xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Utils/PredicateInfo.cpp (revision 32a95656b51ebefcdf3e0b02c110825f59abd26f)
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/ADT/Statistic.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/AssumptionCache.h"
21 #include "llvm/Analysis/CFG.h"
22 #include "llvm/IR/AssemblyAnnotationWriter.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/GlobalVariable.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/InstIterator.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/InitializePasses.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/DebugCounter.h"
37 #include "llvm/Support/FormattedStream.h"
38 #include "llvm/Transforms/Utils.h"
39 #include <algorithm>
40 #define DEBUG_TYPE "predicateinfo"
41 using namespace llvm;
42 using namespace PatternMatch;
43 
44 INITIALIZE_PASS_BEGIN(PredicateInfoPrinterLegacyPass, "print-predicateinfo",
45                       "PredicateInfo Printer", false, false)
46 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
47 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
48 INITIALIZE_PASS_END(PredicateInfoPrinterLegacyPass, "print-predicateinfo",
49                     "PredicateInfo Printer", false, false)
50 static cl::opt<bool> VerifyPredicateInfo(
51     "verify-predicateinfo", cl::init(false), cl::Hidden,
52     cl::desc("Verify PredicateInfo in legacy printer pass."));
53 DEBUG_COUNTER(RenameCounter, "predicateinfo-rename",
54               "Controls which variables are renamed with predicateinfo");
55 
56 // Maximum number of conditions considered for renaming for each branch/assume.
57 // This limits renaming of deep and/or chains.
58 static const unsigned MaxCondsPerBranch = 8;
59 
60 namespace {
61 // Given a predicate info that is a type of branching terminator, get the
62 // branching block.
63 const BasicBlock *getBranchBlock(const PredicateBase *PB) {
64   assert(isa<PredicateWithEdge>(PB) &&
65          "Only branches and switches should have PHIOnly defs that "
66          "require branch blocks.");
67   return cast<PredicateWithEdge>(PB)->From;
68 }
69 
70 // Given a predicate info that is a type of branching terminator, get the
71 // branching terminator.
72 static Instruction *getBranchTerminator(const PredicateBase *PB) {
73   assert(isa<PredicateWithEdge>(PB) &&
74          "Not a predicate info type we know how to get a terminator from.");
75   return cast<PredicateWithEdge>(PB)->From->getTerminator();
76 }
77 
78 // Given a predicate info that is a type of branching terminator, get the
79 // edge this predicate info represents
80 std::pair<BasicBlock *, BasicBlock *> getBlockEdge(const PredicateBase *PB) {
81   assert(isa<PredicateWithEdge>(PB) &&
82          "Not a predicate info type we know how to get an edge from.");
83   const auto *PEdge = cast<PredicateWithEdge>(PB);
84   return std::make_pair(PEdge->From, PEdge->To);
85 }
86 }
87 
88 namespace llvm {
89 enum LocalNum {
90   // Operations that must appear first in the block.
91   LN_First,
92   // Operations that are somewhere in the middle of the block, and are sorted on
93   // demand.
94   LN_Middle,
95   // Operations that must appear last in a block, like successor phi node uses.
96   LN_Last
97 };
98 
99 // Associate global and local DFS info with defs and uses, so we can sort them
100 // into a global domination ordering.
101 struct ValueDFS {
102   int DFSIn = 0;
103   int DFSOut = 0;
104   unsigned int LocalNum = LN_Middle;
105   // Only one of Def or Use will be set.
106   Value *Def = nullptr;
107   Use *U = nullptr;
108   // Neither PInfo nor EdgeOnly participate in the ordering
109   PredicateBase *PInfo = nullptr;
110   bool EdgeOnly = false;
111 };
112 
113 // Perform a strict weak ordering on instructions and arguments.
114 static bool valueComesBefore(const Value *A, const Value *B) {
115   auto *ArgA = dyn_cast_or_null<Argument>(A);
116   auto *ArgB = dyn_cast_or_null<Argument>(B);
117   if (ArgA && !ArgB)
118     return true;
119   if (ArgB && !ArgA)
120     return false;
121   if (ArgA && ArgB)
122     return ArgA->getArgNo() < ArgB->getArgNo();
123   return cast<Instruction>(A)->comesBefore(cast<Instruction>(B));
124 }
125 
126 // This compares ValueDFS structures. Doing so allows us to walk the minimum
127 // number of instructions necessary to compute our def/use ordering.
128 struct ValueDFS_Compare {
129   DominatorTree &DT;
130   ValueDFS_Compare(DominatorTree &DT) : DT(DT) {}
131 
132   bool operator()(const ValueDFS &A, const ValueDFS &B) const {
133     if (&A == &B)
134       return false;
135     // The only case we can't directly compare them is when they in the same
136     // block, and both have localnum == middle.  In that case, we have to use
137     // comesbefore to see what the real ordering is, because they are in the
138     // same basic block.
139 
140     assert((A.DFSIn != B.DFSIn || A.DFSOut == B.DFSOut) &&
141            "Equal DFS-in numbers imply equal out numbers");
142     bool SameBlock = A.DFSIn == B.DFSIn;
143 
144     // We want to put the def that will get used for a given set of phi uses,
145     // before those phi uses.
146     // So we sort by edge, then by def.
147     // Note that only phi nodes uses and defs can come last.
