llvm-6502/include/llvm/Analysis/MemoryDependenceAnalysis.h

312 lines
14 KiB
C
Raw Normal View History

//===- llvm/Analysis/MemoryDependenceAnalysis.h - Memory Deps --*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the MemoryDependenceAnalysis analysis pass.
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_ANALYSIS_MEMORY_DEPENDENCE_H
#define LLVM_ANALYSIS_MEMORY_DEPENDENCE_H
#include "llvm/BasicBlock.h"
#include "llvm/Pass.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/OwningPtr.h"
#include "llvm/ADT/PointerIntPair.h"
namespace llvm {
class Function;
class FunctionPass;
class Instruction;
class CallSite;
class AliasAnalysis;
class TargetData;
class MemoryDependenceAnalysis;
class PredIteratorCache;
class DominatorTree;
/// MemDepResult - A memory dependence query can return one of three different
/// answers, described below.
class MemDepResult {
enum DepType {
/// Invalid - Clients of MemDep never see this.
Invalid = 0,
/// Clobber - This is a dependence on the specified instruction which
/// clobbers the desired value. The pointer member of the MemDepResult
/// pair holds the instruction that clobbers the memory. For example,
/// this occurs when we see a may-aliased store to the memory location we
/// care about.
Clobber,
/// Def - This is a dependence on the specified instruction which
/// defines/produces the desired memory location. The pointer member of
/// the MemDepResult pair holds the instruction that defines the memory.
/// Cases of interest:
/// 1. This could be a load or store for dependence queries on
/// load/store. The value loaded or stored is the produced value.
/// Note that the pointer operand may be different than that of the
/// queried pointer due to must aliases and phi translation. Note
/// that the def may not be the same type as the query, the pointers
/// may just be must aliases.
/// 2. For loads and stores, this could be an allocation instruction. In
/// this case, the load is loading an undef value or a store is the
/// first store to (that part of) the allocation.
/// 3. Dependence queries on calls return Def only when they are
/// readonly calls or memory use intrinsics with identical callees
/// and no intervening clobbers. No validation is done that the
/// operands to the calls are the same.
Def,
/// NonLocal - This marker indicates that the query has no dependency in
/// the specified block. To find out more, the client should query other
/// predecessor blocks.
NonLocal
};
typedef PointerIntPair<Instruction*, 2, DepType> PairTy;
PairTy Value;
explicit MemDepResult(PairTy V) : Value(V) {}
public:
MemDepResult() : Value(0, Invalid) {}
/// get methods: These are static ctor methods for creating various
/// MemDepResult kinds.
static MemDepResult getDef(Instruction *Inst) {
return MemDepResult(PairTy(Inst, Def));
}
static MemDepResult getClobber(Instruction *Inst) {
return MemDepResult(PairTy(Inst, Clobber));
}
static MemDepResult getNonLocal() {
return MemDepResult(PairTy(0, NonLocal));
}
/// isClobber - Return true if this MemDepResult represents a query that is
/// a instruction clobber dependency.
bool isClobber() const { return Value.getInt() == Clobber; }
/// isDef - Return true if this MemDepResult represents a query that is
/// a instruction definition dependency.
bool isDef() const { return Value.getInt() == Def; }
/// isNonLocal - Return true if this MemDepResult represents a query that
/// is transparent to the start of the block, but where a non-local hasn't
/// been done.
bool isNonLocal() const { return Value.getInt() == NonLocal; }
/// getInst() - If this is a normal dependency, return the instruction that
/// is depended on. Otherwise, return null.
Instruction *getInst() const { return Value.getPointer(); }
bool operator==(const MemDepResult &M) const { return Value == M.Value; }
bool operator!=(const MemDepResult &M) const { return Value != M.Value; }
bool operator<(const MemDepResult &M) const { return Value < M.Value; }
bool operator>(const MemDepResult &M) const { return Value > M.Value; }
private:
friend class MemoryDependenceAnalysis;
/// Dirty - Entries with this marker occur in a LocalDeps map or
/// NonLocalDeps map when the instruction they previously referenced was
/// removed from MemDep. In either case, the entry may include an
/// instruction pointer. If so, the pointer is an instruction in the
/// block where scanning can start from, saving some work.
///
/// In a default-constructed MemDepResult object, the type will be Dirty
/// and the instruction pointer will be null.
///
/// isDirty - Return true if this is a MemDepResult in its dirty/invalid.
/// state.
bool isDirty() const { return Value.getInt() == Invalid; }
static MemDepResult getDirty(Instruction *Inst) {
return MemDepResult(PairTy(Inst, Invalid));
}
};
/// MemoryDependenceAnalysis - This is an analysis that determines, for a
/// given memory operation, what preceding memory operations it depends on.
