1 | //===-- PredicateInfo.cpp - PredicateInfo Builder--------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------===// |
8 | // |
9 | // This file implements the PredicateInfo class. |
10 | // |
11 | //===----------------------------------------------------------------===// |
12 | |
13 | #include "llvm/Transforms/Utils/PredicateInfo.h" |
14 | #include "llvm/ADT/DenseMap.h" |
15 | #include "llvm/ADT/DepthFirstIterator.h" |
16 | #include "llvm/ADT/STLExtras.h" |
17 | #include "llvm/ADT/SmallPtrSet.h" |
18 | #include "llvm/Analysis/AssumptionCache.h" |
19 | #include "llvm/IR/AssemblyAnnotationWriter.h" |
20 | #include "llvm/IR/Dominators.h" |
21 | #include "llvm/IR/IRBuilder.h" |
22 | #include "llvm/IR/InstIterator.h" |
23 | #include "llvm/IR/IntrinsicInst.h" |
24 | #include "llvm/IR/Module.h" |
25 | #include "llvm/IR/PatternMatch.h" |
26 | #include "llvm/Support/CommandLine.h" |
27 | #include "llvm/Support/Debug.h" |
28 | #include "llvm/Support/DebugCounter.h" |
29 | #include "llvm/Support/FormattedStream.h" |
30 | #define DEBUG_TYPE "predicateinfo" |
31 | using namespace llvm; |
32 | using namespace PatternMatch; |
33 | |
34 | static cl::opt<bool> VerifyPredicateInfo( |
35 | "verify-predicateinfo" , cl::init(Val: false), cl::Hidden, |
36 | cl::desc("Verify PredicateInfo in legacy printer pass." )); |
37 | DEBUG_COUNTER(RenameCounter, "predicateinfo-rename" , |
38 | "Controls which variables are renamed with predicateinfo" ); |
39 | |
40 | // Maximum number of conditions considered for renaming for each branch/assume. |
41 | // This limits renaming of deep and/or chains. |
42 | static const unsigned MaxCondsPerBranch = 8; |
43 | |
44 | namespace { |
45 | // Given a predicate info that is a type of branching terminator, get the |
46 | // branching block. |
47 | const BasicBlock *getBranchBlock(const PredicateBase *PB) { |
48 | assert(isa<PredicateWithEdge>(PB) && |
49 | "Only branches and switches should have PHIOnly defs that " |
50 | "require branch blocks." ); |
51 | return cast<PredicateWithEdge>(Val: PB)->From; |
52 | } |
53 | |
54 | // Given a predicate info that is a type of branching terminator, get the |
55 | // branching terminator. |
56 | static Instruction *getBranchTerminator(const PredicateBase *PB) { |
57 | assert(isa<PredicateWithEdge>(PB) && |
58 | "Not a predicate info type we know how to get a terminator from." ); |
59 | return cast<PredicateWithEdge>(Val: PB)->From->getTerminator(); |
60 | } |
61 | |
62 | // Given a predicate info that is a type of branching terminator, get the |
63 | // edge this predicate info represents |
64 | std::pair<BasicBlock *, BasicBlock *> getBlockEdge(const PredicateBase *PB) { |
65 | assert(isa<PredicateWithEdge>(PB) && |
66 | "Not a predicate info type we know how to get an edge from." ); |
67 | const auto *PEdge = cast<PredicateWithEdge>(Val: PB); |
68 | return std::make_pair(x: PEdge->From, y: PEdge->To); |
69 | } |
70 | } |
71 | |
72 | namespace llvm { |
73 | enum LocalNum { |
74 | // Operations that must appear first in the block. |
75 | LN_First, |
76 | // Operations that are somewhere in the middle of the block, and are sorted on |
77 | // demand. |
78 | LN_Middle, |
79 | // Operations that must appear last in a block, like successor phi node uses. |
80 | LN_Last |
81 | }; |
82 | |
83 | // Associate global and local DFS info with defs (PInfo set) and uses (U set), |
84 | // so we can sort them into a global domination ordering. |
85 | struct ValueDFS { |
86 | int DFSIn = 0; |
87 | int DFSOut = 0; |
88 | unsigned int LocalNum = LN_Middle; |
89 | // Only one of U or PInfo will be set. |
90 | Use *U = nullptr; |
91 | PredicateBase *PInfo = nullptr; |
92 | }; |
93 | |
94 | // This compares ValueDFS structures. Doing so allows us to walk the minimum |
95 | // number of instructions necessary to compute our def/use ordering. |
96 | struct ValueDFS_Compare { |
97 | DominatorTree &DT; |
98 | ValueDFS_Compare(DominatorTree &DT) : DT(DT) {} |
99 | |
100 | bool operator()(const ValueDFS &A, const ValueDFS &B) const { |
101 | if (&A == &B) |
102 | return false; |
103 | |
104 | // Order by block first. |
105 | if (A.DFSIn != B.DFSIn) |
106 | return A.DFSIn < B.DFSIn; |
107 | assert(A.DFSOut == B.DFSOut && |
108 | "Equal DFS-in numbers imply equal out numbers" ); |
109 | |
110 | // Then order by first/middle/last. |
111 | if (A.LocalNum != B.LocalNum) |
112 | return A.LocalNum < B.LocalNum; |
113 | |
114 | // We want to put the def that will get used for a given set of phi uses, |
115 | // before those phi uses. |
116 | // So we sort by edge, then by def. |
117 | // Note that only phi nodes uses and defs can come last. |
