1 //===-- PredicateInfo.cpp - PredicateInfo Builder--------------------===//
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
7 //===----------------------------------------------------------------===//
9 // This file implements the PredicateInfo class.
11 //===----------------------------------------------------------------===//
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"
40 #define DEBUG_TYPE "predicateinfo"
42 using namespace PatternMatch
;
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");
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;
61 // Given a predicate info that is a type of branching terminator, get the
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
;
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();
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
);
90 // Operations that must appear first in the block.
92 // Operations that are somewhere in the middle of the block, and are sorted on
95 // Operations that must appear last in a block, like successor phi node uses.
99 // Associate global and local DFS info with defs and uses, so we can sort them
100 // into a global domination ordering.
104 unsigned int LocalNum
= LN_Middle
;
105 // Only one of Def or Use will be set.
106 Value
*Def
= nullptr;
108 // Neither PInfo nor EdgeOnly participate in the ordering
109 PredicateBase
*PInfo
= nullptr;
110 bool EdgeOnly
= false;
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
);
122 return ArgA
->getArgNo() < ArgB
->getArgNo();
123 return cast
<Instruction
>(A
)->comesBefore(cast
<Instruction
>(B
));
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
{
130 ValueDFS_Compare(DominatorTree
&DT
) : DT(DT
) {}
132 bool operator()(const ValueDFS
&A
, const ValueDFS
&B
) const {
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
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
;
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
);
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
);
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());
165 // This is really a non-materialized def.
166 return ::getBlockEdge(VD
.PInfo
);
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
);
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");
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();
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
);
202 // Get the definition of an instruction that occurs in the middle of a block.
203 Value
*getMiddleDef(const ValueDFS
&VD
) const {
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.
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();
222 // Return either the Def, if it's not null, or the user of the Use, if the def
224 const Instruction
*getDefOrUser(const Value
*Def
, const Use
*U
) const {
226 return cast
<Instruction
>(Def
);
227 return cast
<Instruction
>(U
->getUser());
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
);
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
);
243 return valueComesBefore(ArgA
, ArgB
);
245 auto *AInst
= getDefOrUser(ADef
, A
.U
);
246 auto *BInst
= getDefOrUser(BDef
, B
.U
);
247 return valueComesBefore(AInst
, BInst
);
251 class PredicateInfoBuilder
{
252 // Used to store information about each value we might rename.
254 SmallVector
<PredicateBase
*, 4> Infos
;
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
;
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
;
272 // The set of edges along which we can only handle phi uses, due to critical
274 DenseSet
<std::pair
<BasicBlock
*, BasicBlock
*>> EdgeUsesOnly
;
276 ValueInfo
&getOrCreateValueInfo(Value
*);
277 const ValueInfo
&getValueInfo(Value
*) const;
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
,
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
&);
296 PredicateInfoBuilder(PredicateInfo
&PI
, Function
&F
, DominatorTree
&DT
,
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);
303 void buildPredicateInfo();
306 bool PredicateInfoBuilder::stackIsInScope(const ValueDFSStack
&Stack
,
307 const ValueDFS
&VDUse
) const {
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
) {
318 auto *PHI
= dyn_cast
<PHINode
>(VDUse
.U
->getUser());
322 BasicBlock
*EdgePred
= PHI
->getIncomingBlock(*VDUse
.U
);
323 if (EdgePred
!= getBranchBlock(Stack
.back().PInfo
))
326 // Use dominates, which knows how to handle edge dominance.
327 return DT
.dominates(getBlockEdge(Stack
.back().PInfo
), *VDUse
.U
);
330 return (VDUse
.DFSIn
>= Stack
.back().DFSIn
&&
331 VDUse
.DFSOut
<= Stack
.back().DFSOut
);
334 void PredicateInfoBuilder::popStackUntilDFSScope(ValueDFSStack
&Stack
,
335 const ValueDFS
&VD
) {
336 while (!Stack
.empty() && !stackIsInScope(Stack
, VD
))
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())) {
347 // Put the phi node uses in the incoming block.
349 if (auto *PN
= dyn_cast
<PHINode
>(I
)) {
350 IBlock
= PN
->getIncomingBlock(U
);
351 // Make phi node users appear last in the incoming block
353 VD
.LocalNum
= LN_Last
;
355 // If it's not a phi node use, it is somewhere in the middle of the
357 IBlock
= I
->getParent();
358 VD
.LocalNum
= LN_Middle
;
360 DomTreeNode
*DomNode
= DT
.getNode(IBlock
);
361 // It's possible our use is in an unreachable block. Skip it if so.