/// It builds on alias analysis information, and tries to provide a lazy,
/// caching interface to a common kind of alias information query.
///
/// The dependency information returned is somewhat unusual, but is pragmatic.
/// If queried about a store or call that might modify memory, the analysis
/// will return the instruction[s] that may either load from that memory or
/// store to it. If queried with a load or call that can never modify memory,
/// the analysis will return calls and stores that might modify the pointer,
/// but generally does not return loads unless a) they are volatile, or
/// b) they load from *must-aliased* pointers. Returning a dependence on
/// must-alias'd pointers instead of all pointers interacts well with the
/// internal caching mechanism.
///
class MemoryDependenceAnalysis : public FunctionPass {
// A map from instructions to their dependency.
typedef DenseMap<Instruction*, MemDepResult> LocalDepMapType;
LocalDepMapType LocalDeps;
public:
typedef std::pair<BasicBlock*, MemDepResult> NonLocalDepEntry;
typedef std::vector<NonLocalDepEntry> NonLocalDepInfo;
private:
/// ValueIsLoadPair - This is a pair<Value*, bool> where the bool is true if
/// the dependence is a read only dependence, false if read/write.
typedef PointerIntPair<Value*, 1, bool> ValueIsLoadPair;
Implement initial support for PHI translation in memdep. This means that memdep keeps track of how PHIs affect the pointer in dep queries, which allows it to eliminate the load in cases like rle-phi-translate.ll, which basically end up being: BB1: X = load P br BB3 BB2: Y = load Q br BB3 BB3: R = phi [P] [Q] load R turning "load R" into a phi of X/Y. In addition to additional exposed opportunities, this makes memdep safe in many cases that it wasn't before (which is required for load PRE) and also makes it substantially more efficient. For example, consider: bb1: // has many predecessors. P = some_operator() load P In this example, previously memdep would scan all the predecessors of BB1 to see if they had something that would mustalias P. In some cases (e.g. test/Transforms/GVN/rle-must-alias.ll) it would actually find them and end up eliminating something. In many other cases though, it would scan and not find anything useful. MemDep now stops at a block if the pointer is defined in that block and cannot be phi translated to predecessors. This causes it to miss the (rare) cases like rle-must-alias.ll, but makes it faster by not scanning tons of stuff that is unlikely to be useful. For example, this speeds up GVN as a whole from 3.928s to 2.448s (60%)!. IMO, scalar GVN should be enhanced to simplify the rle-must-alias pointer base anyway, which would allow the loads to be eliminated. In the future, this should be enhanced to phi translate through geps and bitcasts as well (as indicated by FIXMEs) making memdep even more powerful. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61022 91177308-0d34-0410-b5e6-96231b3b80d8
2008-12-15 03:35:32 +00:00
/// BBSkipFirstBlockPair - This pair is used when caching information for a
/// block. If the pointer is null, the cache value is not a full query that
/// starts at the specified block. If non-null, the bool indicates whether
/// or not the contents of the block was skipped.
typedef PointerIntPair<BasicBlock*, 1, bool> BBSkipFirstBlockPair;
/// CachedNonLocalPointerInfo - This map stores the cached results of doing
/// a pointer lookup at the bottom of a block. The key of this map is the
/// pointer+isload bit, the value is a list of <bb->result> mappings.
Implement initial support for PHI translation in memdep. This means that memdep keeps track of how PHIs affect the pointer in dep queries, which allows it to eliminate the load in cases like rle-phi-translate.ll, which basically end up being: BB1: X = load P br BB3 BB2: Y = load Q br BB3 BB3: R = phi [P] [Q] load R turning "load R" into a phi of X/Y. In addition to additional exposed opportunities, this makes memdep safe in many cases that it wasn't before (which is required for load PRE) and also makes it substantially more efficient. For example, consider: bb1: // has many predecessors. P = some_operator() load P In this example, previously memdep would scan all the predecessors of BB1 to see if they had something that would mustalias P. In some cases (e.g. test/Transforms/GVN/rle-must-alias.ll) it would actually find them and end up eliminating something. In many other cases though, it would scan and not find anything useful. MemDep now stops at a block if the pointer is defined in that block and cannot be phi translated to predecessors. This causes it to miss the (rare) cases like rle-must-alias.ll, but makes it faster by not scanning tons of stuff that is unlikely to be useful. For example, this speeds up GVN as a whole from 3.928s to 2.448s (60%)!. IMO, scalar GVN should be enhanced to simplify the rle-must-alias pointer base anyway, which would allow the loads to be eliminated. In the future, this should be enhanced to phi translate through geps and bitcasts as well (as indicated by FIXMEs) making memdep even more powerful. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61022 91177308-0d34-0410-b5e6-96231b3b80d8
2008-12-15 03:35:32 +00:00
typedef DenseMap<ValueIsLoadPair, std::pair<BBSkipFirstBlockPair,
NonLocalDepInfo> > CachedNonLocalPointerInfo;
CachedNonLocalPointerInfo NonLocalPointerDeps;
// A map from instructions to their non-local pointer dependencies.