118 | if (A.LocalNum == LN_Last) |
119 | return comparePHIRelated(A, B); |
120 | |
121 | // Use block-local ordering for instructions in the middle. |
122 | if (A.LocalNum == LN_Middle) |
123 | return localComesBefore(A, B); |
124 | |
125 | // The order of PredicateInfo definitions at the start of the block does not |
126 | // matter. |
127 | assert(A.LocalNum == LN_First); |
128 | assert(A.PInfo && B.PInfo && "Must be predicate info def" ); |
129 | return false; |
130 | } |
131 | |
132 | // For a phi use, or a non-materialized def, return the edge it represents. |
133 | std::pair<BasicBlock *, BasicBlock *> getBlockEdge(const ValueDFS &VD) const { |
134 | if (VD.U) { |
135 | auto *PHI = cast<PHINode>(Val: VD.U->getUser()); |
136 | return std::make_pair(x: PHI->getIncomingBlock(U: *VD.U), y: PHI->getParent()); |
137 | } |
138 | // This is really a non-materialized def. |
139 | return ::getBlockEdge(PB: VD.PInfo); |
140 | } |
141 | |
142 | // For two phi related values, return the ordering. |
143 | bool comparePHIRelated(const ValueDFS &A, const ValueDFS &B) const { |
144 | BasicBlock *ASrc, *ADest, *BSrc, *BDest; |
145 | std::tie(args&: ASrc, args&: ADest) = getBlockEdge(VD: A); |
146 | std::tie(args&: BSrc, args&: BDest) = getBlockEdge(VD: B); |
147 | |
148 | #ifndef NDEBUG |
149 | // This function should only be used for values in the same BB, check that. |
150 | DomTreeNode *DomASrc = DT.getNode(ASrc); |
151 | DomTreeNode *DomBSrc = DT.getNode(BSrc); |
152 | assert(DomASrc->getDFSNumIn() == (unsigned)A.DFSIn && |
153 | "DFS numbers for A should match the ones of the source block" ); |
154 | assert(DomBSrc->getDFSNumIn() == (unsigned)B.DFSIn && |
155 | "DFS numbers for B should match the ones of the source block" ); |
156 | assert(A.DFSIn == B.DFSIn && "Values must be in the same block" ); |
157 | #endif |
158 | (void)ASrc; |
159 | (void)BSrc; |
160 | |
161 | // Use DFS numbers to compare destination blocks, to guarantee a |
162 | // deterministic order. |
163 | DomTreeNode *DomADest = DT.getNode(BB: ADest); |
164 | DomTreeNode *DomBDest = DT.getNode(BB: BDest); |
165 | unsigned AIn = DomADest->getDFSNumIn(); |
166 | unsigned BIn = DomBDest->getDFSNumIn(); |
167 | bool isAUse = A.U; |
168 | bool isBUse = B.U; |
169 | assert((!A.PInfo || !A.U) && (!B.PInfo || !B.U) && |
170 | "Def and U cannot be set at the same time" ); |
171 | // Now sort by edge destination and then defs before uses. |
172 | return std::tie(args&: AIn, args&: isAUse) < std::tie(args&: BIn, args&: isBUse); |
173 | } |
174 | |
175 | const Instruction *getDefOrUser(const ValueDFS &VD) const { |
176 | if (VD.U) |
177 | return cast<Instruction>(Val: VD.U->getUser()); |
178 | |
179 | // For the purpose of ordering, we pretend the def is right after the |
180 | // assume, because that is where we will insert the info. |
181 | assert(VD.PInfo && "No use, and no predicateinfo should not occur" ); |
182 | assert(isa<PredicateAssume>(VD.PInfo) && |
183 | "Middle of block should only occur for assumes" ); |
184 | return cast<PredicateAssume>(Val: VD.PInfo)->AssumeInst->getNextNode(); |
185 | } |
186 | |
187 | // This performs the necessary local basic block ordering checks to tell |
188 | // whether A comes before B, where both are in the same basic block. |
189 | bool localComesBefore(const ValueDFS &A, const ValueDFS &B) const { |
190 | const Instruction *AInst = getDefOrUser(VD: A); |
191 | const Instruction *BInst = getDefOrUser(VD: B); |
192 | return AInst->comesBefore(Other: BInst); |
193 | } |
194 | }; |
195 | |
196 | class PredicateInfoBuilder { |
197 | // Used to store information about each value we might rename. |
198 | struct ValueInfo { |
199 | SmallVector<PredicateBase *, 4> Infos; |
200 | }; |
201 | |
202 | PredicateInfo &PI; |
203 | Function &F; |
204 | DominatorTree &DT; |
205 | AssumptionCache &AC; |
206 | |
207 | // This stores info about each operand or comparison result we make copies |
208 | // of. The real ValueInfos start at index 1, index 0 is unused so that we |
209 | // can more easily detect invalid indexing. |
210 | SmallVector<ValueInfo, 32> ValueInfos; |
211 | |
212 | // This gives the index into the ValueInfos array for a given Value. Because |
213 | // 0 is not a valid Value Info index, you can use DenseMap::lookup and tell |
214 | // whether it returned a valid result. |
215 | DenseMap<Value *, unsigned int> ValueInfoNums; |
216 | |
217 | BumpPtrAllocator &Allocator; |
218 | |
219 | ValueInfo &getOrCreateValueInfo(Value *); |
220 | const ValueInfo &getValueInfo(Value *) const; |
221 | |
222 | void processAssume(IntrinsicInst *, BasicBlock *, |
223 | SmallVectorImpl<Value *> &OpsToRename); |
