlock *, 2> CoroSuspendBBs{}; |
1617 | bool MayWriteBeforeCoroBegin{false}; |
1618 | bool ShouldUseLifetimeStartInfo{true}; |
1619 | |
1620 | mutable std::optional<bool> ShouldLiveOnFrame{}; |
1621 | |
1622 | bool computeShouldLiveOnFrame() const { |
1623 | // If lifetime information is available, we check it first since it's |
1624 | // more precise. We look at every pair of lifetime.start intrinsic and |
1625 | // every basic block that uses the pointer to see if they cross suspension |
1626 | // points. The uses cover both direct uses as well as indirect uses. |
1627 | if (ShouldUseLifetimeStartInfo && !LifetimeStarts.empty()) { |
1628 | // If there is no explicit lifetime.end, then assume the address can |
1629 | // cross suspension points. |
1630 | if (LifetimeEndBBs.empty()) |
1631 | return true; |
1632 | |
1633 | // If there is a path from a lifetime.start to a suspend without a |
1634 | // corresponding lifetime.end, then the alloca's lifetime persists |
1635 | // beyond that suspension point and the alloca must go on the frame. |
1636 | llvm::SmallVector<BasicBlock *> Worklist(LifetimeStartBBs); |
1637 | if (isManyPotentiallyReachableFromMany(Worklist, StopSet: CoroSuspendBBs, |
1638 | ExclusionSet: &LifetimeEndBBs, DT: &DT)) |
1639 | return true; |
1640 | |
1641 | // Addresses are guaranteed to be identical after every lifetime.start so |
1642 | // we cannot use the local stack if the address escaped and there is a |
1643 | // suspend point between lifetime markers. This should also cover the |
1644 | // case of a single lifetime.start intrinsic in a loop with suspend point. |
1645 | if (PI.isEscaped()) { |
1646 | for (auto *A : LifetimeStarts) { |
1647 | for (auto *B : LifetimeStarts) { |
1648 | if (Checker.hasPathOrLoopCrossingSuspendPoint(From: A->getParent(), |
1649 | To: B->getParent())) |
1650 | return true; |
1651 | } |
1652 | } |
1653 | } |
1654 | return false; |
1655 | } |
1656 | // FIXME: Ideally the isEscaped check should come at the beginning. |
1657 | // However there are a few loose ends that need to be fixed first before |
1658 | // we can do that. We need to make sure we are not over-conservative, so |
1659 | // that the data accessed in-between await_suspend and symmetric transfer |
1660 | // is always put on the stack, and also data accessed after coro.end is |
1661 | // always put on the stack (esp the return object). To fix that, we need |
1662 | // to: |
1663 | // 1) Potentially treat sret as nocapture in calls |
1664 | // 2) Special handle the return object and put it on the stack |
1665 | // 3) Utilize lifetime.end intrinsic |
1666 | if (PI.isEscaped()) |
1667 | return true; |
1668 | |
1669 | for (auto *U1 : Users) |
1670 | for (auto *U2 : Users) |
1671 | if (Checker.isDefinitionAcrossSuspend(I&: *U1, U: U2)) |
1672 | return true; |
1673 | |
1674 | return false; |
1675 | } |
1676 | |
1677 | void handleMayWrite(const Instruction &I) { |
1678 | if (!DT.dominates(Def: CoroShape.CoroBegin, User: &I)) |
1679 | MayWriteBeforeCoroBegin = true; |
1680 | } |
1681 | |
1682 | bool usedAfterCoroBegin(Instruction &I) { |
1683 | for (auto &U : I.uses()) |
1684 | if (DT.dominates(Def: CoroShape.CoroBegin, U)) |
1685 | return true; |
1686 | return false; |
1687 | } |
1688 | |
1689 | void handleAlias(Instruction &I) { |
1690 | // We track all aliases created prior to CoroBegin but used after. |
1691 | // These aliases may need to be recreated after CoroBegin if the alloca |
1692 | // need to live on the frame. |
1693 | if (DT.dominates(Def: CoroShape.CoroBegin, User: &I) || !usedAfterCoroBegin(I)) |
1694 | return; |
1695 | |
1696 | if (!IsOffsetKnown) { |
1697 | AliasOffetMap[&I].reset(); |
1698 | } else { |
1699 | auto Itr = AliasOffetMap.find(Val: &I); |
1700 | if (Itr == AliasOffetMap.end()) { |
1701 | AliasOffetMap[&I] = Offset; |
1702 | } else if (Itr->second && *Itr->second != Offset) { |
1703 | // If we have seen two different possible values for this alias, we set |
1704 | // it to empty. |
1705 | AliasOffetMap[&I].reset(); |
1706 | } |
1707 | } |
1708 | } |
1709 | }; |
1710 | } // namespace |
1711 | |
1712 | // We need to make room to insert a spill after initial PHIs, but before |
1713 | // catchswitch instruction. Placing it before violates the requirement that |
1714 | // catchswitch, like all other EHPads must be the first nonPHI in a block. |
1715 | // |
1716 | // Split away catchswitch into a separate block and insert in its place: |
1717 | // |
1718 | // cleanuppad <InsertPt> cleanupret. |
1719 | // |
1720 | // cleanupret instruction will act as an insert point for the spill. |
1721 | static Instruction *splitBeforeCatchSwitch(CatchSwitchInst *CatchSwitch) { |
1722 | BasicBlock *CurrentBlock = CatchSwitch->getParent(); |
1723 | BasicBlock *NewBlock = CurrentBlock->splitBasicBlock(I: CatchSwitch); |
1724 | CurrentBlock->getTerminator()->eraseFromParent(); |
1725 | |
1726 | auto *CleanupPad = |
1727 | CleanupPadInst::Create(ParentPad: CatchSwitch->getParentPad(), Args: {}, NameStr: "" , InsertBefore: CurrentBlock); |
1728 | auto *CleanupRet = |
1729 | CleanupReturnInst::Create(CleanupPad, UnwindBB: NewBlock, InsertBefore: CurrentBlock); |
1730 | return CleanupRet; |
1731 | } |
1732 | |
1733 | // Replace all alloca and SSA values that are accessed across suspend points |
1734 | // with GetElementPointer from coroutine frame + loads and stores. Create an |
1735 | // AllocaSpillBB that will become the new entry block for the resume parts of |
1736 | // the coroutine: |
1737 | // |
1738 | // %hdl = coro.begin(...) |
1739 | // whatever |
1740 | // |
1741 | // becomes: |
1742 | // |
1743 | // %hdl = coro.begin(...) |
1744 | // br label %AllocaSpillBB |
1745 | // |
1746 | // AllocaSpillBB: |
1747 | // ; geps corresponding to allocas that were moved to coroutine frame |
1748 | // br label PostSpill |
1749 | // |
1750 | // PostSpill: |
1751 | // whatever |
1752 | // |
1753 | // |
1754 | static void insertSpills(const FrameDataInfo &FrameData, coro::Shape &Shape) { |
1755 | auto *CB = Shape.CoroBegin; |
1756 | LLVMContext &C = CB->getContext(); |
1757 | Function *F = CB->getFunction(); |
1758 | IRBuilder<> Builder(C); |
1759 | StructType *FrameTy = Shape.FrameTy; |
1760 | Value *FramePtr = Shape.FramePtr; |
1761 | DominatorTree DT(*F); |
1762 | SmallDenseMap<Argument *, AllocaInst *, 4> ArgToAllocaMap; |
1763 | |
1764 | // Create a GEP with the given index into the coroutine frame for the original |
1765 | // value Orig. Appends an extra 0 index for array-allocas, preserving the |
1766 | // original type. |
1767 | auto GetFramePointer = [&](Value *Orig) -> Value * { |
1768 | FieldIDType Index = FrameData.getFieldIndex(V: Orig); |
1769 | SmallVector<Value *, 3> Indices = { |
1770 | ConstantInt::get(Ty: Type::getInt32Ty(C), V: 0), |
1771 | ConstantInt::get(Ty: Type::getInt32Ty(C), V: Index), |
1772 | }; |
1773 | |
1774 | if (auto *AI = dyn_cast<AllocaInst>(Val: Orig)) { |
1775 | if (auto *CI = dyn_cast<ConstantInt>(Val: AI->getArraySize())) { |
1776 | auto Count = CI->getValue().getZExtValue(); |
1777 | if (Count > 1) { |
1778 | Indices.push_back(Elt: ConstantInt::get(Ty: Type::getInt32Ty(C), V: 0)); |
1779 | } |
1780 | } else { |
1781 | report_fatal_error(reason: "Coroutines cannot handle non static allocas yet" ); |
1782 | } |
1783 | } |
1784 | |
1785 | auto GEP = cast<GetElementPtrInst>( |
1786 | Val: Builder.CreateInBoundsGEP(Ty: FrameTy, Ptr: FramePtr, IdxList: Indices)); |
1787 | if (auto *AI = dyn_cast<AllocaInst>(Val: Orig)) { |
1788 | if (FrameData.getDynamicAlign(V: Orig) != 0) { |
1789 | assert(FrameData.getDynamicAlign(Orig) == AI->getAlign().value()); |
1790 | auto *M = AI->getModule(); |
1791 | auto *IntPtrTy = M->getDataLayout().getIntPtrType(AI->getType()); |
1792 | auto *PtrValue = Builder.CreatePtrToInt(V: GEP, DestTy: IntPtrTy); |
1793 | auto *AlignMask = |
1794 | ConstantInt::get(Ty: IntPtrTy, V: AI->getAlign().value() - 1); |
1795 | PtrValue = Builder.CreateAdd(LHS: PtrValue, RHS: AlignMask); |
1796 | PtrValue = Builder.CreateAnd(LHS: PtrValue, RHS: Builder.CreateNot(V: AlignMask)); |
1797 | return Builder.CreateIntToPtr(V: PtrValue, DestTy: AI->getType()); |
1798 | } |
1799 | // If the type of GEP is not equal to the type of AllocaInst, it implies |
1800 | // that the AllocaInst may be reused in the Frame slot of other |
1801 | // AllocaInst. So We cast GEP to the AllocaInst here to re-use |
1802 | // the Frame storage. |
1803 | // |
1804 | // Note: If we change the strategy dealing with alignment, we need to refine |
1805 | // this casting. |
1806 | if (GEP->getType() != Orig->getType()) |
1807 | return Builder.CreateAddrSpaceCast(V: GEP, DestTy: Orig->getType(), |
1808 | Name: Orig->getName() + Twine(".cast" )); |
1809 | } |
1810 | return GEP; |
1811 | }; |
1812 | |
1813 | for (auto const &E : FrameData.Spills) { |
1814 | Value *Def = E.first; |
1815 | auto SpillAlignment = Align(FrameData.getAlign(V: Def)); |
1816 | // Create a store instruction storing the value into the |
1817 | // coroutine frame. |
1818 | BasicBlock::iterator InsertPt; |
1819 | Type *ByValTy = nullptr; |
1820 | if (auto *Arg = dyn_cast<Argument>(Val: Def)) { |
1821 | // For arguments, we will place the store instruction right after |