148     if (SameBlock && A.LocalNum == LN_Last && B.LocalNum == LN_Last)
149       return comparePHIRelated(A, B);
150 
151     bool isADef = A.Def;
152     bool isBDef = B.Def;
153     if (!SameBlock || A.LocalNum != LN_Middle || B.LocalNum != LN_Middle)
154       return std::tie(A.DFSIn, A.LocalNum, isADef) <
155              std::tie(B.DFSIn, B.LocalNum, isBDef);
156     return localComesBefore(A, B);
157   }
158 
159   // For a phi use, or a non-materialized def, return the edge it represents.
160   std::pair<BasicBlock *, BasicBlock *> getBlockEdge(const ValueDFS &VD) const {
161     if (!VD.Def && VD.U) {
162       auto *PHI = cast<PHINode>(VD.U->getUser());
163       return std::make_pair(PHI->getIncomingBlock(*VD.U), PHI->getParent());
164     }
165     // This is really a non-materialized def.
166     return ::getBlockEdge(VD.PInfo);
167   }
168 
169   // For two phi related values, return the ordering.
170   bool comparePHIRelated(const ValueDFS &A, const ValueDFS &B) const {
171     BasicBlock *ASrc, *ADest, *BSrc, *BDest;
172     std::tie(ASrc, ADest) = getBlockEdge(A);
173     std::tie(BSrc, BDest) = getBlockEdge(B);
174 
175 #ifndef NDEBUG
176     // This function should only be used for values in the same BB, check that.
177     DomTreeNode *DomASrc = DT.getNode(ASrc);
178     DomTreeNode *DomBSrc = DT.getNode(BSrc);
179     assert(DomASrc->getDFSNumIn() == (unsigned)A.DFSIn &&
180            "DFS numbers for A should match the ones of the source block");
181     assert(DomBSrc->getDFSNumIn() == (unsigned)B.DFSIn &&
182            "DFS numbers for B should match the ones of the source block");
183     assert(A.DFSIn == B.DFSIn && "Values must be in the same block");
184 #endif
185     (void)ASrc;
186     (void)BSrc;
187 
188     // Use DFS numbers to compare destination blocks, to guarantee a
189     // deterministic order.
190     DomTreeNode *DomADest = DT.getNode(ADest);
191     DomTreeNode *DomBDest = DT.getNode(BDest);
192     unsigned AIn = DomADest->getDFSNumIn();
193     unsigned BIn = DomBDest->getDFSNumIn();
194     bool isADef = A.Def;
195     bool isBDef = B.Def;
196     assert((!A.Def || !A.U) && (!B.Def || !B.U) &&
197            "Def and U cannot be set at the same time");
198     // Now sort by edge destination and then defs before uses.
199     return std::tie(AIn, isADef) < std::tie(BIn, isBDef);
200   }
201 
202   // Get the definition of an instruction that occurs in the middle of a block.
203   Value *getMiddleDef(const ValueDFS &VD) const {
204     if (VD.Def)
205       return VD.Def;
206     // It's possible for the defs and uses to be null.  For branches, the local
207     // numbering will say the placed predicaeinfos should go first (IE
208     // LN_beginning), so we won't be in this function. For assumes, we will end
209     // up here, beause we need to order the def we will place relative to the
210     // assume.  So for the purpose of ordering, we pretend the def is right
211     // after the assume, because that is where we will insert the info.
212     if (!VD.U) {
213       assert(VD.PInfo &&
214              "No def, no use, and no predicateinfo should not occur");
215       assert(isa<PredicateAssume>(VD.PInfo) &&
216              "Middle of block should only occur for assumes");
217       return cast<PredicateAssume>(VD.PInfo)->AssumeInst->getNextNode();
218     }
219     return nullptr;
220   }
221 
222   // Return either the Def, if it's not null, or the user of the Use, if the def
223   // is null.
224   const Instruction *getDefOrUser(const Value *Def, const Use *U) const {
225     if (Def)
226       return cast<Instruction>(Def);
227     return cast<Instruction>(U->getUser());
228   }
229 
230   // This performs the necessary local basic block ordering checks to tell
231   // whether A comes before B, where both are in the same basic block.
232   bool localComesBefore(const ValueDFS &A, const ValueDFS &B) const {
233     auto *ADef = getMiddleDef(A);
234     auto *BDef = getMiddleDef(B);
235 
236     // See if we have real values or uses. If we have real values, we are
237     // guaranteed they are instructions or arguments. No matter what, we are
238     // guaranteed they are in the same block if they are instructions.
239     auto *ArgA = dyn_cast_or_null<Argument>(ADef);
240     auto *ArgB = dyn_cast_or_null<Argument>(BDef);
241 
242     if (ArgA || ArgB)
243       return valueComesBefore(ArgA, ArgB);
244 
245     auto *AInst = getDefOrUser(ADef, A.U);
246     auto *BInst = getDefOrUser(BDef, B.U);
247     return valueComesBefore(AInst, BInst);
248   }
249 };
250 
251 class PredicateInfoBuilder {
252   // Used to store information about each value we might rename.