224 | void processBranch(BranchInst *, BasicBlock *, |
225 | SmallVectorImpl<Value *> &OpsToRename); |
226 | void processSwitch(SwitchInst *, BasicBlock *, |
227 | SmallVectorImpl<Value *> &OpsToRename); |
228 | void renameUses(SmallVectorImpl<Value *> &OpsToRename); |
229 | void addInfoFor(SmallVectorImpl<Value *> &OpsToRename, Value *Op, |
230 | PredicateBase *PB); |
231 | |
232 | struct StackEntry { |
233 | const ValueDFS *V; |
234 | Value *Def = nullptr; |
235 | |
236 | StackEntry(const ValueDFS *V) : V(V) {} |
237 | }; |
238 | |
239 | using ValueDFSStack = SmallVectorImpl<StackEntry>; |
240 | void convertUsesToDFSOrdered(Value *, SmallVectorImpl<ValueDFS> &); |
241 | Value *materializeStack(unsigned int &, ValueDFSStack &, Value *); |
242 | bool stackIsInScope(const ValueDFSStack &, const ValueDFS &) const; |
243 | void popStackUntilDFSScope(ValueDFSStack &, const ValueDFS &); |
244 | |
245 | public: |
246 | PredicateInfoBuilder(PredicateInfo &PI, Function &F, DominatorTree &DT, |
247 | AssumptionCache &AC, BumpPtrAllocator &Allocator) |
248 | : PI(PI), F(F), DT(DT), AC(AC), Allocator(Allocator) { |
249 | // Push an empty operand info so that we can detect 0 as not finding one |
250 | ValueInfos.resize(N: 1); |
251 | } |
252 | |
253 | void buildPredicateInfo(); |
254 | }; |
255 | |
256 | bool PredicateInfoBuilder::stackIsInScope(const ValueDFSStack &Stack, |
257 | const ValueDFS &VDUse) const { |
258 | assert(!Stack.empty() && "Should not be called with empty stack" ); |
259 | // If it's a phi only use, make sure it's for this phi node edge, and that the |
260 | // use is in a phi node. If it's anything else, and the top of the stack is |
261 | // a LN_Last def, we need to pop the stack. We deliberately sort phi uses |
262 | // next to the defs they must go with so that we can know it's time to pop |
263 | // the stack when we hit the end of the phi uses for a given def. |
264 | const ValueDFS &Top = *Stack.back().V; |
265 | if (Top.LocalNum == LN_Last && Top.PInfo) { |
266 | if (!VDUse.U) |
267 | return false; |
268 | auto *PHI = dyn_cast<PHINode>(Val: VDUse.U->getUser()); |
269 | if (!PHI) |
270 | return false; |
271 | // Check edge |
272 | BasicBlock *EdgePred = PHI->getIncomingBlock(U: *VDUse.U); |
273 | if (EdgePred != getBranchBlock(PB: Top.PInfo)) |
274 | return false; |
275 | |
276 | // Use dominates, which knows how to handle edge dominance. |
277 | return DT.dominates(BBE: getBlockEdge(PB: Top.PInfo), U: *VDUse.U); |
278 | } |
279 | |
280 | return VDUse.DFSIn >= Top.DFSIn && VDUse.DFSOut <= Top.DFSOut; |
281 | } |
282 | |
283 | void PredicateInfoBuilder::popStackUntilDFSScope(ValueDFSStack &Stack, |
284 | const ValueDFS &VD) { |
285 | while (!Stack.empty() && !stackIsInScope(Stack, VDUse: VD)) |
286 | Stack.pop_back(); |
287 | } |
288 | |
289 | // Convert the uses of Op into a vector of uses, associating global and local |
290 | // DFS info with each one. |
291 | void PredicateInfoBuilder::convertUsesToDFSOrdered( |
292 | Value *Op, SmallVectorImpl<ValueDFS> &DFSOrderedSet) { |
293 | for (auto &U : Op->uses()) { |
294 | if (auto *I = dyn_cast<Instruction>(Val: U.getUser())) { |
295 | ValueDFS VD; |
296 | // Put the phi node uses in the incoming block. |
297 | BasicBlock *IBlock; |
298 | if (auto *PN = dyn_cast<PHINode>(Val: I)) { |
299 | IBlock = PN->getIncomingBlock(U); |
300 | // Make phi node users appear last in the incoming block |
301 | // they are from. |
302 | VD.LocalNum = LN_Last; |
303 | } else { |
304 | // If it's not a phi node use, it is somewhere in the middle of the |
305 | // block. |
306 | IBlock = I->getParent(); |
307 | VD.LocalNum = LN_Middle; |
308 | } |
309 | DomTreeNode *DomNode = DT.getNode(BB: IBlock); |
310 | // It's possible our use is in an unreachable block. Skip it if so. |
311 | if (!DomNode) |
312 | continue; |
313 | VD.DFSIn = DomNode->getDFSNumIn(); |
314 | VD.DFSOut = DomNode->getDFSNumOut(); |
315 | VD.U = &U; |
316 | DFSOrderedSet.push_back(Elt: VD); |
317 | } |
318 | } |
319 | } |
320 | |
321 | bool shouldRename(Value *V) { |
322 | // Only want real values, not constants. Additionally, operands with one use |
323 | // are only being used in the comparison, which means they will not be useful |
324 | // for us to consider for predicateinfo. |
325 | return (isa<Instruction>(Val: V) || isa<Argument>(Val: V)) && !V->hasOneUse(); |
326 | } |
327 | |
328 | // Collect relevant operations from Comparison that we may want to insert copies |
329 | // for. |
330 | void collectCmpOps(CmpInst *Comparison, SmallVectorImpl<Value *> &CmpOperands) { |
331 | auto *Op0 = Comparison->getOperand(i_nocapture: 0); |