1822 | // the coroutine frame pointer instruction, i.e. coro.begin. |
1823 | InsertPt = Shape.getInsertPtAfterFramePtr(); |
1824 | |
1825 | // If we're spilling an Argument, make sure we clear 'nocapture' |
1826 | // from the coroutine function. |
1827 | Arg->getParent()->removeParamAttr(ArgNo: Arg->getArgNo(), Kind: Attribute::NoCapture); |
1828 | |
1829 | if (Arg->hasByValAttr()) |
1830 | ByValTy = Arg->getParamByValType(); |
1831 | } else if (auto *CSI = dyn_cast<AnyCoroSuspendInst>(Val: Def)) { |
1832 | // Don't spill immediately after a suspend; splitting assumes |
1833 | // that the suspend will be followed by a branch. |
1834 | InsertPt = CSI->getParent()->getSingleSuccessor()->getFirstNonPHIIt(); |
1835 | } else { |
1836 | auto *I = cast<Instruction>(Val: Def); |
1837 | if (!DT.dominates(Def: CB, User: I)) { |
1838 | // If it is not dominated by CoroBegin, then spill should be |
1839 | // inserted immediately after CoroFrame is computed. |
1840 | InsertPt = Shape.getInsertPtAfterFramePtr(); |
1841 | } else if (auto *II = dyn_cast<InvokeInst>(Val: I)) { |
1842 | // If we are spilling the result of the invoke instruction, split |
1843 | // the normal edge and insert the spill in the new block. |
1844 | auto *NewBB = SplitEdge(From: II->getParent(), To: II->getNormalDest()); |
1845 | InsertPt = NewBB->getTerminator()->getIterator(); |
1846 | } else if (isa<PHINode>(Val: I)) { |
1847 | // Skip the PHINodes and EH pads instructions. |
1848 | BasicBlock *DefBlock = I->getParent(); |
1849 | if (auto *CSI = dyn_cast<CatchSwitchInst>(Val: DefBlock->getTerminator())) |
1850 | InsertPt = splitBeforeCatchSwitch(CatchSwitch: CSI)->getIterator(); |
1851 | else |
1852 | InsertPt = DefBlock->getFirstInsertionPt(); |
1853 | } else { |
1854 | assert(!I->isTerminator() && "unexpected terminator" ); |
1855 | // For all other values, the spill is placed immediately after |
1856 | // the definition. |
1857 | InsertPt = I->getNextNode()->getIterator(); |
1858 | } |
1859 | } |
1860 | |
1861 | auto Index = FrameData.getFieldIndex(V: Def); |
1862 | Builder.SetInsertPoint(TheBB: InsertPt->getParent(), IP: InsertPt); |
1863 | auto *G = Builder.CreateConstInBoundsGEP2_32( |
1864 | Ty: FrameTy, Ptr: FramePtr, Idx0: 0, Idx1: Index, Name: Def->getName() + Twine(".spill.addr" )); |
1865 | if (ByValTy) { |
1866 | // For byval arguments, we need to store the pointed value in the frame, |
1867 | // instead of the pointer itself. |
1868 | auto *Value = Builder.CreateLoad(Ty: ByValTy, Ptr: Def); |
1869 | Builder.CreateAlignedStore(Val: Value, Ptr: G, Align: SpillAlignment); |
1870 | } else { |
1871 | Builder.CreateAlignedStore(Val: Def, Ptr: G, Align: SpillAlignment); |
1872 | } |
1873 | |
1874 | BasicBlock *CurrentBlock = nullptr; |
1875 | Value *CurrentReload = nullptr; |
1876 | for (auto *U : E.second) { |
1877 | // If we have not seen the use block, create a load instruction to reload |
1878 | // the spilled value from the coroutine frame. Populates the Value pointer |
1879 | // reference provided with the frame GEP. |
1880 | if (CurrentBlock != U->getParent()) { |
1881 | CurrentBlock = U->getParent(); |
1882 | Builder.SetInsertPoint(TheBB: CurrentBlock, |
1883 | IP: CurrentBlock->getFirstInsertionPt()); |
1884 | |
1885 | auto *GEP = GetFramePointer(E.first); |
1886 | GEP->setName(E.first->getName() + Twine(".reload.addr" )); |
1887 | if (ByValTy) |
1888 | CurrentReload = GEP; |
1889 | else |
1890 | CurrentReload = Builder.CreateAlignedLoad( |
1891 | Ty: FrameTy->getElementType(N: FrameData.getFieldIndex(V: E.first)), Ptr: GEP, |
1892 | Align: SpillAlignment, Name: E.first->getName() + Twine(".reload" )); |
1893 | |
1894 | TinyPtrVector<DbgDeclareInst *> DIs = findDbgDeclares(V: Def); |
1895 | TinyPtrVector<DbgVariableRecord *> DVRs = findDVRDeclares(V: Def); |
1896 | // Try best to find dbg.declare. If the spill is a temp, there may not |
1897 | // be a direct dbg.declare. Walk up the load chain to find one from an |
1898 | // alias. |
1899 | if (F->getSubprogram()) { |
1900 | auto *CurDef = Def; |
1901 | while (DIs.empty() && DVRs.empty() && isa<LoadInst>(Val: CurDef)) { |
1902 | auto *LdInst = cast<LoadInst>(Val: CurDef); |
1903 | // Only consider ptr to ptr same type load. |
1904 | if (LdInst->getPointerOperandType() != LdInst->getType()) |
1905 | break; |
1906 | CurDef = LdInst->getPointerOperand(); |
1907 | if (!isa<AllocaInst, LoadInst>(Val: CurDef)) |
1908 | break; |
1909 | DIs = findDbgDeclares(V: CurDef); |
1910 | DVRs = findDVRDeclares(V: CurDef); |
1911 | } |
1912 | } |
1913 | |
1914 | auto SalvageOne = [&](auto *DDI) { |
1915 | bool AllowUnresolved = false; |
1916 | // This dbg.declare is preserved for all coro-split function |
1917 | // fragments. It will be unreachable in the main function, and |
1918 | // processed by coro::salvageDebugInfo() by CoroCloner. |
1919 | if (UseNewDbgInfoFormat) { |
1920 | DbgVariableRecord *NewDVR = new DbgVariableRecord( |
1921 | ValueAsMetadata::get(V: CurrentReload), DDI->getVariable(), |
1922 | DDI->getExpression(), DDI->getDebugLoc(), |
1923 | DbgVariableRecord::LocationType::Declare); |
1924 | Builder.GetInsertPoint()->getParent()->insertDbgRecordBefore( |
1925 | DR: NewDVR, Here: Builder.GetInsertPoint()); |
1926 | } else { |
1927 | DIBuilder(*CurrentBlock->getParent()->getParent(), AllowUnresolved) |
1928 | .insertDeclare(CurrentReload, DDI->getVariable(), |
1929 | DDI->getExpression(), DDI->getDebugLoc(), |
1930 | &*Builder.GetInsertPoint()); |
1931 | } |
1932 | // This dbg.declare is for the main function entry point. It |
1933 | // will be deleted in all coro-split functions. |
1934 | coro::salvageDebugInfo(ArgToAllocaMap, *DDI, Shape.OptimizeFrame, |
1935 | false /*UseEntryValue*/); |
1936 | }; |
1937 | for_each(Range&: DIs, F: SalvageOne); |
1938 | for_each(Range&: DVRs, F: SalvageOne); |
1939 | } |
1940 | |
1941 | // If we have a single edge PHINode, remove it and replace it with a |
1942 | // reload from the coroutine frame. (We already took care of multi edge |
1943 | // PHINodes by rewriting them in the rewritePHIs function). |
1944 | if (auto *PN = dyn_cast<PHINode>(Val: U)) { |
1945 | assert(PN->getNumIncomingValues() == 1 && |
1946 | "unexpected number of incoming " |
1947 | "values in the PHINode" ); |
1948 | PN->replaceAllUsesWith(V: CurrentReload); |
1949 | PN->eraseFromParent(); |
1950 | continue; |
1951 | } |
1952 | |
1953 | // Replace all uses of CurrentValue in the current instruction with |
1954 | // reload. |
1955 | U->replaceUsesOfWith(From: Def, To: CurrentReload); |
1956 | // Instructions are added to Def's user list if the attached |
1957 | // debug records use Def. Update those now. |
1958 | for (DbgVariableRecord &DVR : filterDbgVars(R: U->getDbgRecordRange())) |
1959 | DVR.replaceVariableLocationOp(OldValue: Def, NewValue: CurrentReload, AllowEmpty: true); |
1960 | } |
1961 | } |
1962 | |
1963 | BasicBlock *FramePtrBB = Shape.getInsertPtAfterFramePtr()->getParent(); |
1964 | |
1965 | auto SpillBlock = FramePtrBB->splitBasicBlock( |
1966 | I: Shape.getInsertPtAfterFramePtr(), BBName: "AllocaSpillBB" ); |
1967 | SpillBlock->splitBasicBlock(I: &SpillBlock->front(), BBName: "PostSpill" ); |
1968 | Shape.AllocaSpillBlock = SpillBlock; |
1969 | |
1970 | // retcon and retcon.once lowering assumes all uses have been sunk. |
1971 | if (Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce || |
1972 | Shape.ABI == coro::ABI::Async) { |
1973 | // If we found any allocas, replace all of their remaining uses with Geps. |
1974 | Builder.SetInsertPoint(TheBB: SpillBlock, IP: SpillBlock->begin()); |
1975 | for (const auto &P : FrameData.Allocas) { |
1976 | AllocaInst *Alloca = P.Alloca; |
1977 | auto *G = GetFramePointer(Alloca); |
1978 | |
1979 | // We are not using ReplaceInstWithInst(P.first, cast<Instruction>(G)) |
1980 | // here, as we are changing location of the instruction. |
1981 | G->takeName(V: Alloca); |
1982 | Alloca->replaceAllUsesWith(V: G); |
1983 | Alloca->eraseFromParent(); |
1984 | } |
1985 | return; |
1986 | } |
1987 | |
1988 | // If we found any alloca, replace all of their remaining uses with GEP |
1989 | // instructions. To remain debugbility, we replace the uses of allocas for |
1990 | // dbg.declares and dbg.values with the reload from the frame. |
1991 | // Note: We cannot replace the alloca with GEP instructions indiscriminately, |
1992 | // as some of the uses may not be dominated by CoroBegin. |
1993 | Builder.SetInsertPoint(TheBB: Shape.AllocaSpillBlock, |
1994 | IP: Shape.AllocaSpillBlock->begin()); |
1995 | SmallVector<Instruction *, 4> UsersToUpdate; |
1996 | for (const auto &A : FrameData.Allocas) { |
1997 | AllocaInst *Alloca = A.Alloca; |
1998 | UsersToUpdate.clear(); |
1999 | for (User *U : Alloca->users()) { |
2000 | auto *I = cast<Instruction>(Val: U); |
2001 | if (DT.dominates(Def: CB, User: I)) |
2002 | UsersToUpdate.push_back(Elt: I); |
2003 | } |
2004 | if (UsersToUpdate.empty()) |
2005 | continue; |
2006 | auto *G = GetFramePointer(Alloca); |
2007 | G->setName(Alloca->getName() + Twine(".reload.addr" )); |
2008 | |
2009 | SmallVector<DbgVariableIntrinsic *, 4> DIs; |
2010 | SmallVector<DbgVariableRecord *> DbgVariableRecords; |