typedef DenseMap<Instruction*,
SmallPtrSet<ValueIsLoadPair, 4> > ReverseNonLocalPtrDepTy;
ReverseNonLocalPtrDepTy ReverseNonLocalPtrDeps;
/// PerInstNLInfo - This is the instruction we keep for each cached access
/// that we have for an instruction. The pointer is an owning pointer and
/// the bool indicates whether we have any dirty bits in the set.
typedef std::pair<NonLocalDepInfo, bool> PerInstNLInfo;
// A map from instructions to their non-local dependencies.
typedef DenseMap<Instruction*, PerInstNLInfo> NonLocalDepMapType;
NonLocalDepMapType NonLocalDeps;
// A reverse mapping from dependencies to the dependees. This is
// used when removing instructions to keep the cache coherent.
typedef DenseMap<Instruction*,
SmallPtrSet<Instruction*, 4> > ReverseDepMapType;
ReverseDepMapType ReverseLocalDeps;
// A reverse mapping form dependencies to the non-local dependees.
ReverseDepMapType ReverseNonLocalDeps;
/// Current AA implementation, just a cache.
AliasAnalysis *AA;
TargetData *TD;
OwningPtr<PredIteratorCache> PredCache;
public:
MemoryDependenceAnalysis();
~MemoryDependenceAnalysis();
static char ID;
/// Pass Implementation stuff. This doesn't do any analysis eagerly.
bool runOnFunction(Function &);
/// Clean up memory in between runs
void releaseMemory();
/// getAnalysisUsage - Does not modify anything. It uses Value Numbering
/// and Alias Analysis.
///
virtual void getAnalysisUsage(AnalysisUsage &AU) const;
/// getDependency - Return the instruction on which a memory operation
/// depends. See the class comment for more details. It is illegal to call
/// this on non-memory instructions.
MemDepResult getDependency(Instruction *QueryInst);
/// getNonLocalCallDependency - Perform a full dependency query for the
/// specified call, returning the set of blocks that the value is
/// potentially live across. The returned set of results will include a
/// "NonLocal" result for all blocks where the value is live across.
///
/// This method assumes the instruction returns a "NonLocal" dependency
/// within its own block.
///
/// This returns a reference to an internal data structure that may be
/// invalidated on the next non-local query or when an instruction is
/// removed. Clients must copy this data if they want it around longer than
/// that.
const NonLocalDepInfo &getNonLocalCallDependency(CallSite QueryCS);
/// getNonLocalPointerDependency - Perform a full dependency query for an
/// access to the specified (non-volatile) memory location, returning the
/// set of instructions that either define or clobber the value.
///
/// This method assumes the pointer has a "NonLocal" dependency within BB.
void getNonLocalPointerDependency(Value *Pointer, bool isLoad,
BasicBlock *BB,
SmallVectorImpl<NonLocalDepEntry> &Result);
/// GetPHITranslatedValue - Find an available version of the specified value
/// PHI translated across the specified edge. If MemDep isn't able to
/// satisfy this request, it returns null.
Value *GetPHITranslatedValue(Value *V,
BasicBlock *CurBB, BasicBlock *PredBB,
const TargetData *TD) const;
/// GetAvailablePHITranslatedValue - Return the value computed by
/// PHITranslatePointer if it dominates PredBB, otherwise return null.
Value *GetAvailablePHITranslatedValue(Value *V,
BasicBlock *CurBB, BasicBlock *PredBB,
const TargetData *TD,
const DominatorTree &DT) const;
/// InsertPHITranslatedPointer - Insert a computation of the PHI translated
/// version of 'V' for the edge PredBB->CurBB into the end of the PredBB
/// block. All newly created instructions are added to the NewInsts list.