332 | auto *Op1 = Comparison->getOperand(i_nocapture: 1); |
333 | if (Op0 == Op1) |
334 | return; |
335 | |
336 | CmpOperands.push_back(Elt: Op0); |
337 | CmpOperands.push_back(Elt: Op1); |
338 | } |
339 | |
340 | // Add Op, PB to the list of value infos for Op, and mark Op to be renamed. |
341 | void PredicateInfoBuilder::addInfoFor(SmallVectorImpl<Value *> &OpsToRename, |
342 | Value *Op, PredicateBase *PB) { |
343 | auto &OperandInfo = getOrCreateValueInfo(Op); |
344 | if (OperandInfo.Infos.empty()) |
345 | OpsToRename.push_back(Elt: Op); |
346 | OperandInfo.Infos.push_back(Elt: PB); |
347 | } |
348 | |
349 | // Process an assume instruction and place relevant operations we want to rename |
350 | // into OpsToRename. |
351 | void PredicateInfoBuilder::processAssume( |
352 | IntrinsicInst *II, BasicBlock *AssumeBB, |
353 | SmallVectorImpl<Value *> &OpsToRename) { |
354 | SmallVector<Value *, 4> Worklist; |
355 | SmallPtrSet<Value *, 4> Visited; |
356 | Worklist.push_back(Elt: II->getOperand(i_nocapture: 0)); |
357 | while (!Worklist.empty()) { |
358 | Value *Cond = Worklist.pop_back_val(); |
359 | if (!Visited.insert(Ptr: Cond).second) |
360 | continue; |
361 | if (Visited.size() > MaxCondsPerBranch) |
362 | break; |
363 | |
364 | Value *Op0, *Op1; |
365 | if (match(V: Cond, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
366 | Worklist.push_back(Elt: Op1); |
367 | Worklist.push_back(Elt: Op0); |
368 | } |
369 | |
370 | SmallVector<Value *, 4> Values; |
371 | Values.push_back(Elt: Cond); |
372 | if (auto *Cmp = dyn_cast<CmpInst>(Val: Cond)) |
373 | collectCmpOps(Comparison: Cmp, CmpOperands&: Values); |
374 | |
375 | for (Value *V : Values) { |
376 | if (shouldRename(V)) { |
377 | auto *PA = new (Allocator) PredicateAssume(V, II, Cond); |
378 | addInfoFor(OpsToRename, Op: V, PB: PA); |
379 | } |
380 | } |
381 | } |
382 | } |
383 | |
384 | // Process a block terminating branch, and place relevant operations to be |
385 | // renamed into OpsToRename. |
386 | void PredicateInfoBuilder::processBranch( |
387 | BranchInst *BI, BasicBlock *BranchBB, |
388 | SmallVectorImpl<Value *> &OpsToRename) { |
389 | BasicBlock *FirstBB = BI->getSuccessor(i: 0); |
390 | BasicBlock *SecondBB = BI->getSuccessor(i: 1); |
391 | |
392 | for (BasicBlock *Succ : {FirstBB, SecondBB}) { |
393 | bool TakenEdge = Succ == FirstBB; |
394 | // Don't try to insert on a self-edge. This is mainly because we will |
395 | // eliminate during renaming anyway. |
396 | if (Succ == BranchBB) |
397 | continue; |
398 | |
399 | SmallVector<Value *, 4> Worklist; |
400 | SmallPtrSet<Value *, 4> Visited; |
401 | Worklist.push_back(Elt: BI->getCondition()); |
402 | while (!Worklist.empty()) { |
403 | Value *Cond = Worklist.pop_back_val(); |
404 | if (!Visited.insert(Ptr: Cond).second) |
405 | continue; |
406 | if (Visited.size() > MaxCondsPerBranch) |
407 | break; |
408 | |
409 | Value *Op0, *Op1; |
410 | if (TakenEdge ? match(V: Cond, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1))) |
411 | : match(V: Cond, P: m_LogicalOr(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
412 | Worklist.push_back(Elt: Op1); |
413 | Worklist.push_back(Elt: Op0); |
414 | } |
415 | |
416 | SmallVector<Value *, 4> Values; |
417 | Values.push_back(Elt: Cond); |
418 | if (auto *Cmp = dyn_cast<CmpInst>(Val: Cond)) |
419 | collectCmpOps(Comparison: Cmp, CmpOperands&: Values); |
420 | |
421 | for (Value *V : Values) { |
422 | if (shouldRename(V)) { |
423 | PredicateBase *PB = new (Allocator) |
424 | PredicateBranch(V, BranchBB, Succ, Cond, TakenEdge); |
425 | addInfoFor(OpsToRename, Op: V, PB); |
426 | } |
427 | } |
428 | } |
429 | } |
430 | } |
431 | // Process a block terminating switch, and place relevant operations to be |
432 | // renamed into OpsToRename. |
433 | void PredicateInfoBuilder::processSwitch( |
434 | SwitchInst *SI, BasicBlock *BranchBB, |
435 | SmallVectorImpl<Value *> &OpsToRename) { |
436 | Value *Op = SI->getCondition(); |
437 | if ((!isa<Instruction>(Val: Op) && !isa<Argument>(Val: Op)) || Op->hasOneUse()) |
438 | return; |
439 | |
440 | // Remember how many outgoing edges there are to every successor. |
441 | SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; |
442 | for (BasicBlock *TargetBlock : successors(BB: BranchBB)) |
443 | ++SwitchEdges[TargetBlock]; |
444 | |
445 | // Now propagate info for each case value |
446 | for (auto C : SI->cases()) { |
447 | BasicBlock *TargetBlock = C.getCaseSuccessor(); |
448 | if (SwitchEdges.lookup(Val: TargetBlock) == 1) { |
449 | PredicateSwitch *PS = new (Allocator) PredicateSwitch( |
450 | Op, SI->getParent(), TargetBlock, C.getCaseValue(), SI); |