364 VD
.DFSIn
= DomNode
->getDFSNumIn();
365 VD
.DFSOut
= DomNode
->getDFSNumOut();
367 DFSOrderedSet
.push_back(VD
);
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();
379 // Collect relevant operations from Comparison that we may want to insert copies
381 void collectCmpOps(CmpInst
*Comparison
, SmallVectorImpl
<Value
*> &CmpOperands
) {
382 auto *Op0
= Comparison
->getOperand(0);
383 auto *Op1
= Comparison
->getOperand(1);
387 CmpOperands
.push_back(Op0
);
388 CmpOperands
.push_back(Op1
);
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
);
401 // Process an assume instruction and place relevant operations we want to rename
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
)
413 if (Visited
.size() > MaxCondsPerBranch
)
417 if (match(Cond
, m_LogicalAnd(m_Value(Op0
), m_Value(Op1
)))) {
418 Worklist
.push_back(Op1
);
419 Worklist
.push_back(Op0
);
422 SmallVector
<Value
*, 4> Values
;
423 Values
.push_back(Cond
);
424 if (auto *Cmp
= dyn_cast
<CmpInst
>(Cond
))
425 collectCmpOps(Cmp
, Values
);
427 for (Value
*V
: Values
) {
428 if (shouldRename(V
)) {
429 auto *PA
= new PredicateAssume(V
, II
, Cond
);
430 addInfoFor(OpsToRename
, V
, PA
);
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);
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
)
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
)
458 if (Visited
.size() > MaxCondsPerBranch
)
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
);
468 SmallVector
<Value
*, 4> Values
;
469 Values
.push_back(Cond
);
470 if (auto *Cmp
= dyn_cast
<CmpInst
>(Cond
))
471 collectCmpOps(Cmp
, Values
);
473 for (Value
*V
: Values
) {
474 if (shouldRename(V
)) {
476 new PredicateBranch(V
, BranchBB
, Succ
, Cond
, TakenEdge
);
477 addInfoFor(OpsToRename
, V
, PB
);
478 if (!Succ
->getSinglePredecessor())
479 EdgeUsesOnly
.insert({BranchBB
, Succ
});
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())
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
];
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
});
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())
525 // Can't insert conditional information if they all go to the same place.
526 if (BI
->getSuccessor(0) == BI
->getSuccessor(1))
528 processBranch(BI
, BranchBB
, OpsToRename
);
529 } else if (auto *SI
= dyn_cast
<SwitchInst
>(BranchBB
->getTerminator())) {
530 processSwitch(SI
, BranchBB
, OpsToRename
);
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
);
538 // Now rename all our operations.
539 renameUses(OpsToRename
);
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
,
547 // Find the first thing we have to materialize
548 auto RevIter
= RenameStack
.rbegin();
549 for (; RevIter
!= RenameStack
.rend(); ++RevIter
)
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
) {
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()
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
);
584 B
.CreateCall(IF
, Op
, Op
->getName() + "." + Twine(Counter
++));
585 PI
.PredicateMap
.insert({PIC
, ValInfo
});
588 auto *PAssume
= dyn_cast
<PredicateAssume
>(ValInfo
);
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
});
604 return RenameStack
.back().Def
;
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
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
) {
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());
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
);
662 VD
.DFSIn
= DomNode
->getDFSNumIn();
663 VD
.DFSOut
= DomNode
->getDFSNumOut();
664 VD
.PInfo
= PossibleCopy
;
666 OrderedUses
.push_back(VD
);
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
);
675 VD
.DFSIn
= DomNode
->getDFSNumIn();
676 VD
.DFSOut
= DomNode
->getDFSNumOut();
677 VD
.PInfo
= PossibleCopy
;
678 OrderedUses
.push_back(VD
);
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
697 bool PossibleCopy
= VD
.PInfo
!= nullptr;
698 if (RenameStack
.empty()) {
699 LLVM_DEBUG(dbgs() << "Rename Stack is empty\n");
701 LLVM_DEBUG(dbgs() << "Rename Stack Top DFS numbers are ("
702 << RenameStack
.back().DFSIn
<< ","
703 << RenameStack
.back().DFSOut
<< ")\n");
706 LLVM_DEBUG(dbgs() << "Current DFS numbers are (" << VD
.DFSIn
<< ","
707 << VD
.DFSOut
<< ")\n");
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
);
715 RenameStack
.push_back(VD
);
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())
722 // Skip values, only want to rename the uses
723 if (VD
.Def
|| PossibleCopy
)
725 if (!DebugCounter::shouldExecute(RenameCounter
)) {
726 LLVM_DEBUG(dbgs() << "Skipping execution due to debug counter\n");
729 ValueDFS
&Result
= RenameStack
.back();
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.