253   struct ValueInfo {
254     SmallVector<PredicateBase *, 4> Infos;
255   };
256 
257   PredicateInfo &PI;
258   Function &F;
259   DominatorTree &DT;
260   AssumptionCache &AC;
261 
262   // This stores info about each operand or comparison result we make copies
263   // of. The real ValueInfos start at index 1, index 0 is unused so that we
264   // can more easily detect invalid indexing.
265   SmallVector<ValueInfo, 32> ValueInfos;
266 
267   // This gives the index into the ValueInfos array for a given Value. Because
268   // 0 is not a valid Value Info index, you can use DenseMap::lookup and tell
269   // whether it returned a valid result.
270   DenseMap<Value *, unsigned int> ValueInfoNums;
271 
272   // The set of edges along which we can only handle phi uses, due to critical
273   // edges.
274   DenseSet<std::pair<BasicBlock *, BasicBlock *>> EdgeUsesOnly;
275 
276   ValueInfo &getOrCreateValueInfo(Value *);
277   const ValueInfo &getValueInfo(Value *) const;
278 
279   void processAssume(IntrinsicInst *, BasicBlock *,
280                      SmallVectorImpl<Value *> &OpsToRename);
281   void processBranch(BranchInst *, BasicBlock *,
282                      SmallVectorImpl<Value *> &OpsToRename);
283   void processSwitch(SwitchInst *, BasicBlock *,
284                      SmallVectorImpl<Value *> &OpsToRename);
285   void renameUses(SmallVectorImpl<Value *> &OpsToRename);
286   void addInfoFor(SmallVectorImpl<Value *> &OpsToRename, Value *Op,
287                   PredicateBase *PB);
288 
289   typedef SmallVectorImpl<ValueDFS> ValueDFSStack;
290   void convertUsesToDFSOrdered(Value *, SmallVectorImpl<ValueDFS> &);
291   Value *materializeStack(unsigned int &, ValueDFSStack &, Value *);
292   bool stackIsInScope(const ValueDFSStack &, const ValueDFS &) const;
293   void popStackUntilDFSScope(ValueDFSStack &, const ValueDFS &);
294 
295 public:
296   PredicateInfoBuilder(PredicateInfo &PI, Function &F, DominatorTree &DT,
297                        AssumptionCache &AC)
298       : PI(PI), F(F), DT(DT), AC(AC) {
299     // Push an empty operand info so that we can detect 0 as not finding one
300     ValueInfos.resize(1);
301   }
302 
303   void buildPredicateInfo();
304 };
305 
306 bool PredicateInfoBuilder::stackIsInScope(const ValueDFSStack &Stack,
307                                           const ValueDFS &VDUse) const {
308   if (Stack.empty())
309     return false;
310   // If it's a phi only use, make sure it's for this phi node edge, and that the
311   // use is in a phi node.  If it's anything else, and the top of the stack is
312   // EdgeOnly, we need to pop the stack.  We deliberately sort phi uses next to
313   // the defs they must go with so that we can know it's time to pop the stack
314   // when we hit the end of the phi uses for a given def.
315   if (Stack.back().EdgeOnly) {
316     if (!VDUse.U)
317       return false;
318     auto *PHI = dyn_cast<PHINode>(VDUse.U->getUser());
319     if (!PHI)
320       return false;
321     // Check edge
322     BasicBlock *EdgePred = PHI->getIncomingBlock(*VDUse.U);
323     if (EdgePred != getBranchBlock(Stack.back().PInfo))
324       return false;
325 
326     // Use dominates, which knows how to handle edge dominance.
327     return DT.dominates(getBlockEdge(Stack.back().PInfo), *VDUse.U);
328   }
329 
330   return (VDUse.DFSIn >= Stack.back().DFSIn &&
331           VDUse.DFSOut <= Stack.back().DFSOut);
332 }
333 
334 void PredicateInfoBuilder::popStackUntilDFSScope(ValueDFSStack &Stack,
335                                                  const ValueDFS &VD) {
336   while (!Stack.empty() && !stackIsInScope(Stack, VD))
337     Stack.pop_back();
338 }
339 
340 // Convert the uses of Op into a vector of uses, associating global and local
341 // DFS info with each one.
342 void PredicateInfoBuilder::convertUsesToDFSOrdered(
343     Value *Op, SmallVectorImpl<ValueDFS> &DFSOrderedSet) {
344   for (auto &U : Op->uses()) {
345     if (auto *I = dyn_cast<Instruction>(U.getUser())) {
346       ValueDFS VD;
347       // Put the phi node uses in the incoming block.
348       BasicBlock *IBlock;
349       if (auto *PN = dyn_cast<PHINode>(I)) {
350         IBlock = PN->getIncomingBlock(U);
351         // Make phi node users appear last in the incoming block
352         // they are from.
353         VD.LocalNum = LN_Last;
354       } else {
355         // If it's not a phi node use, it is somewhere in the middle of the
356         // block.
357         IBlock = I->getParent();
358         VD.LocalNum = LN_Middle;
359       }
360       DomTreeNode *DomNode = DT.getNode(IBlock);
361       // It's possible our use is in an unreachable block. Skip it if so.
362       if (!DomNode)
363         continue;
364       VD.DFSIn = DomNode->getDFSNumIn();
365       VD.DFSOut = DomNode->getDFSNumOut();
366       VD.U = &U;
367       DFSOrderedSet.push_back(VD);
368     }
369   }
370 }
371 
372 bool shouldRename(Value *V) {
373   // Only want real values, not constants.  Additionally, operands with one use
374   // are only being used in the comparison, which means they will not be useful
375   // for us to consider for predicateinfo.