451 | addInfoFor(OpsToRename, Op, PB: PS); |
452 | } |
453 | } |
454 | } |
455 | |
456 | // Build predicate info for our function |
457 | void PredicateInfoBuilder::buildPredicateInfo() { |
458 | DT.updateDFSNumbers(); |
459 | // Collect operands to rename from all conditional branch terminators, as well |
460 | // as assume statements. |
461 | SmallVector<Value *, 8> OpsToRename; |
462 | for (BasicBlock &BB : F) { |
463 | if (!DT.isReachableFromEntry(A: &BB)) |
464 | continue; |
465 | |
466 | if (auto *BI = dyn_cast<BranchInst>(Val: BB.getTerminator())) { |
467 | if (!BI->isConditional()) |
468 | continue; |
469 | // Can't insert conditional information if they all go to the same place. |
470 | if (BI->getSuccessor(i: 0) == BI->getSuccessor(i: 1)) |
471 | continue; |
472 | processBranch(BI, BranchBB: &BB, OpsToRename); |
473 | } else if (auto *SI = dyn_cast<SwitchInst>(Val: BB.getTerminator())) { |
474 | processSwitch(SI, BranchBB: &BB, OpsToRename); |
475 | } |
476 | } |
477 | for (auto &Assume : AC.assumptions()) { |
478 | if (auto *II = dyn_cast_or_null<IntrinsicInst>(Val&: Assume)) |
479 | if (DT.isReachableFromEntry(A: II->getParent())) |
480 | processAssume(II, AssumeBB: II->getParent(), OpsToRename); |
481 | } |
482 | // Now rename all our operations. |
483 | renameUses(OpsToRename); |
484 | } |
485 | |
486 | // Given the renaming stack, make all the operands currently on the stack real |
487 | // by inserting them into the IR. Return the last operation's value. |
488 | Value *PredicateInfoBuilder::materializeStack(unsigned int &Counter, |
489 | ValueDFSStack &RenameStack, |
490 | Value *OrigOp) { |
491 | // Find the first thing we have to materialize |
492 | auto RevIter = RenameStack.rbegin(); |
493 | for (; RevIter != RenameStack.rend(); ++RevIter) |
494 | if (RevIter->Def) |
495 | break; |
496 | |
497 | size_t Start = RevIter - RenameStack.rbegin(); |
498 | // The maximum number of things we should be trying to materialize at once |
499 | // right now is 4, depending on if we had an assume, a branch, and both used |
500 | // and of conditions. |
501 | for (auto RenameIter = RenameStack.end() - Start; |
502 | RenameIter != RenameStack.end(); ++RenameIter) { |
503 | auto *Op = |
504 | RenameIter == RenameStack.begin() ? OrigOp : (RenameIter - 1)->Def; |
505 | StackEntry &Result = *RenameIter; |
506 | auto *ValInfo = Result.V->PInfo; |
507 | ValInfo->RenamedOp = (RenameStack.end() - Start) == RenameStack.begin() |
508 | ? OrigOp |
509 | : (RenameStack.end() - Start - 1)->Def; |
510 | auto CreateSSACopy = [this](IRBuilderBase &B, Value *Op, |
511 | const Twine &Name = "" ) { |
512 | auto It = PI.DeclarationCache.try_emplace(Key: Op->getType()); |
513 | if (It.second) { |
514 | // The number of named values is used to detect if a new declaration |
515 | // was added. If so, that declaration is tracked so that it can be |
516 | // removed when the analysis is done. The corner case were a new |
517 | // declaration results in a name clash and the old name being renamed |
518 | // is not considered as that represents an invalid module. |
519 | auto NumDecls = F.getParent()->getNumNamedValues(); |
520 | Function *IF = Intrinsic::getOrInsertDeclaration( |
521 | M: F.getParent(), id: Intrinsic::ssa_copy, Tys: Op->getType()); |
522 | if (NumDecls != F.getParent()->getNumNamedValues()) |
523 | PI.CreatedDeclarations.insert(V: IF); |
524 | It.first->second = IF; |
525 | } |
526 | return B.CreateCall(Callee: It.first->second, Args: Op, Name); |
527 | }; |
528 | // For edge predicates, we can just place the operand in the block before |
529 | // the terminator. For assume, we have to place it right after the assume |
530 | // to ensure we dominate all uses except assume itself. Always insert |
531 | // right before the terminator or after the assume, so that we insert in |
532 | // proper order in the case of multiple predicateinfo in the same block. |
533 | if (isa<PredicateWithEdge>(Val: ValInfo)) { |
534 | IRBuilder<> B(getBranchTerminator(PB: ValInfo)); |
535 | CallInst *PIC = |
536 | CreateSSACopy(B, Op, Op->getName() + "." + Twine(Counter++)); |
537 | PI.PredicateMap.insert(KV: {PIC, ValInfo}); |
538 | Result.Def = PIC; |
539 | } else { |
540 | auto *PAssume = dyn_cast<PredicateAssume>(Val: ValInfo); |
541 | assert(PAssume && |
542 | "Should not have gotten here without it being an assume" ); |
543 | // Insert the predicate directly after the assume. While it also holds |
544 | // directly before it, assume(i1 true) is not a useful fact. |
545 | IRBuilder<> B(PAssume->AssumeInst->getNextNode()); |
546 | CallInst *PIC = CreateSSACopy(B, Op); |
547 | PI.PredicateMap.insert(KV: {PIC, ValInfo}); |