2011 | findDbgUsers(DbgInsts&: DIs, V: Alloca, DbgVariableRecords: &DbgVariableRecords); |
2012 | for (auto *DVI : DIs) |
2013 | DVI->replaceUsesOfWith(From: Alloca, To: G); |
2014 | for (auto *DVR : DbgVariableRecords) |
2015 | DVR->replaceVariableLocationOp(OldValue: Alloca, NewValue: G); |
2016 | |
2017 | for (Instruction *I : UsersToUpdate) { |
2018 | // It is meaningless to retain the lifetime intrinsics refer for the |
2019 | // member of coroutine frames and the meaningless lifetime intrinsics |
2020 | // are possible to block further optimizations. |
2021 | if (I->isLifetimeStartOrEnd()) { |
2022 | I->eraseFromParent(); |
2023 | continue; |
2024 | } |
2025 | |
2026 | I->replaceUsesOfWith(From: Alloca, To: G); |
2027 | } |
2028 | } |
2029 | Builder.SetInsertPoint(&*Shape.getInsertPtAfterFramePtr()); |
2030 | for (const auto &A : FrameData.Allocas) { |
2031 | AllocaInst *Alloca = A.Alloca; |
2032 | if (A.MayWriteBeforeCoroBegin) { |
2033 | // isEscaped really means potentially modified before CoroBegin. |
2034 | if (Alloca->isArrayAllocation()) |
2035 | report_fatal_error( |
2036 | reason: "Coroutines cannot handle copying of array allocas yet" ); |
2037 | |
2038 | auto *G = GetFramePointer(Alloca); |
2039 | auto *Value = Builder.CreateLoad(Ty: Alloca->getAllocatedType(), Ptr: Alloca); |
2040 | Builder.CreateStore(Val: Value, Ptr: G); |
2041 | } |
2042 | // For each alias to Alloca created before CoroBegin but used after |
2043 | // CoroBegin, we recreate them after CoroBegin by appplying the offset |
2044 | // to the pointer in the frame. |
2045 | for (const auto &Alias : A.Aliases) { |
2046 | auto *FramePtr = GetFramePointer(Alloca); |
2047 | auto &Value = *Alias.second; |
2048 | auto ITy = IntegerType::get(C, NumBits: Value.getBitWidth()); |
2049 | auto *AliasPtr = |
2050 | Builder.CreatePtrAdd(Ptr: FramePtr, Offset: ConstantInt::get(Ty: ITy, V: Value)); |
2051 | Alias.first->replaceUsesWithIf( |
2052 | New: AliasPtr, ShouldReplace: [&](Use &U) { return DT.dominates(Def: CB, U); }); |
2053 | } |
2054 | } |
2055 | |
2056 | // PromiseAlloca is not collected in FrameData.Allocas. So we don't handle |
2057 | // the case that the PromiseAlloca may have writes before CoroBegin in the |
2058 | // above codes. And it may be problematic in edge cases. See |
2059 | // https://github.com/llvm/llvm-project/issues/57861 for an example. |
2060 | if (Shape.ABI == coro::ABI::Switch && Shape.SwitchLowering.PromiseAlloca) { |
2061 | AllocaInst *PA = Shape.SwitchLowering.PromiseAlloca; |
2062 | // If there is memory accessing to promise alloca before CoroBegin; |
2063 | bool HasAccessingPromiseBeforeCB = llvm::any_of(Range: PA->uses(), P: [&](Use &U) { |
2064 | auto *Inst = dyn_cast<Instruction>(Val: U.getUser()); |
2065 | if (!Inst || DT.dominates(Def: CB, User: Inst)) |
2066 | return false; |
2067 | |
2068 | if (auto *CI = dyn_cast<CallInst>(Val: Inst)) { |
2069 | // It is fine if the call wouldn't write to the Promise. |
2070 | // This is possible for @llvm.coro.id intrinsics, which |
2071 | // would take the promise as the second argument as a |
2072 | // marker. |
2073 | if (CI->onlyReadsMemory() || |
2074 | CI->onlyReadsMemory(OpNo: CI->getArgOperandNo(U: &U))) |
2075 | return false; |
2076 | return true; |
2077 | } |
2078 | |
2079 | return isa<StoreInst>(Val: Inst) || |
2080 | // It may take too much time to track the uses. |
2081 | // Be conservative about the case the use may escape. |
2082 | isa<GetElementPtrInst>(Val: Inst) || |
2083 | // There would always be a bitcast for the promise alloca |
2084 | // before we enabled Opaque pointers. And now given |
2085 | // opaque pointers are enabled by default. This should be |
2086 | // fine. |
2087 | isa<BitCastInst>(Val: Inst); |
2088 | }); |
2089 | if (HasAccessingPromiseBeforeCB) { |
2090 | Builder.SetInsertPoint(&*Shape.getInsertPtAfterFramePtr()); |
2091 | auto *G = GetFramePointer(PA); |
2092 | auto *Value = Builder.CreateLoad(Ty: PA->getAllocatedType(), Ptr: PA); |
2093 | Builder.CreateStore(Val: Value, Ptr: G); |
2094 | } |
2095 | } |
2096 | } |
2097 | |
2098 | // Moves the values in the PHIs in SuccBB that correspong to PredBB into a new |
2099 | // PHI in InsertedBB. |
2100 | static void movePHIValuesToInsertedBlock(BasicBlock *SuccBB, |
2101 | BasicBlock *InsertedBB, |
2102 | BasicBlock *PredBB, |
2103 | PHINode *UntilPHI = nullptr) { |
2104 | auto *PN = cast<PHINode>(Val: &SuccBB->front()); |
2105 | do { |
2106 | int Index = PN->getBasicBlockIndex(BB: InsertedBB); |
2107 | Value *V = PN->getIncomingValue(i: Index); |
2108 | PHINode *InputV = PHINode::Create( |
2109 | Ty: V->getType(), NumReservedValues: 1, NameStr: V->getName() + Twine("." ) + SuccBB->getName()); |
2110 | InputV->insertBefore(InsertPos: InsertedBB->begin()); |
2111 | InputV->addIncoming(V, BB: PredBB); |
2112 | PN->setIncomingValue(i: Index, V: InputV); |
2113 | PN = dyn_cast<PHINode>(Val: PN->getNextNode()); |
2114 | } while (PN != UntilPHI); |
2115 | } |
2116 | |
2117 | // Rewrites the PHI Nodes in a cleanuppad. |
2118 | static void rewritePHIsForCleanupPad(BasicBlock *CleanupPadBB, |
2119 | CleanupPadInst *CleanupPad) { |
2120 | // For every incoming edge to a CleanupPad we will create a new block holding |
2121 | // all incoming values in single-value PHI nodes. We will then create another |
2122 | // block to act as a dispather (as all unwind edges for related EH blocks |
2123 | // must be the same). |
2124 | // |
2125 | // cleanuppad: |
2126 | // %2 = phi i32[%0, %catchswitch], [%1, %catch.1] |
2127 | // %3 = cleanuppad within none [] |
2128 | // |
2129 | // It will create: |
2130 | // |
2131 | // cleanuppad.corodispatch |
2132 | // %2 = phi i8[0, %catchswitch], [1, %catch.1] |
2133 | // %3 = cleanuppad within none [] |
2134 | // switch i8 % 2, label %unreachable |
2135 | // [i8 0, label %cleanuppad.from.catchswitch |
2136 | // i8 1, label %cleanuppad.from.catch.1] |
2137 | // cleanuppad.from.catchswitch: |
2138 | // %4 = phi i32 [%0, %catchswitch] |
2139 | // br %label cleanuppad |
2140 | // cleanuppad.from.catch.1: |
2141 | // %6 = phi i32 [%1, %catch.1] |
2142 | // br %label cleanuppad |
2143 | // cleanuppad: |
2144 | // %8 = phi i32 [%4, %cleanuppad.from.catchswitch], |
2145 | // [%6, %cleanuppad.from.catch.1] |
2146 | |
2147 | // Unreachable BB, in case switching on an invalid value in the dispatcher. |
2148 | auto *UnreachBB = BasicBlock::Create( |
2149 | Context&: CleanupPadBB->getContext(), Name: "unreachable" , Parent: CleanupPadBB->getParent()); |
2150 | IRBuilder<> Builder(UnreachBB); |
2151 | Builder.CreateUnreachable(); |
2152 | |
2153 | // Create a new cleanuppad which will be the dispatcher. |
2154 | auto *NewCleanupPadBB = |
2155 | BasicBlock::Create(Context&: CleanupPadBB->getContext(), |
2156 | Name: CleanupPadBB->getName() + Twine(".corodispatch" ), |
2157 | Parent: CleanupPadBB->getParent(), InsertBefore: CleanupPadBB); |
2158 | Builder.SetInsertPoint(NewCleanupPadBB); |
2159 | auto *SwitchType = Builder.getInt8Ty(); |
2160 | auto *SetDispatchValuePN = |
2161 | Builder.CreatePHI(Ty: SwitchType, NumReservedValues: pred_size(BB: CleanupPadBB)); |
2162 | CleanupPad->removeFromParent(); |
2163 | CleanupPad->insertAfter(InsertPos: SetDispatchValuePN); |
2164 | auto *SwitchOnDispatch = Builder.CreateSwitch(V: SetDispatchValuePN, Dest: UnreachBB, |
2165 | NumCases: pred_size(BB: CleanupPadBB)); |
2166 | |
2167 | int SwitchIndex = 0; |
2168 | SmallVector<BasicBlock *, 8> Preds(predecessors(BB: CleanupPadBB)); |
2169 | for (BasicBlock *Pred : Preds) { |
2170 | // Create a new cleanuppad and move the PHI values to there. |
2171 | auto *CaseBB = BasicBlock::Create(Context&: CleanupPadBB->getContext(), |
2172 | Name: CleanupPadBB->getName() + |
2173 | Twine(".from." ) + Pred->getName(), |
2174 | Parent: CleanupPadBB->getParent(), InsertBefore: CleanupPadBB); |
2175 | updatePhiNodes(DestBB: CleanupPadBB, OldPred: Pred, NewPred: CaseBB); |
2176 | CaseBB->setName(CleanupPadBB->getName() + Twine(".from." ) + |
2177 | Pred->getName()); |
2178 | Builder.SetInsertPoint(CaseBB); |
2179 | Builder.CreateBr(Dest: CleanupPadBB); |
2180 | movePHIValuesToInsertedBlock(SuccBB: CleanupPadBB, InsertedBB: CaseBB, PredBB: NewCleanupPadBB); |
2181 | |
2182 | // Update this Pred to the new unwind point. |
2183 | setUnwindEdgeTo(TI: Pred->getTerminator(), Succ: NewCleanupPadBB); |
2184 | |
2185 | // Setup the switch in the dispatcher. |
2186 | auto *SwitchConstant = ConstantInt::get(Ty: SwitchType, V: SwitchIndex); |
2187 | SetDispatchValuePN->addIncoming(V: SwitchConstant, BB: Pred); |
2188 | SwitchOnDispatch->addCase(OnVal: SwitchConstant, Dest: CaseBB); |
2189 | SwitchIndex++; |
2190 | } |
2191 | } |
2192 | |
2193 | static void cleanupSinglePredPHIs(Function &F) { |
2194 | SmallVector<PHINode *, 32> Worklist; |
2195 | for (auto &BB : F) { |
2196 | for (auto &Phi : BB.phis()) { |
2197 | if (Phi.getNumIncomingValues() == 1) { |
2198 | Worklist.push_back(Elt: &Phi); |
2199 | } else |
2200 | break; |
2201 | } |
2202 | } |
2203 | while (!Worklist.empty()) { |
2204 | auto *Phi = Worklist.pop_back_val(); |
2205 | auto *OriginalValue = Phi->getIncomingValue(i: 0); |