376   return (isa<Instruction>(V) || isa<Argument>(V)) && !V->hasOneUse();
377 }
378 
379 // Collect relevant operations from Comparison that we may want to insert copies
380 // for.
381 void collectCmpOps(CmpInst *Comparison, SmallVectorImpl<Value *> &CmpOperands) {
382   auto *Op0 = Comparison->getOperand(0);
383   auto *Op1 = Comparison->getOperand(1);
384   if (Op0 == Op1)
385     return;
386 
387   CmpOperands.push_back(Op0);
388   CmpOperands.push_back(Op1);
389 }
390 
391 // Add Op, PB to the list of value infos for Op, and mark Op to be renamed.
392 void PredicateInfoBuilder::addInfoFor(SmallVectorImpl<Value *> &OpsToRename,
393                                       Value *Op, PredicateBase *PB) {
394   auto &OperandInfo = getOrCreateValueInfo(Op);
395   if (OperandInfo.Infos.empty())
396     OpsToRename.push_back(Op);
397   PI.AllInfos.push_back(PB);
398   OperandInfo.Infos.push_back(PB);
399 }
400 
401 // Process an assume instruction and place relevant operations we want to rename
402 // into OpsToRename.
403 void PredicateInfoBuilder::processAssume(
404     IntrinsicInst *II, BasicBlock *AssumeBB,
405     SmallVectorImpl<Value *> &OpsToRename) {
406   SmallVector<Value *, 4> Worklist;
407   SmallPtrSet<Value *, 4> Visited;
408   Worklist.push_back(II->getOperand(0));
409   while (!Worklist.empty()) {
410     Value *Cond = Worklist.pop_back_val();
411     if (!Visited.insert(Cond).second)
412       continue;
413     if (Visited.size() > MaxCondsPerBranch)
414       break;
415 
416     Value *Op0, *Op1;
417     if (match(Cond, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
418       Worklist.push_back(Op1);
419       Worklist.push_back(Op0);
420     }
421 
422     SmallVector<Value *, 4> Values;
423     Values.push_back(Cond);
424     if (auto *Cmp = dyn_cast<CmpInst>(Cond))
425       collectCmpOps(Cmp, Values);
426 
427     for (Value *V : Values) {
428       if (shouldRename(V)) {
429         auto *PA = new PredicateAssume(V, II, Cond);
430         addInfoFor(OpsToRename, V, PA);
431       }
432     }
433   }
434 }
435 
436 // Process a block terminating branch, and place relevant operations to be
437 // renamed into OpsToRename.
438 void PredicateInfoBuilder::processBranch(
439     BranchInst *BI, BasicBlock *BranchBB,
440     SmallVectorImpl<Value *> &OpsToRename) {
441   BasicBlock *FirstBB = BI->getSuccessor(0);
442   BasicBlock *SecondBB = BI->getSuccessor(1);
443 
444   for (BasicBlock *Succ : {FirstBB, SecondBB}) {
445     bool TakenEdge = Succ == FirstBB;
446     // Don't try to insert on a self-edge. This is mainly because we will
447     // eliminate during renaming anyway.
448     if (Succ == BranchBB)
449       continue;
450 
451     SmallVector<Value *, 4> Worklist;
452     SmallPtrSet<Value *, 4> Visited;
453     Worklist.push_back(BI->getCondition());
454     while (!Worklist.empty()) {
455       Value *Cond = Worklist.pop_back_val();
456       if (!Visited.insert(Cond).second)
457         continue;
458       if (Visited.size() > MaxCondsPerBranch)
459         break;
460 
461       Value *Op0, *Op1;
462       if (TakenEdge ? match(Cond, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))
463                     : match(Cond, m_LogicalOr(m_Value(Op0), m_Value(Op1)))) {
464         Worklist.push_back(Op1);
465         Worklist.push_back(Op0);
466       }
467 
468       SmallVector<Value *, 4> Values;
469       Values.push_back(Cond);
470       if (auto *Cmp = dyn_cast<CmpInst>(Cond))
471         collectCmpOps(Cmp, Values);
472 
473       for (Value *V : Values) {
474         if (shouldRename(V)) {
475           PredicateBase *PB =
476               new PredicateBranch(V, BranchBB, Succ, Cond, TakenEdge);
477           addInfoFor(OpsToRename, V, PB);
478           if (!Succ->getSinglePredecessor())
479             EdgeUsesOnly.insert({BranchBB, Succ});
480         }
481       }
482     }
483   }
484 }
485 // Process a block terminating switch, and place relevant operations to be
486 // renamed into OpsToRename.
487 void PredicateInfoBuilder::processSwitch(
488     SwitchInst *SI, BasicBlock *BranchBB,
489     SmallVectorImpl<Value *> &OpsToRename) {
490   Value *Op = SI->getCondition();
491   if ((!isa<Instruction>(Op) && !isa<Argument>(Op)) || Op->hasOneUse())
492     return;
493 
494   // Remember how many outgoing edges there are to every successor.