548 | Result.Def = PIC; |
549 | } |
550 | } |
551 | return RenameStack.back().Def; |
552 | } |
553 | |
554 | // Instead of the standard SSA renaming algorithm, which is O(Number of |
555 | // instructions), and walks the entire dominator tree, we walk only the defs + |
556 | // uses. The standard SSA renaming algorithm does not really rely on the |
557 | // dominator tree except to order the stack push/pops of the renaming stacks, so |
558 | // that defs end up getting pushed before hitting the correct uses. This does |
559 | // not require the dominator tree, only the *order* of the dominator tree. The |
560 | // complete and correct ordering of the defs and uses, in dominator tree is |
561 | // contained in the DFS numbering of the dominator tree. So we sort the defs and |
562 | // uses into the DFS ordering, and then just use the renaming stack as per |
563 | // normal, pushing when we hit a def (which is a predicateinfo instruction), |
564 | // popping when we are out of the dfs scope for that def, and replacing any uses |
565 | // with top of stack if it exists. In order to handle liveness without |
566 | // propagating liveness info, we don't actually insert the predicateinfo |
567 | // instruction def until we see a use that it would dominate. Once we see such |
568 | // a use, we materialize the predicateinfo instruction in the right place and |
569 | // use it. |
570 | // |
571 | // TODO: Use this algorithm to perform fast single-variable renaming in |
572 | // promotememtoreg and memoryssa. |
573 | void PredicateInfoBuilder::renameUses(SmallVectorImpl<Value *> &OpsToRename) { |
574 | ValueDFS_Compare Compare(DT); |
575 | // Compute liveness, and rename in O(uses) per Op. |
576 | for (auto *Op : OpsToRename) { |
577 | LLVM_DEBUG(dbgs() << "Visiting " << *Op << "\n" ); |
578 | unsigned Counter = 0; |
579 | SmallVector<ValueDFS, 16> OrderedUses; |
580 | const auto &ValueInfo = getValueInfo(Op); |
581 | // Insert the possible copies into the def/use list. |
582 | // They will become real copies if we find a real use for them, and never |
583 | // created otherwise. |
584 | for (const auto &PossibleCopy : ValueInfo.Infos) { |
585 | ValueDFS VD; |
586 | // Determine where we are going to place the copy by the copy type. |
587 | // The predicate info for branches always come first, they will get |
588 | // materialized in the split block at the top of the block. |
589 | // The predicate info for assumes will be somewhere in the middle, |
590 | // it will get materialized right after the assume. |
591 | if (const auto *PAssume = dyn_cast<PredicateAssume>(Val: PossibleCopy)) { |
592 | VD.LocalNum = LN_Middle; |
593 | DomTreeNode *DomNode = DT.getNode(BB: PAssume->AssumeInst->getParent()); |
594 | if (!DomNode) |
595 | continue; |
596 | VD.DFSIn = DomNode->getDFSNumIn(); |
597 | VD.DFSOut = DomNode->getDFSNumOut(); |
598 | VD.PInfo = PossibleCopy; |
599 | OrderedUses.push_back(Elt: VD); |
600 | } else if (isa<PredicateWithEdge>(Val: PossibleCopy)) { |
601 | // If we can only do phi uses, we treat it like it's in the branch |
602 | // block, and handle it specially. We know that it goes last, and only |
603 | // dominate phi uses. |
604 | auto BlockEdge = getBlockEdge(PB: PossibleCopy); |
605 | if (!BlockEdge.second->getSinglePredecessor()) { |
606 | VD.LocalNum = LN_Last; |
607 | auto *DomNode = DT.getNode(BB: BlockEdge.first); |
608 | if (DomNode) { |
609 | VD.DFSIn = DomNode->getDFSNumIn(); |
610 | VD.DFSOut = DomNode->getDFSNumOut(); |
611 | VD.PInfo = PossibleCopy; |
612 | OrderedUses.push_back(Elt: VD); |
613 | } |
614 | } else { |
615 | // Otherwise, we are in the split block (even though we perform |
616 | // insertion in the branch block). |
617 | // Insert a possible copy at the split block and before the branch. |
618 | VD.LocalNum = LN_First; |
619 | auto *DomNode = DT.getNode(BB: BlockEdge.second); |
620 | if (DomNode) { |
621 | VD.DFSIn = DomNode->getDFSNumIn(); |
622 | VD.DFSOut = DomNode->getDFSNumOut(); |
623 | VD.PInfo = PossibleCopy; |
624 | OrderedUses.push_back(Elt: VD); |
625 | } |
626 | } |
627 | } |
628 | } |
629 | |
630 | convertUsesToDFSOrdered(Op, DFSOrderedSet&: OrderedUses); |
631 | // Here we require a stable sort because we do not bother to try to |
632 | // assign an order to the operands the uses represent. Thus, two |
633 | // uses in the same instruction do not have a strict sort order |
634 | // currently and will be considered equal. We could get rid of the |
635 | // stable sort by creating one if we wanted. |
636 | llvm::stable_sort(Range&: OrderedUses, C: Compare); |
637 | SmallVector<StackEntry, 8> RenameStack; |