2206 | Phi->replaceAllUsesWith(V: OriginalValue); |
2207 | } |
2208 | } |
2209 | |
2210 | static void rewritePHIs(BasicBlock &BB) { |
2211 | // For every incoming edge we will create a block holding all |
2212 | // incoming values in a single PHI nodes. |
2213 | // |
2214 | // loop: |
2215 | // %n.val = phi i32[%n, %entry], [%inc, %loop] |
2216 | // |
2217 | // It will create: |
2218 | // |
2219 | // loop.from.entry: |
2220 | // %n.loop.pre = phi i32 [%n, %entry] |
2221 | // br %label loop |
2222 | // loop.from.loop: |
2223 | // %inc.loop.pre = phi i32 [%inc, %loop] |
2224 | // br %label loop |
2225 | // |
2226 | // After this rewrite, further analysis will ignore any phi nodes with more |
2227 | // than one incoming edge. |
2228 | |
2229 | // TODO: Simplify PHINodes in the basic block to remove duplicate |
2230 | // predecessors. |
2231 | |
2232 | // Special case for CleanupPad: all EH blocks must have the same unwind edge |
2233 | // so we need to create an additional "dispatcher" block. |
2234 | if (auto *CleanupPad = |
2235 | dyn_cast_or_null<CleanupPadInst>(Val: BB.getFirstNonPHI())) { |
2236 | SmallVector<BasicBlock *, 8> Preds(predecessors(BB: &BB)); |
2237 | for (BasicBlock *Pred : Preds) { |
2238 | if (CatchSwitchInst *CS = |
2239 | dyn_cast<CatchSwitchInst>(Val: Pred->getTerminator())) { |
2240 | // CleanupPad with a CatchSwitch predecessor: therefore this is an |
2241 | // unwind destination that needs to be handle specially. |
2242 | assert(CS->getUnwindDest() == &BB); |
2243 | (void)CS; |
2244 | rewritePHIsForCleanupPad(CleanupPadBB: &BB, CleanupPad); |
2245 | return; |
2246 | } |
2247 | } |
2248 | } |
2249 | |
2250 | LandingPadInst *LandingPad = nullptr; |
2251 | PHINode *ReplPHI = nullptr; |
2252 | if ((LandingPad = dyn_cast_or_null<LandingPadInst>(Val: BB.getFirstNonPHI()))) { |
2253 | // ehAwareSplitEdge will clone the LandingPad in all the edge blocks. |
2254 | // We replace the original landing pad with a PHINode that will collect the |
2255 | // results from all of them. |
2256 | ReplPHI = PHINode::Create(Ty: LandingPad->getType(), NumReservedValues: 1, NameStr: "" ); |
2257 | ReplPHI->insertBefore(InsertPos: LandingPad->getIterator()); |
2258 | ReplPHI->takeName(V: LandingPad); |
2259 | LandingPad->replaceAllUsesWith(V: ReplPHI); |
2260 | // We will erase the original landing pad at the end of this function after |
2261 | // ehAwareSplitEdge cloned it in the transition blocks. |
2262 | } |
2263 | |
2264 | SmallVector<BasicBlock *, 8> Preds(predecessors(BB: &BB)); |
2265 | for (BasicBlock *Pred : Preds) { |
2266 | auto *IncomingBB = ehAwareSplitEdge(BB: Pred, Succ: &BB, OriginalPad: LandingPad, LandingPadReplacement: ReplPHI); |
2267 | IncomingBB->setName(BB.getName() + Twine(".from." ) + Pred->getName()); |
2268 | |
2269 | // Stop the moving of values at ReplPHI, as this is either null or the PHI |
2270 | // that replaced the landing pad. |
2271 | movePHIValuesToInsertedBlock(SuccBB: &BB, InsertedBB: IncomingBB, PredBB: Pred, UntilPHI: ReplPHI); |
2272 | } |
2273 | |
2274 | if (LandingPad) { |
2275 | // Calls to ehAwareSplitEdge function cloned the original lading pad. |
2276 | // No longer need it. |
2277 | LandingPad->eraseFromParent(); |
2278 | } |
2279 | } |
2280 | |
2281 | static void rewritePHIs(Function &F) { |
2282 | SmallVector<BasicBlock *, 8> WorkList; |
2283 | |
2284 | for (BasicBlock &BB : F) |
2285 | if (auto *PN = dyn_cast<PHINode>(Val: &BB.front())) |
2286 | if (PN->getNumIncomingValues() > 1) |
2287 | WorkList.push_back(Elt: &BB); |
2288 | |
2289 | for (BasicBlock *BB : WorkList) |
2290 | rewritePHIs(BB&: *BB); |
2291 | } |
2292 | |
2293 | /// Default materializable callback |
2294 | // Check for instructions that we can recreate on resume as opposed to spill |
2295 | // the result into a coroutine frame. |
2296 | bool coro::defaultMaterializable(Instruction &V) { |
2297 | return (isa<CastInst>(Val: &V) || isa<GetElementPtrInst>(Val: &V) || |
2298 | isa<BinaryOperator>(Val: &V) || isa<CmpInst>(Val: &V) || isa<SelectInst>(Val: &V)); |
2299 | } |
2300 | |
2301 | // Check for structural coroutine intrinsics that should not be spilled into |
2302 | // the coroutine frame. |
2303 | static bool isCoroutineStructureIntrinsic(Instruction &I) { |
2304 | return isa<CoroIdInst>(Val: &I) || isa<CoroSaveInst>(Val: &I) || |
2305 | isa<CoroSuspendInst>(Val: &I); |
2306 | } |
2307 | |
2308 | // For each instruction identified as materializable across the suspend point, |
2309 | // and its associated DAG of other rematerializable instructions, |
2310 | // recreate the DAG of instructions after the suspend point. |
2311 | static void rewriteMaterializableInstructions( |
2312 | const SmallMapVector<Instruction *, std::unique_ptr<RematGraph>, 8> |
2313 | &AllRemats) { |
2314 | // This has to be done in 2 phases |
2315 | // Do the remats and record the required defs to be replaced in the |
2316 | // original use instructions |
2317 | // Once all the remats are complete, replace the uses in the final |
2318 | // instructions with the new defs |
2319 | typedef struct { |
2320 | Instruction *Use; |
2321 | Instruction *Def; |
2322 | Instruction *Remat; |
2323 | } ProcessNode; |
2324 | |
2325 | SmallVector<ProcessNode> FinalInstructionsToProcess; |
2326 | |
2327 | for (const auto &E : AllRemats) { |
2328 | Instruction *Use = E.first; |
2329 | Instruction *CurrentMaterialization = nullptr; |
2330 | RematGraph *RG = E.second.get(); |
2331 | ReversePostOrderTraversal<RematGraph *> RPOT(RG); |
2332 | SmallVector<Instruction *> InstructionsToProcess; |
2333 | |
2334 | // If the target use is actually a suspend instruction then we have to |
2335 | // insert the remats into the end of the predecessor (there should only be |
2336 | // one). This is so that suspend blocks always have the suspend instruction |
2337 | // as the first instruction. |
2338 | auto InsertPoint = &*Use->getParent()->getFirstInsertionPt(); |
2339 | if (isa<AnyCoroSuspendInst>(Val: Use)) { |
2340 | BasicBlock *SuspendPredecessorBlock = |
2341 | Use->getParent()->getSinglePredecessor(); |
2342 | assert(SuspendPredecessorBlock && "malformed coro suspend instruction" ); |
2343 | InsertPoint = SuspendPredecessorBlock->getTerminator(); |
2344 | } |
2345 | |
2346 | // Note: skip the first instruction as this is the actual use that we're |
2347 | // rematerializing everything for. |
2348 | auto I = RPOT.begin(); |
2349 | ++I; |
2350 | for (; I != RPOT.end(); ++I) { |
2351 | Instruction *D = (*I)->Node; |
2352 | CurrentMaterialization = D->clone(); |
2353 | CurrentMaterialization->setName(D->getName()); |
2354 | CurrentMaterialization->insertBefore(InsertPos: InsertPoint); |
2355 | InsertPoint = CurrentMaterialization; |
2356 | |
2357 | // Replace all uses of Def in the instructions being added as part of this |
2358 | // rematerialization group |
2359 | for (auto &I : InstructionsToProcess) |
2360 | I->replaceUsesOfWith(From: D, To: CurrentMaterialization); |
2361 | |
2362 | // Don't replace the final use at this point as this can cause problems |
2363 | // for other materializations. Instead, for any final use that uses a |
2364 | // define that's being rematerialized, record the replace values |
2365 | for (unsigned i = 0, E = Use->getNumOperands(); i != E; ++i) |
2366 | if (Use->getOperand(i) == D) // Is this operand pointing to oldval? |
2367 | FinalInstructionsToProcess.push_back( |
2368 | Elt: {.Use: Use, .Def: D, .Remat: CurrentMaterialization}); |
2369 | |
2370 | InstructionsToProcess.push_back(Elt: CurrentMaterialization); |
2371 | } |
2372 | } |
2373 | |
2374 | // Finally, replace the uses with the defines that we've just rematerialized |
2375 | for (auto &R : FinalInstructionsToProcess) { |
2376 | if (auto *PN = dyn_cast<PHINode>(Val: R.Use)) { |
2377 | assert(PN->getNumIncomingValues() == 1 && "unexpected number of incoming " |
2378 | "values in the PHINode" ); |
2379 | PN->replaceAllUsesWith(V: R.Remat); |
2380 | PN->eraseFromParent(); |
2381 | continue; |
2382 | } |
2383 | R.Use->replaceUsesOfWith(From: R.Def, To: R.Remat); |
2384 | } |
2385 | } |
2386 | |
2387 | // Splits the block at a particular instruction unless it is the first |
2388 | // instruction in the block with a single predecessor. |
2389 | static BasicBlock *splitBlockIfNotFirst(Instruction *I, const Twine &Name) { |
2390 | auto *BB = I->getParent(); |
2391 | if (&BB->front() == I) { |
2392 | if (BB->getSinglePredecessor()) { |
2393 | BB->setName(Name); |
2394 | return BB; |
2395 | } |
2396 | } |
2397 | return BB->splitBasicBlock(I, BBName: Name); |
2398 | } |
2399 | |
2400 | // Split above and below a particular instruction so that it |
2401 | // will be all alone by itself in a block. |
2402 | static void splitAround(Instruction *I, const Twine &Name) { |
2403 | splitBlockIfNotFirst(I, Name); |
2404 | splitBlockIfNotFirst(I: I->getNextNode(), Name: "After" + Name); |
2405 | } |
2406 | |
2407 | static bool isSuspendBlock(BasicBlock *BB) { |
2408 | return isa<AnyCoroSuspendInst>(Val: BB->front()); |
2409 | } |
2410 | |
2411 | typedef SmallPtrSet<BasicBlock*, 8> VisitedBlocksSet; |
2412 | |
2413 | /// Does control flow starting at the given block ever reach a suspend |