495   SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
496   for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {
497     BasicBlock *TargetBlock = SI->getSuccessor(i);
498     ++SwitchEdges[TargetBlock];
499   }
500 
501   // Now propagate info for each case value
502   for (auto C : SI->cases()) {
503     BasicBlock *TargetBlock = C.getCaseSuccessor();
504     if (SwitchEdges.lookup(TargetBlock) == 1) {
505       PredicateSwitch *PS = new PredicateSwitch(
506           Op, SI->getParent(), TargetBlock, C.getCaseValue(), SI);
507       addInfoFor(OpsToRename, Op, PS);
508       if (!TargetBlock->getSinglePredecessor())
509         EdgeUsesOnly.insert({BranchBB, TargetBlock});
510     }
511   }
512 }
513 
514 // Build predicate info for our function
515 void PredicateInfoBuilder::buildPredicateInfo() {
516   DT.updateDFSNumbers();
517   // Collect operands to rename from all conditional branch terminators, as well
518   // as assume statements.
519   SmallVector<Value *, 8> OpsToRename;
520   for (auto DTN : depth_first(DT.getRootNode())) {
521     BasicBlock *BranchBB = DTN->getBlock();
522     if (auto *BI = dyn_cast<BranchInst>(BranchBB->getTerminator())) {
523       if (!BI->isConditional())
524         continue;
525       // Can't insert conditional information if they all go to the same place.
526       if (BI->getSuccessor(0) == BI->getSuccessor(1))
527         continue;
528       processBranch(BI, BranchBB, OpsToRename);
529     } else if (auto *SI = dyn_cast<SwitchInst>(BranchBB->getTerminator())) {
530       processSwitch(SI, BranchBB, OpsToRename);
531     }
532   }
533   for (auto &Assume : AC.assumptions()) {
534     if (auto *II = dyn_cast_or_null<IntrinsicInst>(Assume))
535       if (DT.isReachableFromEntry(II->getParent()))
536         processAssume(II, II->getParent(), OpsToRename);
537   }
538   // Now rename all our operations.
539   renameUses(OpsToRename);
540 }
541 
542 // Given the renaming stack, make all the operands currently on the stack real
543 // by inserting them into the IR.  Return the last operation's value.
544 Value *PredicateInfoBuilder::materializeStack(unsigned int &Counter,
545                                              ValueDFSStack &RenameStack,
546                                              Value *OrigOp) {
547   // Find the first thing we have to materialize
548   auto RevIter = RenameStack.rbegin();
549   for (; RevIter != RenameStack.rend(); ++RevIter)
550     if (RevIter->Def)
551       break;
552 
553   size_t Start = RevIter - RenameStack.rbegin();
554   // The maximum number of things we should be trying to materialize at once
555   // right now is 4, depending on if we had an assume, a branch, and both used
556   // and of conditions.
557   for (auto RenameIter = RenameStack.end() - Start;
558        RenameIter != RenameStack.end(); ++RenameIter) {
559     auto *Op =
560         RenameIter == RenameStack.begin() ? OrigOp : (RenameIter - 1)->Def;
561     ValueDFS &Result = *RenameIter;
562     auto *ValInfo = Result.PInfo;
563     ValInfo->RenamedOp = (RenameStack.end() - Start) == RenameStack.begin()
564                              ? OrigOp
565                              : (RenameStack.end() - Start - 1)->Def;
566     // For edge predicates, we can just place the operand in the block before
567     // the terminator.  For assume, we have to place it right before the assume
568     // to ensure we dominate all of our uses.  Always insert right before the
569     // relevant instruction (terminator, assume), so that we insert in proper
570     // order in the case of multiple predicateinfo in the same block.
571     // The number of named values is used to detect if a new declaration was
572     // added. If so, that declaration is tracked so that it can be removed when
573     // the analysis is done. The corner case were a new declaration results in
574     // a name clash and the old name being renamed is not considered as that
575     // represents an invalid module.
576     if (isa<PredicateWithEdge>(ValInfo)) {
577       IRBuilder<> B(getBranchTerminator(ValInfo));
578       auto NumDecls = F.getParent()->getNumNamedValues();
579       Function *IF = Intrinsic::getDeclaration(
580           F.getParent(), Intrinsic::ssa_copy, Op->getType());
581       if (NumDecls != F.getParent()->getNumNamedValues())
582         PI.CreatedDeclarations.insert(IF);
583       CallInst *PIC =
584           B.CreateCall(IF, Op, Op->getName() + "." + Twine(Counter++));
585       PI.PredicateMap.insert({PIC, ValInfo});
586       Result.Def = PIC;
587     } else {
588       auto *PAssume = dyn_cast<PredicateAssume>(ValInfo);
589       assert(PAssume &&
590              "Should not have gotten here without it being an assume");
591       // Insert the predicate directly after the assume. While it also holds
592       // directly before it, assume(i1 true) is not a useful fact.