735 Result
.Def
= materializeStack(Counter
, RenameStack
, Op
);
737 LLVM_DEBUG(dbgs() << "Found replacement " << *Result
.Def
<< " for "
738 << *VD
.U
->get() << " in " << *(VD
.U
->getUser())
740 assert(DT
.dominates(cast
<Instruction
>(Result
.Def
), *VD
.U
) &&
741 "Predicateinfo def should have dominated this use");
742 VD
.U
->set(Result
.Def
);
747 PredicateInfoBuilder::ValueInfo
&
748 PredicateInfoBuilder::getOrCreateValueInfo(Value
*Operand
) {
749 auto OIN
= ValueInfoNums
.find(Operand
);
750 if (OIN
== ValueInfoNums
.end()) {
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
];
758 return ValueInfos
[OIN
->second
];
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
];
770 PredicateInfo::PredicateInfo(Function
&F
, DominatorTree
&DT
,
773 PredicateInfoBuilder
Builder(*this, F
, DT
, AC
);
774 Builder
.buildPredicateInfo();
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();
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();
794 Optional
<PredicateConstraint
> PredicateBase::getConstraint() const {
798 bool TrueEdge
= true;
799 if (auto *PBranch
= dyn_cast
<PredicateBranch
>(this))
800 TrueEdge
= PBranch
->TrueEdge
;
802 if (Condition
== RenamedOp
) {
803 return {{CmpInst::ICMP_EQ
,
804 TrueEdge
? ConstantInt::getTrue(Condition
->getType())
805 : ConstantInt::getFalse(Condition
->getType())}};
808 CmpInst
*Cmp
= dyn_cast
<CmpInst
>(Condition
);
810 // TODO: Make this an assertion once RenamedOp is fully accurate.
814 CmpInst::Predicate Pred
;
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);
823 // TODO: Make this an assertion once RenamedOp is fully accurate.
827 // Invert predicate along false edge.
829 Pred
= CmpInst::getInversePredicate(Pred
);
831 return {{Pred
, OtherOp
}};
834 if (Condition
!= RenamedOp
) {
835 // TODO: Make this an assertion once RenamedOp is fully accurate.
839 return {{CmpInst::ICMP_EQ
, cast
<PredicateSwitch
>(this)->CaseValue
}};
841 llvm_unreachable("Unknown predicate type");
844 void PredicateInfo::verifyPredicateInfo() const {}
846 char PredicateInfoPrinterLegacyPass::ID
= 0;
848 PredicateInfoPrinterLegacyPass::PredicateInfoPrinterLegacyPass()
850 initializePredicateInfoPrinterLegacyPassPass(
851 *PassRegistry::getPassRegistry());
854 void PredicateInfoPrinterLegacyPass::getAnalysisUsage(AnalysisUsage
&AU
) const {
855 AU
.setPreservesAll();
856 AU
.addRequiredTransitive
<DominatorTreeWrapperPass
>();
857 AU
.addRequired
<AssumptionCacheTracker
>();
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
)
868 Inst
.replaceAllUsesWith(II
->getOperand(0));
869 Inst
.eraseFromParent();
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();
881 replaceCreatedSSACopys(*PredInfo
, F
);
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
);
893 replaceCreatedSSACopys(*PredInfo
, F
);
894 return PreservedAnalyses::all();
897 /// An assembly annotator class to print PredicateInfo information in
899 class PredicateInfoAnnotatedWriter
: public AssemblyAnnotationWriter
{
900 friend class PredicateInfo
;
901 const PredicateInfo
*PredInfo
;
904 PredicateInfoAnnotatedWriter(const PredicateInfo
*M
) : PredInfo(M
) {}
906 void emitBasicBlockStartAnnot(const BasicBlock
*BB
,
907 formatted_raw_ostream
&OS
) override
{}
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
);
918 PB
->To
->printAsOperand(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
);
925 PS
->To
->printAsOperand(OS
);
927 } else if (const auto *PA
= dyn_cast
<PredicateAssume
>(PI
)) {
928 OS
<< "; assume predicate info {"
929 << " Comparison:" << *PA
->Condition
;
931 OS
<< ", RenamedOp: ";
932 PI
->RenamedOp
->printAsOperand(OS
, false);
938 void PredicateInfo::print(raw_ostream
&OS
) const {
939 PredicateInfoAnnotatedWriter
Writer(this);
940 F
.print(OS
, &Writer
);
943 void PredicateInfo::dump() const {
944 PredicateInfoAnnotatedWriter
Writer(this);
945 F
.print(dbgs(), &Writer
);
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();
954 return PreservedAnalyses::all();