2414 | /// instruction before reaching a block in VisitedOrFreeBBs? |
2415 | static bool isSuspendReachableFrom(BasicBlock *From, |
2416 | VisitedBlocksSet &VisitedOrFreeBBs) { |
2417 | // Eagerly try to add this block to the visited set. If it's already |
2418 | // there, stop recursing; this path doesn't reach a suspend before |
2419 | // either looping or reaching a freeing block. |
2420 | if (!VisitedOrFreeBBs.insert(Ptr: From).second) |
2421 | return false; |
2422 | |
2423 | // We assume that we'll already have split suspends into their own blocks. |
2424 | if (isSuspendBlock(BB: From)) |
2425 | return true; |
2426 | |
2427 | // Recurse on the successors. |
2428 | for (auto *Succ : successors(BB: From)) { |
2429 | if (isSuspendReachableFrom(From: Succ, VisitedOrFreeBBs)) |
2430 | return true; |
2431 | } |
2432 | |
2433 | return false; |
2434 | } |
2435 | |
2436 | /// Is the given alloca "local", i.e. bounded in lifetime to not cross a |
2437 | /// suspend point? |
2438 | static bool isLocalAlloca(CoroAllocaAllocInst *AI) { |
2439 | // Seed the visited set with all the basic blocks containing a free |
2440 | // so that we won't pass them up. |
2441 | VisitedBlocksSet VisitedOrFreeBBs; |
2442 | for (auto *User : AI->users()) { |
2443 | if (auto FI = dyn_cast<CoroAllocaFreeInst>(Val: User)) |
2444 | VisitedOrFreeBBs.insert(Ptr: FI->getParent()); |
2445 | } |
2446 | |
2447 | return !isSuspendReachableFrom(From: AI->getParent(), VisitedOrFreeBBs); |
2448 | } |
2449 | |
2450 | /// After we split the coroutine, will the given basic block be along |
2451 | /// an obvious exit path for the resumption function? |
2452 | static bool willLeaveFunctionImmediatelyAfter(BasicBlock *BB, |
2453 | unsigned depth = 3) { |
2454 | // If we've bottomed out our depth count, stop searching and assume |
2455 | // that the path might loop back. |
2456 | if (depth == 0) return false; |
2457 | |
2458 | // If this is a suspend block, we're about to exit the resumption function. |
2459 | if (isSuspendBlock(BB)) return true; |
2460 | |
2461 | // Recurse into the successors. |
2462 | for (auto *Succ : successors(BB)) { |
2463 | if (!willLeaveFunctionImmediatelyAfter(BB: Succ, depth: depth - 1)) |
2464 | return false; |
2465 | } |
2466 | |
2467 | // If none of the successors leads back in a loop, we're on an exit/abort. |
2468 | return true; |
2469 | } |
2470 | |
2471 | static bool localAllocaNeedsStackSave(CoroAllocaAllocInst *AI) { |
2472 | // Look for a free that isn't sufficiently obviously followed by |
2473 | // either a suspend or a termination, i.e. something that will leave |
2474 | // the coro resumption frame. |
2475 | for (auto *U : AI->users()) { |
2476 | auto FI = dyn_cast<CoroAllocaFreeInst>(Val: U); |
2477 | if (!FI) continue; |
2478 | |
2479 | if (!willLeaveFunctionImmediatelyAfter(BB: FI->getParent())) |
2480 | return true; |
2481 | } |
2482 | |
2483 | // If we never found one, we don't need a stack save. |
2484 | return false; |
2485 | } |
2486 | |
2487 | /// Turn each of the given local allocas into a normal (dynamic) alloca |
2488 | /// instruction. |
2489 | static void lowerLocalAllocas(ArrayRef<CoroAllocaAllocInst*> LocalAllocas, |
2490 | SmallVectorImpl<Instruction*> &DeadInsts) { |
2491 | for (auto *AI : LocalAllocas) { |
2492 | IRBuilder<> Builder(AI); |
2493 | |
2494 | // Save the stack depth. Try to avoid doing this if the stackrestore |
2495 | // is going to immediately precede a return or something. |
2496 | Value *StackSave = nullptr; |
2497 | if (localAllocaNeedsStackSave(AI)) |
2498 | StackSave = Builder.CreateStackSave(); |
2499 | |
2500 | // Allocate memory. |
2501 | auto Alloca = Builder.CreateAlloca(Ty: Builder.getInt8Ty(), ArraySize: AI->getSize()); |
2502 | Alloca->setAlignment(AI->getAlignment()); |
2503 | |
2504 | for (auto *U : AI->users()) { |
2505 | // Replace gets with the allocation. |
2506 | if (isa<CoroAllocaGetInst>(Val: U)) { |
2507 | U->replaceAllUsesWith(V: Alloca); |
2508 | |
2509 | // Replace frees with stackrestores. This is safe because |
2510 | // alloca.alloc is required to obey a stack discipline, although we |
2511 | // don't enforce that structurally. |
2512 | } else { |
2513 | auto FI = cast<CoroAllocaFreeInst>(Val: U); |
2514 | if (StackSave) { |
2515 | Builder.SetInsertPoint(FI); |
2516 | Builder.CreateStackRestore(Ptr: StackSave); |
2517 | } |
2518 | } |
2519 | DeadInsts.push_back(Elt: cast<Instruction>(Val: U)); |
2520 | } |
2521 | |
2522 | DeadInsts.push_back(Elt: AI); |
2523 | } |
2524 | } |
2525 | |
2526 | /// Turn the given coro.alloca.alloc call into a dynamic allocation. |
2527 | /// This happens during the all-instructions iteration, so it must not |
2528 | /// delete the call. |
2529 | static Instruction *lowerNonLocalAlloca(CoroAllocaAllocInst *AI, |
2530 | coro::Shape &Shape, |
2531 | SmallVectorImpl<Instruction*> &DeadInsts) { |
2532 | IRBuilder<> Builder(AI); |
2533 | auto Alloc = Shape.emitAlloc(Builder, Size: AI->getSize(), CG: nullptr); |
2534 | |
2535 | for (User *U : AI->users()) { |
2536 | if (isa<CoroAllocaGetInst>(Val: U)) { |
2537 | U->replaceAllUsesWith(V: Alloc); |
2538 | } else { |
2539 | auto FI = cast<CoroAllocaFreeInst>(Val: U); |
2540 | Builder.SetInsertPoint(FI); |
2541 | Shape.emitDealloc(Builder, Ptr: Alloc, CG: nullptr); |
2542 | } |
2543 | DeadInsts.push_back(Elt: cast<Instruction>(Val: U)); |
2544 | } |
2545 | |
2546 | // Push this on last so that it gets deleted after all the others. |
2547 | DeadInsts.push_back(Elt: AI); |
2548 | |
2549 | // Return the new allocation value so that we can check for needed spills. |
2550 | return cast<Instruction>(Val: Alloc); |
2551 | } |
2552 | |
2553 | /// Get the current swifterror value. |
2554 | static Value *emitGetSwiftErrorValue(IRBuilder<> &Builder, Type *ValueTy, |
2555 | coro::Shape &Shape) { |
2556 | // Make a fake function pointer as a sort of intrinsic. |
2557 | auto FnTy = FunctionType::get(Result: ValueTy, Params: {}, isVarArg: false); |
2558 | auto Fn = ConstantPointerNull::get(T: Builder.getPtrTy()); |
2559 | |
2560 | auto Call = Builder.CreateCall(FTy: FnTy, Callee: Fn, Args: {}); |
2561 | Shape.SwiftErrorOps.push_back(Elt: Call); |
2562 | |
2563 | return Call; |
2564 | } |
2565 | |
2566 | /// Set the given value as the current swifterror value. |
2567 | /// |
2568 | /// Returns a slot that can be used as a swifterror slot. |
2569 | static Value *emitSetSwiftErrorValue(IRBuilder<> &Builder, Value *V, |
2570 | coro::Shape &Shape) { |
2571 | // Make a fake function pointer as a sort of intrinsic. |
2572 | auto FnTy = FunctionType::get(Result: Builder.getPtrTy(), |
2573 | Params: {V->getType()}, isVarArg: false); |
2574 | auto Fn = ConstantPointerNull::get(T: Builder.getPtrTy()); |
2575 | |
2576 | auto Call = Builder.CreateCall(FTy: FnTy, Callee: Fn, Args: { V }); |
2577 | Shape.SwiftErrorOps.push_back(Elt: Call); |
2578 | |
2579 | return Call; |
2580 | } |
2581 | |
2582 | /// Set the swifterror value from the given alloca before a call, |
2583 | /// then put in back in the alloca afterwards. |
2584 | /// |
2585 | /// Returns an address that will stand in for the swifterror slot |
2586 | /// until splitting. |
2587 | static Value *emitSetAndGetSwiftErrorValueAround(Instruction *Call, |
2588 | AllocaInst *Alloca, |
2589 | coro::Shape &Shape) { |
2590 | auto ValueTy = Alloca->getAllocatedType(); |
2591 | IRBuilder<> Builder(Call); |
2592 | |
2593 | // Load the current value from the alloca and set it as the |
2594 | // swifterror value. |
2595 | auto ValueBeforeCall = Builder.CreateLoad(Ty: ValueTy, Ptr: Alloca); |
2596 | auto Addr = emitSetSwiftErrorValue(Builder, V: ValueBeforeCall, Shape); |
2597 | |
2598 | // Move to after the call. Since swifterror only has a guaranteed |
2599 | // value on normal exits, we can ignore implicit and explicit unwind |
2600 | // edges. |
2601 | if (isa<CallInst>(Val: Call)) { |
2602 | Builder.SetInsertPoint(Call->getNextNode()); |
2603 | } else { |
2604 | auto Invoke = cast<InvokeInst>(Val: Call); |
2605 | Builder.SetInsertPoint(Invoke->getNormalDest()->getFirstNonPHIOrDbg()); |
2606 | } |
2607 | |
2608 | // Get the current swifterror value and store it to the alloca. |
2609 | auto ValueAfterCall = emitGetSwiftErrorValue(Builder, ValueTy, Shape); |
2610 | Builder.CreateStore(Val: ValueAfterCall, Ptr: Alloca); |
2611 | |
2612 | return Addr; |
2613 | } |
2614 | |
2615 | /// Eliminate a formerly-swifterror alloca by inserting the get/set |
2616 | /// intrinsics and attempting to MemToReg the alloca away. |
2617 | static void eliminateSwiftErrorAlloca(Function &F, AllocaInst *Alloca, |
2618 | coro::Shape &Shape) { |
2619 | for (Use &Use : llvm::make_early_inc_range(Range: Alloca->uses())) { |
2620 | // swifterror values can only be used in very specific ways. |
2621 | // We take advantage of that here. |
2622 | auto User = Use.getUser(); |
2623 | if (isa<LoadInst>(Val: User) || isa<StoreInst>(Val: User)) |
2624 | continue; |
2625 | |
2626 | assert(isa<CallInst>(User) || isa<InvokeInst>(User)); |
2627 | auto Call = cast<Instruction>(Val: User); |
2628 | |
2629 | auto Addr = emitSetAndGetSwiftErrorValueAround(Call, Alloca, Shape); |
2630 | |
2631 | // Use the returned slot address as the call argument. |