593       IRBuilder<> B(PAssume->AssumeInst->getNextNode());
594       auto NumDecls = F.getParent()->getNumNamedValues();
595       Function *IF = Intrinsic::getDeclaration(
596           F.getParent(), Intrinsic::ssa_copy, Op->getType());
597       if (NumDecls != F.getParent()->getNumNamedValues())
598         PI.CreatedDeclarations.insert(IF);
599       CallInst *PIC = B.CreateCall(IF, Op);
600       PI.PredicateMap.insert({PIC, ValInfo});
601       Result.Def = PIC;
602     }
603   }
604   return RenameStack.back().Def;
605 }
606 
607 // Instead of the standard SSA renaming algorithm, which is O(Number of
608 // instructions), and walks the entire dominator tree, we walk only the defs +
609 // uses.  The standard SSA renaming algorithm does not really rely on the
610 // dominator tree except to order the stack push/pops of the renaming stacks, so
611 // that defs end up getting pushed before hitting the correct uses.  This does
612 // not require the dominator tree, only the *order* of the dominator tree. The
613 // complete and correct ordering of the defs and uses, in dominator tree is
614 // contained in the DFS numbering of the dominator tree. So we sort the defs and
615 // uses into the DFS ordering, and then just use the renaming stack as per
616 // normal, pushing when we hit a def (which is a predicateinfo instruction),
617 // popping when we are out of the dfs scope for that def, and replacing any uses
618 // with top of stack if it exists.  In order to handle liveness without
619 // propagating liveness info, we don't actually insert the predicateinfo
620 // instruction def until we see a use that it would dominate.  Once we see such
621 // a use, we materialize the predicateinfo instruction in the right place and
622 // use it.
623 //
624 // TODO: Use this algorithm to perform fast single-variable renaming in
625 // promotememtoreg and memoryssa.
626 void PredicateInfoBuilder::renameUses(SmallVectorImpl<Value *> &OpsToRename) {
627   ValueDFS_Compare Compare(DT);
628   // Compute liveness, and rename in O(uses) per Op.
629   for (auto *Op : OpsToRename) {
630     LLVM_DEBUG(dbgs() << "Visiting " << *Op << "\n");
631     unsigned Counter = 0;
632     SmallVector<ValueDFS, 16> OrderedUses;
633     const auto &ValueInfo = getValueInfo(Op);
634     // Insert the possible copies into the def/use list.
635     // They will become real copies if we find a real use for them, and never
636     // created otherwise.
637     for (auto &PossibleCopy : ValueInfo.Infos) {
638       ValueDFS VD;
639       // Determine where we are going to place the copy by the copy type.
640       // The predicate info for branches always come first, they will get
641       // materialized in the split block at the top of the block.
642       // The predicate info for assumes will be somewhere in the middle,
643       // it will get materialized in front of the assume.
644       if (const auto *PAssume = dyn_cast<PredicateAssume>(PossibleCopy)) {
645         VD.LocalNum = LN_Middle;
646         DomTreeNode *DomNode = DT.getNode(PAssume->AssumeInst->getParent());
647         if (!DomNode)
648           continue;
649         VD.DFSIn = DomNode->getDFSNumIn();
650         VD.DFSOut = DomNode->getDFSNumOut();
651         VD.PInfo = PossibleCopy;
652         OrderedUses.push_back(VD);
653       } else if (isa<PredicateWithEdge>(PossibleCopy)) {
654         // If we can only do phi uses, we treat it like it's in the branch
655         // block, and handle it specially. We know that it goes last, and only
656         // dominate phi uses.
657         auto BlockEdge = getBlockEdge(PossibleCopy);
658         if (EdgeUsesOnly.count(BlockEdge)) {
659           VD.LocalNum = LN_Last;
660           auto *DomNode = DT.getNode(BlockEdge.first);
661           if (DomNode) {
662             VD.DFSIn = DomNode->getDFSNumIn();
663             VD.DFSOut = DomNode->getDFSNumOut();
664             VD.PInfo = PossibleCopy;
665             VD.EdgeOnly = true;
666             OrderedUses.push_back(VD);
667           }
668         } else {
669           // Otherwise, we are in the split block (even though we perform
670           // insertion in the branch block).
671           // Insert a possible copy at the split block and before the branch.
672           VD.LocalNum = LN_First;
673           auto *DomNode = DT.getNode(BlockEdge.second);
674           if (DomNode) {
675             VD.DFSIn = DomNode->getDFSNumIn();
676             VD.DFSOut = DomNode->getDFSNumOut();
677             VD.PInfo = PossibleCopy;
678             OrderedUses.push_back(VD);
679           }
680         }
681       }
682     }
683 
684     convertUsesToDFSOrdered(Op, OrderedUses);
685     // Here we require a stable sort because we do not bother to try to
686     // assign an order to the operands the uses represent. Thus, two
687     // uses in the same instruction do not have a strict sort order
688     // currently and will be considered equal. We could get rid of the
689     // stable sort by creating one if we wanted.
690     llvm::stable_sort(OrderedUses, Compare);
691     SmallVector<ValueDFS, 8> RenameStack;
692     // For each use, sorted into dfs order, push values and replaces uses with
693     // top of stack, which will represent the reaching def.
694     for (auto &VD : OrderedUses) {
695       // We currently do not materialize copy over copy, but we should decide if
696       // we want to.
697       bool PossibleCopy = VD.PInfo != nullptr;
698       if (RenameStack.empty()) {
699         LLVM_DEBUG(dbgs() << "Rename Stack is empty\n");
700       } else {
701         LLVM_DEBUG(dbgs() << "Rename Stack Top DFS numbers are ("
702                           << RenameStack.back().DFSIn << ","
703                           << RenameStack.back().DFSOut << ")\n");
704       }
705 
706       LLVM_DEBUG(dbgs() << "Current DFS numbers are (" << VD.DFSIn << ","
707                         << VD.DFSOut << ")\n");
708 
709       bool ShouldPush = (VD.Def || PossibleCopy);
710       bool OutOfScope = !stackIsInScope(RenameStack, VD);
711       if (OutOfScope || ShouldPush) {
712         // Sync to our current scope.