638 | // For each use, sorted into dfs order, push values and replaces uses with |
639 | // top of stack, which will represent the reaching def. |
640 | for (const ValueDFS &VD : OrderedUses) { |
641 | // We currently do not materialize copy over copy, but we should decide if |
642 | // we want to. |
643 | if (RenameStack.empty()) { |
644 | LLVM_DEBUG(dbgs() << "Rename Stack is empty\n" ); |
645 | } else { |
646 | LLVM_DEBUG(dbgs() << "Rename Stack Top DFS numbers are (" |
647 | << RenameStack.back().V->DFSIn << "," |
648 | << RenameStack.back().V->DFSOut << ")\n" ); |
649 | } |
650 | |
651 | LLVM_DEBUG(dbgs() << "Current DFS numbers are (" << VD.DFSIn << "," |
652 | << VD.DFSOut << ")\n" ); |
653 | |
654 | // Sync to our current scope. |
655 | popStackUntilDFSScope(Stack&: RenameStack, VD); |
656 | |
657 | if (VD.PInfo) { |
658 | RenameStack.push_back(Elt: &VD); |
659 | continue; |
660 | } |
661 | |
662 | // If we get to this point, and the stack is empty we must have a use |
663 | // with no renaming needed, just skip it. |
664 | if (RenameStack.empty()) |
665 | continue; |
666 | if (!DebugCounter::shouldExecute(CounterName: RenameCounter)) { |
667 | LLVM_DEBUG(dbgs() << "Skipping execution due to debug counter\n" ); |
668 | continue; |
669 | } |
670 | StackEntry &Result = RenameStack.back(); |
671 | |
672 | // If the possible copy dominates something, materialize our stack up to |
673 | // this point. This ensures every comparison that affects our operation |
674 | // ends up with predicateinfo. |
675 | if (!Result.Def) |
676 | Result.Def = materializeStack(Counter, RenameStack, OrigOp: Op); |
677 | |
678 | LLVM_DEBUG(dbgs() << "Found replacement " << *Result.Def << " for " |
679 | << *VD.U->get() << " in " << *(VD.U->getUser()) |
680 | << "\n" ); |
681 | assert(DT.dominates(cast<Instruction>(Result.Def), *VD.U) && |
682 | "Predicateinfo def should have dominated this use" ); |
683 | VD.U->set(Result.Def); |
684 | } |
685 | } |
686 | } |
687 | |
688 | PredicateInfoBuilder::ValueInfo & |
689 | PredicateInfoBuilder::getOrCreateValueInfo(Value *Operand) { |
690 | auto Res = ValueInfoNums.try_emplace(Key: Operand, Args: ValueInfos.size()); |
691 | if (Res.second) { |
692 | // Allocate space for new ValueInfo. |
693 | ValueInfos.resize(N: ValueInfos.size() + 1); |
694 | } |
695 | return ValueInfos[Res.first->second]; |
696 | } |
697 | |
698 | const PredicateInfoBuilder::ValueInfo & |
699 | PredicateInfoBuilder::getValueInfo(Value *Operand) const { |
700 | auto OINI = ValueInfoNums.lookup(Val: Operand); |
701 | assert(OINI != 0 && "Operand was not really in the Value Info Numbers" ); |
702 | assert(OINI < ValueInfos.size() && |
703 | "Value Info Number greater than size of Value Info Table" ); |
704 | return ValueInfos[OINI]; |
705 | } |
706 | |
707 | PredicateInfo::PredicateInfo(Function &F, DominatorTree &DT, |
708 | AssumptionCache &AC, BumpPtrAllocator &Allocator) |
709 | : F(F) { |
710 | PredicateInfoBuilder Builder(*this, F, DT, AC, Allocator); |
711 | Builder.buildPredicateInfo(); |
712 | } |
713 | |
714 | // Remove all declarations we created . The PredicateInfo consumers are |
715 | // responsible for remove the ssa_copy calls created. |
716 | PredicateInfo::~PredicateInfo() { |
717 | // Collect function pointers in set first, as SmallSet uses a SmallVector |
718 | // internally and we have to remove the asserting value handles first. |
719 | SmallPtrSet<Function *, 20> FunctionPtrs; |
720 | for (const auto &F : CreatedDeclarations) |
721 | FunctionPtrs.insert(Ptr: &*F); |
722 | CreatedDeclarations.clear(); |
723 | |
724 | for (Function *F : FunctionPtrs) { |
725 | assert(F->user_begin() == F->user_end() && |
726 | "PredicateInfo consumer did not remove all SSA copies." ); |
727 | F->eraseFromParent(); |
728 | } |
729 | } |
730 | |
731 | std::optional<PredicateConstraint> PredicateBase::getConstraint() const { |
732 | switch (Type) { |
733 | case PT_Assume: |
734 | case PT_Branch: { |
735 | bool TrueEdge = true; |
736 | if (auto *PBranch = dyn_cast<PredicateBranch>(Val: this)) |
737 | TrueEdge = PBranch->TrueEdge; |
738 | |
739 | if (Condition == RenamedOp) { |
740 | return {{.Predicate: CmpInst::ICMP_EQ, |
741 | .OtherOp: TrueEdge ? ConstantInt::getTrue(Ty: Condition->getType()) |
742 | : ConstantInt::getFalse(Ty: Condition->getType())}}; |
743 | } |
744 | |
745 | CmpInst *Cmp = dyn_cast<CmpInst>(Val: Condition); |
746 | if (!Cmp) { |
747 | // TODO: Make this an assertion once RenamedOp is fully accurate. |
748 | return std::nullopt; |
749 | } |
750 | |
751 | CmpInst::Predicate Pred; |
752 | Value *OtherOp; |
753 | if (Cmp->getOperand(i_nocapture: 0) == RenamedOp) { |