2632 | Use.set(Addr); |
2633 | } |
2634 | |
2635 | // All the uses should be loads and stores now. |
2636 | assert(isAllocaPromotable(Alloca)); |
2637 | } |
2638 | |
2639 | /// "Eliminate" a swifterror argument by reducing it to the alloca case |
2640 | /// and then loading and storing in the prologue and epilog. |
2641 | /// |
2642 | /// The argument keeps the swifterror flag. |
2643 | static void eliminateSwiftErrorArgument(Function &F, Argument &Arg, |
2644 | coro::Shape &Shape, |
2645 | SmallVectorImpl<AllocaInst*> &AllocasToPromote) { |
2646 | IRBuilder<> Builder(F.getEntryBlock().getFirstNonPHIOrDbg()); |
2647 | |
2648 | auto ArgTy = cast<PointerType>(Val: Arg.getType()); |
2649 | auto ValueTy = PointerType::getUnqual(C&: F.getContext()); |
2650 | |
2651 | // Reduce to the alloca case: |
2652 | |
2653 | // Create an alloca and replace all uses of the arg with it. |
2654 | auto Alloca = Builder.CreateAlloca(Ty: ValueTy, AddrSpace: ArgTy->getAddressSpace()); |
2655 | Arg.replaceAllUsesWith(V: Alloca); |
2656 | |
2657 | // Set an initial value in the alloca. swifterror is always null on entry. |
2658 | auto InitialValue = Constant::getNullValue(Ty: ValueTy); |
2659 | Builder.CreateStore(Val: InitialValue, Ptr: Alloca); |
2660 | |
2661 | // Find all the suspends in the function and save and restore around them. |
2662 | for (auto *Suspend : Shape.CoroSuspends) { |
2663 | (void) emitSetAndGetSwiftErrorValueAround(Call: Suspend, Alloca, Shape); |
2664 | } |
2665 | |
2666 | // Find all the coro.ends in the function and restore the error value. |
2667 | for (auto *End : Shape.CoroEnds) { |
2668 | Builder.SetInsertPoint(End); |
2669 | auto FinalValue = Builder.CreateLoad(Ty: ValueTy, Ptr: Alloca); |
2670 | (void) emitSetSwiftErrorValue(Builder, V: FinalValue, Shape); |
2671 | } |
2672 | |
2673 | // Now we can use the alloca logic. |
2674 | AllocasToPromote.push_back(Elt: Alloca); |
2675 | eliminateSwiftErrorAlloca(F, Alloca, Shape); |
2676 | } |
2677 | |
2678 | /// Eliminate all problematic uses of swifterror arguments and allocas |
2679 | /// from the function. We'll fix them up later when splitting the function. |
2680 | static void eliminateSwiftError(Function &F, coro::Shape &Shape) { |
2681 | SmallVector<AllocaInst*, 4> AllocasToPromote; |
2682 | |
2683 | // Look for a swifterror argument. |
2684 | for (auto &Arg : F.args()) { |
2685 | if (!Arg.hasSwiftErrorAttr()) continue; |
2686 | |
2687 | eliminateSwiftErrorArgument(F, Arg, Shape, AllocasToPromote); |
2688 | break; |
2689 | } |
2690 | |
2691 | // Look for swifterror allocas. |
2692 | for (auto &Inst : F.getEntryBlock()) { |
2693 | auto Alloca = dyn_cast<AllocaInst>(Val: &Inst); |
2694 | if (!Alloca || !Alloca->isSwiftError()) continue; |
2695 | |
2696 | // Clear the swifterror flag. |
2697 | Alloca->setSwiftError(false); |
2698 | |
2699 | AllocasToPromote.push_back(Elt: Alloca); |
2700 | eliminateSwiftErrorAlloca(F, Alloca, Shape); |
2701 | } |
2702 | |
2703 | // If we have any allocas to promote, compute a dominator tree and |
2704 | // promote them en masse. |
2705 | if (!AllocasToPromote.empty()) { |
2706 | DominatorTree DT(F); |
2707 | PromoteMemToReg(Allocas: AllocasToPromote, DT); |
2708 | } |
2709 | } |
2710 | |
2711 | /// retcon and retcon.once conventions assume that all spill uses can be sunk |
2712 | /// after the coro.begin intrinsic. |
2713 | static void sinkSpillUsesAfterCoroBegin(Function &F, |
2714 | const FrameDataInfo &FrameData, |
2715 | CoroBeginInst *CoroBegin) { |
2716 | DominatorTree Dom(F); |
2717 | |
2718 | SmallSetVector<Instruction *, 32> ToMove; |
2719 | SmallVector<Instruction *, 32> Worklist; |
2720 | |
2721 | // Collect all users that precede coro.begin. |
2722 | for (auto *Def : FrameData.getAllDefs()) { |
2723 | for (User *U : Def->users()) { |
2724 | auto Inst = cast<Instruction>(Val: U); |
2725 | if (Inst->getParent() != CoroBegin->getParent() || |
2726 | Dom.dominates(Def: CoroBegin, User: Inst)) |
2727 | continue; |
2728 | if (ToMove.insert(X: Inst)) |
2729 | Worklist.push_back(Elt: Inst); |
2730 | } |
2731 | } |
2732 | // Recursively collect users before coro.begin. |
2733 | while (!Worklist.empty()) { |
2734 | auto *Def = Worklist.pop_back_val(); |
2735 | for (User *U : Def->users()) { |
2736 | auto Inst = cast<Instruction>(Val: U); |
2737 | if (Dom.dominates(Def: CoroBegin, User: Inst)) |
2738 | continue; |
2739 | if (ToMove.insert(X: Inst)) |
2740 | Worklist.push_back(Elt: Inst); |
2741 | } |
2742 | } |
2743 | |
2744 | // Sort by dominance. |
2745 | SmallVector<Instruction *, 64> InsertionList(ToMove.begin(), ToMove.end()); |
2746 | llvm::sort(C&: InsertionList, Comp: [&Dom](Instruction *A, Instruction *B) -> bool { |
2747 | // If a dominates b it should preceed (<) b. |
2748 | return Dom.dominates(Def: A, User: B); |
2749 | }); |
2750 | |
2751 | Instruction *InsertPt = CoroBegin->getNextNode(); |
2752 | for (Instruction *Inst : InsertionList) |
2753 | Inst->moveBefore(MovePos: InsertPt); |
2754 | } |
2755 | |
2756 | /// For each local variable that all of its user are only used inside one of |
2757 | /// suspended region, we sink their lifetime.start markers to the place where |
2758 | /// after the suspend block. Doing so minimizes the lifetime of each variable, |
2759 | /// hence minimizing the amount of data we end up putting on the frame. |
2760 | static void sinkLifetimeStartMarkers(Function &F, coro::Shape &Shape, |
2761 | SuspendCrossingInfo &Checker, |
2762 | const DominatorTree &DT) { |
2763 | if (F.hasOptNone()) |
2764 | return; |
2765 | |
2766 | // Collect all possible basic blocks which may dominate all uses of allocas. |
2767 | SmallPtrSet<BasicBlock *, 4> DomSet; |
2768 | DomSet.insert(Ptr: &F.getEntryBlock()); |
2769 | for (auto *CSI : Shape.CoroSuspends) { |
2770 | BasicBlock *SuspendBlock = CSI->getParent(); |
2771 | assert(isSuspendBlock(SuspendBlock) && SuspendBlock->getSingleSuccessor() && |
2772 | "should have split coro.suspend into its own block" ); |
2773 | DomSet.insert(Ptr: SuspendBlock->getSingleSuccessor()); |
2774 | } |
2775 | |
2776 | for (Instruction &I : instructions(F)) { |
2777 | AllocaInst* AI = dyn_cast<AllocaInst>(Val: &I); |
2778 | if (!AI) |
2779 | continue; |
2780 | |
2781 | for (BasicBlock *DomBB : DomSet) { |
2782 | bool Valid = true; |
2783 | SmallVector<Instruction *, 1> Lifetimes; |
2784 | |
2785 | auto isLifetimeStart = [](Instruction* I) { |
2786 | if (auto* II = dyn_cast<IntrinsicInst>(Val: I)) |
2787 | return II->getIntrinsicID() == Intrinsic::lifetime_start; |
2788 | return false; |
2789 | }; |
2790 | |
2791 | auto collectLifetimeStart = [&](Instruction *U, AllocaInst *AI) { |
2792 | if (isLifetimeStart(U)) { |
2793 | Lifetimes.push_back(Elt: U); |
2794 | return true; |
2795 | } |
2796 | if (!U->hasOneUse() || U->stripPointerCasts() != AI) |
2797 | return false; |
2798 | if (isLifetimeStart(U->user_back())) { |
2799 | Lifetimes.push_back(Elt: U->user_back()); |
2800 | return true; |
2801 | } |
2802 | return false; |
2803 | }; |
2804 | |
2805 | for (User *U : AI->users()) { |
2806 | Instruction *UI = cast<Instruction>(Val: U); |
2807 | // For all users except lifetime.start markers, if they are all |
2808 | // dominated by one of the basic blocks and do not cross |
2809 | // suspend points as well, then there is no need to spill the |
2810 | // instruction. |
2811 | if (!DT.dominates(A: DomBB, B: UI->getParent()) || |
2812 | Checker.isDefinitionAcrossSuspend(DefBB: DomBB, U: UI)) { |
2813 | // Skip lifetime.start, GEP and bitcast used by lifetime.start |
2814 | // markers. |
2815 | if (collectLifetimeStart(UI, AI)) |
2816 | continue; |
2817 | Valid = false; |
2818 | break; |
2819 | } |
2820 | } |
2821 | // Sink lifetime.start markers to dominate block when they are |
2822 | // only used outside the region. |
2823 | if (Valid && Lifetimes.size() != 0) { |
2824 | auto *NewLifetime = Lifetimes[0]->clone(); |
2825 | NewLifetime->replaceUsesOfWith(From: NewLifetime->getOperand(i: 1), To: AI); |
2826 | NewLifetime->insertBefore(InsertPos: DomBB->getTerminator()); |
2827 | |
2828 | // All the outsided lifetime.start markers are no longer necessary. |
2829 | for (Instruction *S : Lifetimes) |
2830 | S->eraseFromParent(); |
2831 | |
2832 | break; |
2833 | } |
2834 | } |
2835 | } |
2836 | } |
2837 | |
2838 | static void collectFrameAlloca(AllocaInst *AI, coro::Shape &Shape, |
2839 | const SuspendCrossingInfo &Checker, |
2840 | SmallVectorImpl<AllocaInfo> &Allocas, |
2841 | const DominatorTree &DT) { |
2842 | if (Shape.CoroSuspends.empty()) |
2843 | return; |
2844 | |
2845 | // The PromiseAlloca will be specially handled since it needs to be in a |
2846 | // fixed position in the frame. |
2847 | if (AI == Shape.SwitchLowering.PromiseAlloca) |
2848 | return; |
2849 | |
2850 | // The __coro_gro alloca should outlive the promise, make sure we |
2851 | // keep it outside the frame. |
2852 | if (AI->hasMetadata(KindID: LLVMContext::MD_coro_outside_frame)) |
2853 | return; |
2854 | |
2855 | // The code that uses lifetime.start intrinsic does not work for functions |
2856 | // with loops without exit. Disable it on ABIs we know to generate such |
2857 | // code. |