713         popStackUntilDFSScope(RenameStack, VD);
714         if (ShouldPush) {
715           RenameStack.push_back(VD);
716         }
717       }
718       // If we get to this point, and the stack is empty we must have a use
719       // with no renaming needed, just skip it.
720       if (RenameStack.empty())
721         continue;
722       // Skip values, only want to rename the uses
723       if (VD.Def || PossibleCopy)
724         continue;
725       if (!DebugCounter::shouldExecute(RenameCounter)) {
726         LLVM_DEBUG(dbgs() << "Skipping execution due to debug counter\n");
727         continue;
728       }
729       ValueDFS &Result = RenameStack.back();
730 
731       // If the possible copy dominates something, materialize our stack up to
732       // this point. This ensures every comparison that affects our operation
733       // ends up with predicateinfo.
734       if (!Result.Def)
735         Result.Def = materializeStack(Counter, RenameStack, Op);
736 
737       LLVM_DEBUG(dbgs() << "Found replacement " << *Result.Def << " for "
738                         << *VD.U->get() << " in " << *(VD.U->getUser())
739                         << "\n");
740       assert(DT.dominates(cast<Instruction>(Result.Def), *VD.U) &&
741              "Predicateinfo def should have dominated this use");
742       VD.U->set(Result.Def);
743     }
744   }
745 }
746 
747 PredicateInfoBuilder::ValueInfo &
748 PredicateInfoBuilder::getOrCreateValueInfo(Value *Operand) {
749   auto OIN = ValueInfoNums.find(Operand);
750   if (OIN == ValueInfoNums.end()) {
751     // This will grow it
752     ValueInfos.resize(ValueInfos.size() + 1);
753     // This will use the new size and give us a 0 based number of the info
754     auto InsertResult = ValueInfoNums.insert({Operand, ValueInfos.size() - 1});
755     assert(InsertResult.second && "Value info number already existed?");
756     return ValueInfos[InsertResult.first->second];
757   }
758   return ValueInfos[OIN->second];
759 }
760 
761 const PredicateInfoBuilder::ValueInfo &
762 PredicateInfoBuilder::getValueInfo(Value *Operand) const {
763   auto OINI = ValueInfoNums.lookup(Operand);
764   assert(OINI != 0 && "Operand was not really in the Value Info Numbers");
765   assert(OINI < ValueInfos.size() &&
766          "Value Info Number greater than size of Value Info Table");
767   return ValueInfos[OINI];
768 }
769 
770 PredicateInfo::PredicateInfo(Function &F, DominatorTree &DT,
771                              AssumptionCache &AC)
772     : F(F) {
773   PredicateInfoBuilder Builder(*this, F, DT, AC);
774   Builder.buildPredicateInfo();
775 }
776 
777 // Remove all declarations we created . The PredicateInfo consumers are
778 // responsible for remove the ssa_copy calls created.
779 PredicateInfo::~PredicateInfo() {
780   // Collect function pointers in set first, as SmallSet uses a SmallVector
781   // internally and we have to remove the asserting value handles first.
782   SmallPtrSet<Function *, 20> FunctionPtrs;
783   for (auto &F : CreatedDeclarations)
784     FunctionPtrs.insert(&*F);
785   CreatedDeclarations.clear();
786 
787   for (Function *F : FunctionPtrs) {
788     assert(F->user_begin() == F->user_end() &&
789            "PredicateInfo consumer did not remove all SSA copies.");
790     F->eraseFromParent();
791   }
792 }
793 
794 Optional<PredicateConstraint> PredicateBase::getConstraint() const {
795   switch (Type) {
796   case PT_Assume:
797   case PT_Branch: {
798     bool TrueEdge = true;
799     if (auto *PBranch = dyn_cast<PredicateBranch>(this))
800       TrueEdge = PBranch->TrueEdge;
801 
802     if (Condition == RenamedOp) {
803       return {{CmpInst::ICMP_EQ,
804                TrueEdge ? ConstantInt::getTrue(Condition->getType())
805                         : ConstantInt::getFalse(Condition->getType())}};
806     }
807 
808     CmpInst *Cmp = dyn_cast<CmpInst>(Condition);
809     if (!Cmp) {
810       // TODO: Make this an assertion once RenamedOp is fully accurate.
811       return None;
812     }
813 
814     CmpInst::Predicate Pred;
815     Value *OtherOp;
816     if (Cmp->getOperand(0) == RenamedOp) {
817       Pred = Cmp->getPredicate();
818       OtherOp = Cmp->getOperand(1);
819     } else if (Cmp->getOperand(1) == RenamedOp) {
820       Pred = Cmp->getSwappedPredicate();
821       OtherOp = Cmp->getOperand(0);
822     } else {
823       // TODO: Make this an assertion once RenamedOp is fully accurate.
824       return None;
825     }
826 
827     // Invert predicate along false edge.