754 | Pred = Cmp->getPredicate(); |
755 | OtherOp = Cmp->getOperand(i_nocapture: 1); |
756 | } else if (Cmp->getOperand(i_nocapture: 1) == RenamedOp) { |
757 | Pred = Cmp->getSwappedPredicate(); |
758 | OtherOp = Cmp->getOperand(i_nocapture: 0); |
759 | } else { |
760 | // TODO: Make this an assertion once RenamedOp is fully accurate. |
761 | return std::nullopt; |
762 | } |
763 | |
764 | // Invert predicate along false edge. |
765 | if (!TrueEdge) |
766 | Pred = CmpInst::getInversePredicate(pred: Pred); |
767 | |
768 | return {{.Predicate: Pred, .OtherOp: OtherOp}}; |
769 | } |
770 | case PT_Switch: |
771 | if (Condition != RenamedOp) { |
772 | // TODO: Make this an assertion once RenamedOp is fully accurate. |
773 | return std::nullopt; |
774 | } |
775 | |
776 | return {{.Predicate: CmpInst::ICMP_EQ, .OtherOp: cast<PredicateSwitch>(Val: this)->CaseValue}}; |
777 | } |
778 | llvm_unreachable("Unknown predicate type" ); |
779 | } |
780 | |
781 | void PredicateInfo::verifyPredicateInfo() const {} |
782 | |
783 | // Replace ssa_copy calls created by PredicateInfo with their operand. |
784 | static void replaceCreatedSSACopys(PredicateInfo &PredInfo, Function &F) { |
785 | for (Instruction &Inst : llvm::make_early_inc_range(Range: instructions(F))) { |
786 | const auto *PI = PredInfo.getPredicateInfoFor(V: &Inst); |
787 | auto *II = dyn_cast<IntrinsicInst>(Val: &Inst); |
788 | if (!PI || !II || II->getIntrinsicID() != Intrinsic::ssa_copy) |
789 | continue; |
790 | |
791 | Inst.replaceAllUsesWith(V: II->getOperand(i_nocapture: 0)); |
792 | Inst.eraseFromParent(); |
793 | } |
794 | } |
795 | |
796 | PreservedAnalyses PredicateInfoPrinterPass::run(Function &F, |
797 | FunctionAnalysisManager &AM) { |
798 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
799 | auto &AC = AM.getResult<AssumptionAnalysis>(IR&: F); |
800 | OS << "PredicateInfo for function: " << F.getName() << "\n" ; |
801 | BumpPtrAllocator Allocator; |
802 | auto PredInfo = std::make_unique<PredicateInfo>(args&: F, args&: DT, args&: AC, args&: Allocator); |
803 | PredInfo->print(OS); |
804 | |
805 | replaceCreatedSSACopys(PredInfo&: *PredInfo, F); |
806 | return PreservedAnalyses::all(); |
807 | } |
808 | |
809 | /// An assembly annotator class to print PredicateInfo information in |
810 | /// comments. |
811 | class PredicateInfoAnnotatedWriter : public AssemblyAnnotationWriter { |
812 | friend class PredicateInfo; |
813 | const PredicateInfo *PredInfo; |
814 | |
815 | public: |
816 | PredicateInfoAnnotatedWriter(const PredicateInfo *M) : PredInfo(M) {} |
817 | |
818 | void emitBasicBlockStartAnnot(const BasicBlock *BB, |
819 | formatted_raw_ostream &OS) override {} |
820 | |
821 | void emitInstructionAnnot(const Instruction *I, |
822 | formatted_raw_ostream &OS) override { |
823 | if (const auto *PI = PredInfo->getPredicateInfoFor(V: I)) { |
824 | OS << "; Has predicate info\n" ; |
825 | if (const auto *PB = dyn_cast<PredicateBranch>(Val: PI)) { |
826 | OS << "; branch predicate info { TrueEdge: " << PB->TrueEdge |
827 | << " Comparison:" << *PB->Condition << " Edge: [" ; |
828 | PB->From->printAsOperand(O&: OS); |
829 | OS << "," ; |
830 | PB->To->printAsOperand(O&: OS); |
831 | OS << "]" ; |
832 | } else if (const auto *PS = dyn_cast<PredicateSwitch>(Val: PI)) { |
833 | OS << "; switch predicate info { CaseValue: " << *PS->CaseValue |
834 | << " Switch:" << *PS->Switch << " Edge: [" ; |
835 | PS->From->printAsOperand(O&: OS); |
836 | OS << "," ; |
837 | PS->To->printAsOperand(O&: OS); |
838 | OS << "]" ; |
839 | } else if (const auto *PA = dyn_cast<PredicateAssume>(Val: PI)) { |
840 | OS << "; assume predicate info {" |
841 | << " Comparison:" << *PA->Condition; |
842 | } |
843 | OS << ", RenamedOp: " ; |
844 | PI->RenamedOp->printAsOperand(O&: OS, PrintType: false); |
845 | OS << " }\n" ; |
846 | } |
847 | } |
848 | }; |
849 | |
850 | void PredicateInfo::print(raw_ostream &OS) const { |
851 | PredicateInfoAnnotatedWriter Writer(this); |
852 | F.print(OS, AAW: &Writer); |
853 | } |
854 | |
855 | void PredicateInfo::dump() const { |
856 | PredicateInfoAnnotatedWriter Writer(this); |
857 | F.print(OS&: dbgs(), AAW: &Writer); |
858 | } |
859 | |
860 | PreservedAnalyses PredicateInfoVerifierPass::run(Function &F, |
861 | FunctionAnalysisManager &AM) { |
862 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
863 | auto &AC = AM.getResult<AssumptionAnalysis>(IR&: F); |
864 | BumpPtrAllocator Allocator; |
865 | std::make_unique<PredicateInfo>(args&: F, args&: DT, args&: AC, args&: Allocator)->verifyPredicateInfo(); |
866 | |
867 | return PreservedAnalyses::all(); |
868 | } |
869 | } |
870 | |