Value *InsertPHITranslatedPointer(Value *V,
BasicBlock *CurBB, BasicBlock *PredBB,
const TargetData *TD,
const DominatorTree &DT,
SmallVectorImpl<Instruction*> &NewInsts) const;
/// removeInstruction - Remove an instruction from the dependence analysis,
/// updating the dependence of instructions that previously depended on it.
void removeInstruction(Instruction *InstToRemove);
/// invalidateCachedPointerInfo - This method is used to invalidate cached
/// information about the specified pointer, because it may be too
/// conservative in memdep. This is an optional call that can be used when
/// the client detects an equivalence between the pointer and some other
/// value and replaces the other value with ptr. This can make Ptr available
/// in more places that cached info does not necessarily keep.
void invalidateCachedPointerInfo(Value *Ptr);
private:
MemDepResult getPointerDependencyFrom(Value *Pointer, uint64_t MemSize,
bool isLoad,
BasicBlock::iterator ScanIt,
BasicBlock *BB);
MemDepResult getCallSiteDependencyFrom(CallSite C, bool isReadOnlyCall,
BasicBlock::iterator ScanIt,
BasicBlock *BB);
Implement initial support for PHI translation in memdep. This means that memdep keeps track of how PHIs affect the pointer in dep queries, which allows it to eliminate the load in cases like rle-phi-translate.ll, which basically end up being: BB1: X = load P br BB3 BB2: Y = load Q br BB3 BB3: R = phi [P] [Q] load R turning "load R" into a phi of X/Y. In addition to additional exposed opportunities, this makes memdep safe in many cases that it wasn't before (which is required for load PRE) and also makes it substantially more efficient. For example, consider: bb1: // has many predecessors. P = some_operator() load P In this example, previously memdep would scan all the predecessors of BB1 to see if they had something that would mustalias P. In some cases (e.g. test/Transforms/GVN/rle-must-alias.ll) it would actually find them and end up eliminating something. In many other cases though, it would scan and not find anything useful. MemDep now stops at a block if the pointer is defined in that block and cannot be phi translated to predecessors. This causes it to miss the (rare) cases like rle-must-alias.ll, but makes it faster by not scanning tons of stuff that is unlikely to be useful. For example, this speeds up GVN as a whole from 3.928s to 2.448s (60%)!. IMO, scalar GVN should be enhanced to simplify the rle-must-alias pointer base anyway, which would allow the loads to be eliminated. In the future, this should be enhanced to phi translate through geps and bitcasts as well (as indicated by FIXMEs) making memdep even more powerful. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61022 91177308-0d34-0410-b5e6-96231b3b80d8
2008-12-15 03:35:32 +00:00
bool getNonLocalPointerDepFromBB(Value *Pointer, uint64_t Size,
bool isLoad, BasicBlock *BB,
SmallVectorImpl<NonLocalDepEntry> &Result,
Implement initial support for PHI translation in memdep. This means that memdep keeps track of how PHIs affect the pointer in dep queries, which allows it to eliminate the load in cases like rle-phi-translate.ll, which basically end up being: BB1: X = load P br BB3 BB2: Y = load Q br BB3 BB3: R = phi [P] [Q] load R turning "load R" into a phi of X/Y. In addition to additional exposed opportunities, this makes memdep safe in many cases that it wasn't before (which is required for load PRE) and also makes it substantially more efficient. For example, consider: bb1: // has many predecessors. P = some_operator() load P In this example, previously memdep would scan all the predecessors of BB1 to see if they had something that would mustalias P. In some cases (e.g. test/Transforms/GVN/rle-must-alias.ll) it would actually find them and end up eliminating something. In many other cases though, it would scan and not find anything useful. MemDep now stops at a block if the pointer is defined in that block and cannot be phi translated to predecessors. This causes it to miss the (rare) cases like rle-must-alias.ll, but makes it faster by not scanning tons of stuff that is unlikely to be useful. For example, this speeds up GVN as a whole from 3.928s to 2.448s (60%)!. IMO, scalar GVN should be enhanced to simplify the rle-must-alias pointer base anyway, which would allow the loads to be eliminated. In the future, this should be enhanced to phi translate through geps and bitcasts as well (as indicated by FIXMEs) making memdep even more powerful. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61022 91177308-0d34-0410-b5e6-96231b3b80d8
2008-12-15 03:35:32 +00:00
DenseMap<BasicBlock*, Value*> &Visited,
bool SkipFirstBlock = false);
MemDepResult GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
bool isLoad, BasicBlock *BB,
NonLocalDepInfo *Cache,
unsigned NumSortedEntries);
void RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P);
/// verifyRemoved - Verify that the specified instruction does not occur
/// in our internal data structures.
void verifyRemoved(Instruction *Inst) const;
};
} // End llvm namespace
#endif