2858 | bool ShouldUseLifetimeStartInfo = |
2859 | (Shape.ABI != coro::ABI::Async && Shape.ABI != coro::ABI::Retcon && |
2860 | Shape.ABI != coro::ABI::RetconOnce); |
2861 | AllocaUseVisitor Visitor{AI->getDataLayout(), DT, Shape, Checker, |
2862 | ShouldUseLifetimeStartInfo}; |
2863 | Visitor.visitPtr(I&: *AI); |
2864 | if (!Visitor.getShouldLiveOnFrame()) |
2865 | return; |
2866 | Allocas.emplace_back(Args&: AI, Args: Visitor.getAliasesCopy(), |
2867 | Args: Visitor.getMayWriteBeforeCoroBegin()); |
2868 | } |
2869 | |
2870 | static std::optional<std::pair<Value &, DIExpression &>> |
2871 | salvageDebugInfoImpl(SmallDenseMap<Argument *, AllocaInst *, 4> &ArgToAllocaMap, |
2872 | bool OptimizeFrame, bool UseEntryValue, Function *F, |
2873 | Value *Storage, DIExpression *Expr, |
2874 | bool SkipOutermostLoad) { |
2875 | IRBuilder<> Builder(F->getContext()); |
2876 | auto InsertPt = F->getEntryBlock().getFirstInsertionPt(); |
2877 | while (isa<IntrinsicInst>(Val: InsertPt)) |
2878 | ++InsertPt; |
2879 | Builder.SetInsertPoint(TheBB: &F->getEntryBlock(), IP: InsertPt); |
2880 | |
2881 | while (auto *Inst = dyn_cast_or_null<Instruction>(Val: Storage)) { |
2882 | if (auto *LdInst = dyn_cast<LoadInst>(Val: Inst)) { |
2883 | Storage = LdInst->getPointerOperand(); |
2884 | // FIXME: This is a heuristic that works around the fact that |
2885 | // LLVM IR debug intrinsics cannot yet distinguish between |
2886 | // memory and value locations: Because a dbg.declare(alloca) is |
2887 | // implicitly a memory location no DW_OP_deref operation for the |
2888 | // last direct load from an alloca is necessary. This condition |
2889 | // effectively drops the *last* DW_OP_deref in the expression. |
2890 | if (!SkipOutermostLoad) |
2891 | Expr = DIExpression::prepend(Expr, Flags: DIExpression::DerefBefore); |
2892 | } else if (auto *StInst = dyn_cast<StoreInst>(Val: Inst)) { |
2893 | Storage = StInst->getValueOperand(); |
2894 | } else { |
2895 | SmallVector<uint64_t, 16> Ops; |
2896 | SmallVector<Value *, 0> AdditionalValues; |
2897 | Value *Op = llvm::salvageDebugInfoImpl( |
2898 | I&: *Inst, CurrentLocOps: Expr ? Expr->getNumLocationOperands() : 0, Ops, |
2899 | AdditionalValues); |
2900 | if (!Op || !AdditionalValues.empty()) { |
2901 | // If salvaging failed or salvaging produced more than one location |
2902 | // operand, give up. |
2903 | break; |
2904 | } |
2905 | Storage = Op; |
2906 | Expr = DIExpression::appendOpsToArg(Expr, Ops, ArgNo: 0, /*StackValue*/ false); |
2907 | } |
2908 | SkipOutermostLoad = false; |
2909 | } |
2910 | if (!Storage) |
2911 | return std::nullopt; |
2912 | |
2913 | auto *StorageAsArg = dyn_cast<Argument>(Val: Storage); |
2914 | const bool IsSwiftAsyncArg = |
2915 | StorageAsArg && StorageAsArg->hasAttribute(Kind: Attribute::SwiftAsync); |
2916 | |
2917 | // Swift async arguments are described by an entry value of the ABI-defined |
2918 | // register containing the coroutine context. |
2919 | // Entry values in variadic expressions are not supported. |
2920 | if (IsSwiftAsyncArg && UseEntryValue && !Expr->isEntryValue() && |
2921 | Expr->isSingleLocationExpression()) |
2922 | Expr = DIExpression::prepend(Expr, Flags: DIExpression::EntryValue); |
2923 | |
2924 | // If the coroutine frame is an Argument, store it in an alloca to improve |
2925 | // its availability (e.g. registers may be clobbered). |
2926 | // Avoid this if optimizations are enabled (they would remove the alloca) or |
2927 | // if the value is guaranteed to be available through other means (e.g. swift |
2928 | // ABI guarantees). |
2929 | if (StorageAsArg && !OptimizeFrame && !IsSwiftAsyncArg) { |
2930 | auto &Cached = ArgToAllocaMap[StorageAsArg]; |
2931 | if (!Cached) { |
2932 | Cached = Builder.CreateAlloca(Ty: Storage->getType(), AddrSpace: 0, ArraySize: nullptr, |
2933 | Name: Storage->getName() + ".debug" ); |
2934 | Builder.CreateStore(Val: Storage, Ptr: Cached); |
2935 | } |
2936 | Storage = Cached; |
2937 | // FIXME: LLVM lacks nuanced semantics to differentiate between |
2938 | // memory and direct locations at the IR level. The backend will |
2939 | // turn a dbg.declare(alloca, ..., DIExpression()) into a memory |
2940 | // location. Thus, if there are deref and offset operations in the |
2941 | // expression, we need to add a DW_OP_deref at the *start* of the |
2942 | // expression to first load the contents of the alloca before |
2943 | // adjusting it with the expression. |
2944 | Expr = DIExpression::prepend(Expr, Flags: DIExpression::DerefBefore); |
2945 | } |
2946 | |
2947 | return {{*Storage, *Expr}}; |
2948 | } |
2949 | |
2950 | void coro::salvageDebugInfo( |
2951 | SmallDenseMap<Argument *, AllocaInst *, 4> &ArgToAllocaMap, |
2952 | DbgVariableIntrinsic &DVI, bool OptimizeFrame, bool UseEntryValue) { |
2953 | |
2954 | Function *F = DVI.getFunction(); |
2955 | // Follow the pointer arithmetic all the way to the incoming |
2956 | // function argument and convert into a DIExpression. |
2957 | bool SkipOutermostLoad = !isa<DbgValueInst>(Val: DVI); |
2958 | Value *OriginalStorage = DVI.getVariableLocationOp(OpIdx: 0); |
2959 | |
2960 | auto SalvagedInfo = ::salvageDebugInfoImpl( |
2961 | ArgToAllocaMap, OptimizeFrame, UseEntryValue, F, Storage: OriginalStorage, |
2962 | Expr: DVI.getExpression(), SkipOutermostLoad); |
2963 | if (!SalvagedInfo) |
2964 | return; |
2965 | |
2966 | Value *Storage = &SalvagedInfo->first; |
2967 | DIExpression *Expr = &SalvagedInfo->second; |
2968 | |
2969 | DVI.replaceVariableLocationOp(OldValue: OriginalStorage, NewValue: Storage); |
2970 | DVI.setExpression(Expr); |
2971 | // We only hoist dbg.declare today since it doesn't make sense to hoist |
2972 | // dbg.value since it does not have the same function wide guarantees that |
2973 | // dbg.declare does. |
2974 | if (isa<DbgDeclareInst>(Val: DVI)) { |
2975 | std::optional<BasicBlock::iterator> InsertPt; |
2976 | if (auto *I = dyn_cast<Instruction>(Val: Storage)) { |
2977 | InsertPt = I->getInsertionPointAfterDef(); |
2978 | // Update DILocation only if variable was not inlined. |
2979 | DebugLoc ILoc = I->getDebugLoc(); |
2980 | DebugLoc DVILoc = DVI.getDebugLoc(); |
2981 | if (ILoc && DVILoc && |
2982 | DVILoc->getScope()->getSubprogram() == |
2983 | ILoc->getScope()->getSubprogram()) |
2984 | DVI.setDebugLoc(I->getDebugLoc()); |
2985 | } else if (isa<Argument>(Val: Storage)) |
2986 | InsertPt = F->getEntryBlock().begin(); |
2987 | if (InsertPt) |
2988 | DVI.moveBefore(BB&: *(*InsertPt)->getParent(), I: *InsertPt); |
2989 | } |
2990 | } |
2991 | |
2992 | void coro::salvageDebugInfo( |
2993 | SmallDenseMap<Argument *, AllocaInst *, 4> &ArgToAllocaMap, |
2994 | DbgVariableRecord &DVR, bool OptimizeFrame, bool UseEntryValue) { |
2995 | |
2996 | Function *F = DVR.getFunction(); |
2997 | // Follow the pointer arithmetic all the way to the incoming |
2998 | // function argument and convert into a DIExpression. |
2999 | bool SkipOutermostLoad = DVR.isDbgDeclare(); |
3000 | Value *OriginalStorage = DVR.getVariableLocationOp(OpIdx: 0); |
3001 | |
3002 | auto SalvagedInfo = ::salvageDebugInfoImpl( |
3003 | ArgToAllocaMap, OptimizeFrame, UseEntryValue, F, Storage: OriginalStorage, |
3004 | Expr: DVR.getExpression(), SkipOutermostLoad); |
3005 | if (!SalvagedInfo) |
3006 | return; |
3007 | |
3008 | Value *Storage = &SalvagedInfo->first; |
3009 | DIExpression *Expr = &SalvagedInfo->second; |
3010 | |
3011 | DVR.replaceVariableLocationOp(OldValue: OriginalStorage, NewValue: Storage); |
3012 | DVR.setExpression(Expr); |
3013 | // We only hoist dbg.declare today since it doesn't make sense to hoist |
3014 | // dbg.value since it does not have the same function wide guarantees that |
3015 | // dbg.declare does. |
3016 | if (DVR.getType() == DbgVariableRecord::LocationType::Declare) { |
3017 | std::optional<BasicBlock::iterator> InsertPt; |
3018 | if (auto *I = dyn_cast<Instruction>(Val: Storage)) { |
3019 | InsertPt = I->getInsertionPointAfterDef(); |
3020 | // Update DILocation only if variable was not inlined. |
3021 | DebugLoc ILoc = I->getDebugLoc(); |
3022 | DebugLoc DVRLoc = DVR.getDebugLoc(); |
3023 | if (ILoc && DVRLoc && |
3024 | DVRLoc->getScope()->getSubprogram() == |
3025 | ILoc->getScope()->getSubprogram()) |
3026 | DVR.setDebugLoc(ILoc); |
3027 | } else if (isa<Argument>(Val: Storage)) |
3028 | InsertPt = F->getEntryBlock().begin(); |
3029 | if (InsertPt) { |
3030 | DVR.removeFromParent(); |
3031 | (*InsertPt)->getParent()->insertDbgRecordBefore(DR: &DVR, Here: *InsertPt); |
3032 | } |
3033 | } |
3034 | } |
3035 | |
3036 | static void doRematerializations( |
3037 | Function &F, SuspendCrossingInfo &Checker, |
3038 | const std::function<bool(Instruction &)> &MaterializableCallback) { |
3039 | if (F.hasOptNone()) |
3040 | return; |
3041 | |
3042 | SpillInfo Spills; |
3043 | |
3044 | // See if there are materializable instructions across suspend points |
3045 | // We record these as the starting point to also identify materializable |
3046 | // defs of uses in these operations |
3047 | for (Instruction &I : instructions(F)) { |
3048 | if (!MaterializableCallback(I)) |
3049 | continue; |
3050 | for (User *U : I.users()) |
3051 | if (Checker.isDefinitionAcrossSuspend(I, U)) |