828     if (!TrueEdge)
829       Pred = CmpInst::getInversePredicate(Pred);
830 
831     return {{Pred, OtherOp}};
832   }
833   case PT_Switch:
834     if (Condition != RenamedOp) {
835       // TODO: Make this an assertion once RenamedOp is fully accurate.
836       return None;
837     }
838 
839     return {{CmpInst::ICMP_EQ, cast<PredicateSwitch>(this)->CaseValue}};
840   }
841   llvm_unreachable("Unknown predicate type");
842 }
843 
844 void PredicateInfo::verifyPredicateInfo() const {}
845 
846 char PredicateInfoPrinterLegacyPass::ID = 0;
847 
848 PredicateInfoPrinterLegacyPass::PredicateInfoPrinterLegacyPass()
849     : FunctionPass(ID) {
850   initializePredicateInfoPrinterLegacyPassPass(
851       *PassRegistry::getPassRegistry());
852 }
853 
854 void PredicateInfoPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
855   AU.setPreservesAll();
856   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
857   AU.addRequired<AssumptionCacheTracker>();
858 }
859 
860 // Replace ssa_copy calls created by PredicateInfo with their operand.
861 static void replaceCreatedSSACopys(PredicateInfo &PredInfo, Function &F) {
862   for (Instruction &Inst : llvm::make_early_inc_range(instructions(F))) {
863     const auto *PI = PredInfo.getPredicateInfoFor(&Inst);
864     auto *II = dyn_cast<IntrinsicInst>(&Inst);
865     if (!PI || !II || II->getIntrinsicID() != Intrinsic::ssa_copy)
866       continue;
867 
868     Inst.replaceAllUsesWith(II->getOperand(0));
869     Inst.eraseFromParent();
870   }
871 }
872 
873 bool PredicateInfoPrinterLegacyPass::runOnFunction(Function &F) {
874   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
875   auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
876   auto PredInfo = std::make_unique<PredicateInfo>(F, DT, AC);
877   PredInfo->print(dbgs());
878   if (VerifyPredicateInfo)
879     PredInfo->verifyPredicateInfo();
880 
881   replaceCreatedSSACopys(*PredInfo, F);
882   return false;
883 }
884 
885 PreservedAnalyses PredicateInfoPrinterPass::run(Function &F,
886                                                 FunctionAnalysisManager &AM) {
887   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
888   auto &AC = AM.getResult<AssumptionAnalysis>(F);
889   OS << "PredicateInfo for function: " << F.getName() << "\n";
890   auto PredInfo = std::make_unique<PredicateInfo>(F, DT, AC);
891   PredInfo->print(OS);
892 
893   replaceCreatedSSACopys(*PredInfo, F);
894   return PreservedAnalyses::all();
895 }
896 
897 /// An assembly annotator class to print PredicateInfo information in
898 /// comments.
899 class PredicateInfoAnnotatedWriter : public AssemblyAnnotationWriter {
900   friend class PredicateInfo;
901   const PredicateInfo *PredInfo;
902 
903 public:
904   PredicateInfoAnnotatedWriter(const PredicateInfo *M) : PredInfo(M) {}
905 
906   void emitBasicBlockStartAnnot(const BasicBlock *BB,
907                                 formatted_raw_ostream &OS) override {}
908 
909   void emitInstructionAnnot(const Instruction *I,
910                             formatted_raw_ostream &OS) override {
911     if (const auto *PI = PredInfo->getPredicateInfoFor(I)) {
912       OS << "; Has predicate info\n";
913       if (const auto *PB = dyn_cast<PredicateBranch>(PI)) {
914         OS << "; branch predicate info { TrueEdge: " << PB->TrueEdge
915            << " Comparison:" << *PB->Condition << " Edge: [";
916         PB->From->printAsOperand(OS);
917         OS << ",";
918         PB->To->printAsOperand(OS);
919         OS << "]";
920       } else if (const auto *PS = dyn_cast<PredicateSwitch>(PI)) {
921         OS << "; switch predicate info { CaseValue: " << *PS->CaseValue
922            << " Switch:" << *PS->Switch << " Edge: [";
923         PS->From->printAsOperand(OS);
924         OS << ",";
925         PS->To->printAsOperand(OS);
926         OS << "]";
927       } else if (const auto *PA = dyn_cast<PredicateAssume>(PI)) {
928         OS << "; assume predicate info {"
929            << " Comparison:" << *PA->Condition;
930       }
931       OS << ", RenamedOp: ";
932       PI->RenamedOp->printAsOperand(OS, false);
933       OS << " }\n";
934     }
935   }
936 };
937 
938 void PredicateInfo::print(raw_ostream &OS) const {
939   PredicateInfoAnnotatedWriter Writer(this);
940   F.print(OS, &Writer);
941 }
942 
943 void PredicateInfo::dump() const {
944   PredicateInfoAnnotatedWriter Writer(this);
945   F.print(dbgs(), &Writer);
946 }
947 
948 PreservedAnalyses PredicateInfoVerifierPass::run(Function &F,
949                                                  FunctionAnalysisManager &AM) {
950   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
951   auto &AC = AM.getResult<AssumptionAnalysis>(F);
952   std::make_unique<PredicateInfo>(F, DT, AC)->verifyPredicateInfo();
953 
954   return PreservedAnalyses::all();
955 }
956 }
957