3052 | Spills[&I].push_back(Elt: cast<Instruction>(Val: U)); |
3053 | } |
3054 | |
3055 | // Process each of the identified rematerializable instructions |
3056 | // and add predecessor instructions that can also be rematerialized. |
3057 | // This is actually a graph of instructions since we could potentially |
3058 | // have multiple uses of a def in the set of predecessor instructions. |
3059 | // The approach here is to maintain a graph of instructions for each bottom |
3060 | // level instruction - where we have a unique set of instructions (nodes) |
3061 | // and edges between them. We then walk the graph in reverse post-dominator |
3062 | // order to insert them past the suspend point, but ensure that ordering is |
3063 | // correct. We also rely on CSE removing duplicate defs for remats of |
3064 | // different instructions with a def in common (rather than maintaining more |
3065 | // complex graphs for each suspend point) |
3066 | |
3067 | // We can do this by adding new nodes to the list for each suspend |
3068 | // point. Then using standard GraphTraits to give a reverse post-order |
3069 | // traversal when we insert the nodes after the suspend |
3070 | SmallMapVector<Instruction *, std::unique_ptr<RematGraph>, 8> AllRemats; |
3071 | for (auto &E : Spills) { |
3072 | for (Instruction *U : E.second) { |
3073 | // Don't process a user twice (this can happen if the instruction uses |
3074 | // more than one rematerializable def) |
3075 | if (AllRemats.count(Key: U)) |
3076 | continue; |
3077 | |
3078 | // Constructor creates the whole RematGraph for the given Use |
3079 | auto RematUPtr = |
3080 | std::make_unique<RematGraph>(args: MaterializableCallback, args&: U, args&: Checker); |
3081 | |
3082 | LLVM_DEBUG(dbgs() << "***** Next remat group *****\n" ; |
3083 | ReversePostOrderTraversal<RematGraph *> RPOT(RematUPtr.get()); |
3084 | for (auto I = RPOT.begin(); I != RPOT.end(); |
3085 | ++I) { (*I)->Node->dump(); } dbgs() |
3086 | << "\n" ;); |
3087 | |
3088 | AllRemats[U] = std::move(RematUPtr); |
3089 | } |
3090 | } |
3091 | |
3092 | // Rewrite materializable instructions to be materialized at the use |
3093 | // point. |
3094 | LLVM_DEBUG(dumpRemats("Materializations" , AllRemats)); |
3095 | rewriteMaterializableInstructions(AllRemats); |
3096 | } |
3097 | |
3098 | void coro::buildCoroutineFrame( |
3099 | Function &F, Shape &Shape, TargetTransformInfo &TTI, |
3100 | const std::function<bool(Instruction &)> &MaterializableCallback) { |
3101 | // Don't eliminate swifterror in async functions that won't be split. |
3102 | if (Shape.ABI != coro::ABI::Async || !Shape.CoroSuspends.empty()) |
3103 | eliminateSwiftError(F, Shape); |
3104 | |
3105 | if (Shape.ABI == coro::ABI::Switch && |
3106 | Shape.SwitchLowering.PromiseAlloca) { |
3107 | Shape.getSwitchCoroId()->clearPromise(); |
3108 | } |
3109 | |
3110 | // Make sure that all coro.save, coro.suspend and the fallthrough coro.end |
3111 | // intrinsics are in their own blocks to simplify the logic of building up |
3112 | // SuspendCrossing data. |
3113 | for (auto *CSI : Shape.CoroSuspends) { |
3114 | if (auto *Save = CSI->getCoroSave()) |
3115 | splitAround(I: Save, Name: "CoroSave" ); |
3116 | splitAround(I: CSI, Name: "CoroSuspend" ); |
3117 | } |
3118 | |
3119 | // Put CoroEnds into their own blocks. |
3120 | for (AnyCoroEndInst *CE : Shape.CoroEnds) { |
3121 | splitAround(I: CE, Name: "CoroEnd" ); |
3122 | |
3123 | // Emit the musttail call function in a new block before the CoroEnd. |
3124 | // We do this here so that the right suspend crossing info is computed for |
3125 | // the uses of the musttail call function call. (Arguments to the coro.end |
3126 | // instructions would be ignored) |
3127 | if (auto *AsyncEnd = dyn_cast<CoroAsyncEndInst>(Val: CE)) { |
3128 | auto *MustTailCallFn = AsyncEnd->getMustTailCallFunction(); |
3129 | if (!MustTailCallFn) |
3130 | continue; |
3131 | IRBuilder<> Builder(AsyncEnd); |
3132 | SmallVector<Value *, 8> Args(AsyncEnd->args()); |
3133 | auto Arguments = ArrayRef<Value *>(Args).drop_front(N: 3); |
3134 | auto *Call = createMustTailCall(Loc: AsyncEnd->getDebugLoc(), MustTailCallFn, |
3135 | TTI, Arguments, Builder); |
3136 | splitAround(I: Call, Name: "MustTailCall.Before.CoroEnd" ); |
3137 | } |
3138 | } |
3139 | |
3140 | // Later code makes structural assumptions about single predecessors phis e.g |
3141 | // that they are not live across a suspend point. |
3142 | cleanupSinglePredPHIs(F); |
3143 | |
3144 | // Transforms multi-edge PHI Nodes, so that any value feeding into a PHI will |
3145 | // never has its definition separated from the PHI by the suspend point. |
3146 | rewritePHIs(F); |
3147 | |
3148 | // Build suspend crossing info. |
3149 | SuspendCrossingInfo Checker(F, Shape); |
3150 | |
3151 | doRematerializations(F, Checker, MaterializableCallback); |
3152 | |
3153 | const DominatorTree DT(F); |
3154 | FrameDataInfo FrameData; |
3155 | SmallVector<CoroAllocaAllocInst*, 4> LocalAllocas; |
3156 | SmallVector<Instruction*, 4> DeadInstructions; |
3157 | if (Shape.ABI != coro::ABI::Async && Shape.ABI != coro::ABI::Retcon && |
3158 | Shape.ABI != coro::ABI::RetconOnce) |
3159 | sinkLifetimeStartMarkers(F, Shape, Checker, DT); |
3160 | |
3161 | // Collect the spills for arguments and other not-materializable values. |
3162 | for (Argument &A : F.args()) |
3163 | for (User *U : A.users()) |
3164 | if (Checker.isDefinitionAcrossSuspend(A, U)) |
3165 | FrameData.Spills[&A].push_back(Elt: cast<Instruction>(Val: U)); |
3166 | |
3167 | for (Instruction &I : instructions(F)) { |
3168 | // Values returned from coroutine structure intrinsics should not be part |
3169 | // of the Coroutine Frame. |
3170 | if (isCoroutineStructureIntrinsic(I) || &I == Shape.CoroBegin) |
3171 | continue; |
3172 | |
3173 | // Handle alloca.alloc specially here. |
3174 | if (auto AI = dyn_cast<CoroAllocaAllocInst>(Val: &I)) { |
3175 | // Check whether the alloca's lifetime is bounded by suspend points. |
3176 | if (isLocalAlloca(AI)) { |
3177 | LocalAllocas.push_back(Elt: AI); |
3178 | continue; |
3179 | } |
3180 | |
3181 | // If not, do a quick rewrite of the alloca and then add spills of |
3182 | // the rewritten value. The rewrite doesn't invalidate anything in |
3183 | // Spills because the other alloca intrinsics have no other operands |
3184 | // besides AI, and it doesn't invalidate the iteration because we delay |
3185 | // erasing AI. |
3186 | auto Alloc = lowerNonLocalAlloca(AI, Shape, DeadInsts&: DeadInstructions); |
3187 | |
3188 | for (User *U : Alloc->users()) { |
3189 | if (Checker.isDefinitionAcrossSuspend(I&: *Alloc, U)) |
3190 | FrameData.Spills[Alloc].push_back(Elt: cast<Instruction>(Val: U)); |
3191 | } |
3192 | continue; |
3193 | } |
3194 | |
3195 | // Ignore alloca.get; we process this as part of coro.alloca.alloc. |
3196 | if (isa<CoroAllocaGetInst>(Val: I)) |
3197 | continue; |
3198 | |
3199 | if (auto *AI = dyn_cast<AllocaInst>(Val: &I)) { |
3200 | collectFrameAlloca(AI, Shape, Checker, Allocas&: FrameData.Allocas, DT); |
3201 | continue; |
3202 | } |
3203 | |
3204 | for (User *U : I.users()) |
3205 | if (Checker.isDefinitionAcrossSuspend(I, U)) { |
3206 | // We cannot spill a token. |
3207 | if (I.getType()->isTokenTy()) |
3208 | report_fatal_error( |
3209 | reason: "token definition is separated from the use by a suspend point" ); |
3210 | FrameData.Spills[&I].push_back(Elt: cast<Instruction>(Val: U)); |
3211 | } |
3212 | } |
3213 | |
3214 | LLVM_DEBUG(dumpAllocas(FrameData.Allocas)); |
3215 | |
3216 | // We don't want the layout of coroutine frame to be affected |
3217 | // by debug information. So we only choose to salvage DbgValueInst for |
3218 | // whose value is already in the frame. |
3219 | // We would handle the dbg.values for allocas specially |
3220 | for (auto &Iter : FrameData.Spills) { |
3221 | auto *V = Iter.first; |
3222 | SmallVector<DbgValueInst *, 16> DVIs; |
3223 | SmallVector<DbgVariableRecord *, 16> DVRs; |
3224 | findDbgValues(DbgValues&: DVIs, V, DbgVariableRecords: &DVRs); |
3225 | for (DbgValueInst *DVI : DVIs) |
3226 | if (Checker.isDefinitionAcrossSuspend(V&: *V, U: DVI)) |
3227 | FrameData.Spills[V].push_back(Elt: DVI); |
3228 | // Add the instructions which carry debug info that is in the frame. |
3229 | for (DbgVariableRecord *DVR : DVRs) |
3230 | if (Checker.isDefinitionAcrossSuspend(V&: *V, U: DVR->Marker->MarkedInstr)) |
3231 | FrameData.Spills[V].push_back(Elt: DVR->Marker->MarkedInstr); |
3232 | } |
3233 | |
3234 | LLVM_DEBUG(dumpSpills("Spills" , FrameData.Spills)); |
3235 | if (Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce || |
3236 | Shape.ABI == coro::ABI::Async) |
3237 | sinkSpillUsesAfterCoroBegin(F, FrameData, CoroBegin: Shape.CoroBegin); |
3238 | Shape.FrameTy = buildFrameType(F, Shape, FrameData); |
3239 | Shape.FramePtr = Shape.CoroBegin; |
3240 | // For now, this works for C++ programs only. |
3241 | buildFrameDebugInfo(F, Shape, FrameData); |
3242 | insertSpills(FrameData, Shape); |
3243 | lowerLocalAllocas(LocalAllocas, DeadInsts&: DeadInstructions); |
3244 | |
3245 | for (auto *I : DeadInstructions) |
3246 | I->eraseFromParent(); |
3247 | } |
3248 | |