2007-07-06 23:14:35 +00:00
|
|
|
//===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation --*- C++ -*-===//
|
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
2007-12-29 20:36:04 +00:00
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
2007-07-06 23:14:35 +00:00
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This file implements an analysis that determines, for a given memory
|
|
|
|
// operation, what preceding memory operations it depends on. It builds on
|
2007-08-08 22:01:54 +00:00
|
|
|
// alias analysis information, and tries to provide a lazy, caching interface to
|
2007-07-06 23:14:35 +00:00
|
|
|
// a common kind of alias information query.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2008-11-28 21:45:17 +00:00
|
|
|
#define DEBUG_TYPE "memdep"
|
2007-07-06 23:14:35 +00:00
|
|
|
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
|
2007-07-10 17:25:03 +00:00
|
|
|
#include "llvm/Constants.h"
|
2007-07-06 23:14:35 +00:00
|
|
|
#include "llvm/Instructions.h"
|
|
|
|
#include "llvm/Function.h"
|
|
|
|
#include "llvm/Analysis/AliasAnalysis.h"
|
2008-11-28 22:28:27 +00:00
|
|
|
#include "llvm/ADT/Statistic.h"
|
|
|
|
#include "llvm/ADT/STLExtras.h"
|
2007-07-24 21:52:37 +00:00
|
|
|
#include "llvm/Support/CFG.h"
|
2008-02-06 00:54:55 +00:00
|
|
|
#include "llvm/Support/CommandLine.h"
|
2008-11-28 21:45:17 +00:00
|
|
|
#include "llvm/Support/Debug.h"
|
2007-07-06 23:14:35 +00:00
|
|
|
#include "llvm/Target/TargetData.h"
|
2007-09-09 21:43:49 +00:00
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
using namespace llvm;
|
|
|
|
|
2008-05-13 00:00:25 +00:00
|
|
|
// Control the calculation of non-local dependencies by only examining the
|
|
|
|
// predecessors if the basic block has less than X amount (50 by default).
|
|
|
|
static cl::opt<int>
|
|
|
|
PredLimit("nonlocaldep-threshold", cl::Hidden, cl::init(50),
|
|
|
|
cl::desc("Control the calculation of non-local"
|
|
|
|
"dependencies (default = 50)"));
|
2008-02-06 00:54:55 +00:00
|
|
|
|
2007-09-09 21:43:49 +00:00
|
|
|
STATISTIC(NumCacheNonlocal, "Number of cached non-local responses");
|
|
|
|
STATISTIC(NumUncacheNonlocal, "Number of uncached non-local responses");
|
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
char MemoryDependenceAnalysis::ID = 0;
|
|
|
|
|
|
|
|
// Register this pass...
|
2007-07-10 20:21:08 +00:00
|
|
|
static RegisterPass<MemoryDependenceAnalysis> X("memdep",
|
2008-11-28 21:45:17 +00:00
|
|
|
"Memory Dependence Analysis", false, true);
|
2007-07-06 23:14:35 +00:00
|
|
|
|
2008-11-28 21:45:17 +00:00
|
|
|
/// verifyRemoved - Verify that the specified instruction does not occur
|
|
|
|
/// in our internal data structures.
|
2008-11-28 21:42:09 +00:00
|
|
|
void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
|
2008-11-29 01:43:36 +00:00
|
|
|
for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
|
|
|
|
E = LocalDeps.end(); I != E; ++I) {
|
2008-11-28 21:45:17 +00:00
|
|
|
assert(I->first != D && "Inst occurs in data structures");
|
2008-11-29 01:43:36 +00:00
|
|
|
assert(I->second.first.getPointer() != D &&
|
|
|
|
"Inst occurs in data structures");
|
2007-12-08 01:37:09 +00:00
|
|
|
}
|
|
|
|
|
2008-11-28 21:42:09 +00:00
|
|
|
for (nonLocalDepMapType::const_iterator I = depGraphNonLocal.begin(),
|
|
|
|
E = depGraphNonLocal.end(); I != E; ++I) {
|
2008-11-28 21:45:17 +00:00
|
|
|
assert(I->first != D && "Inst occurs in data structures");
|
2008-11-29 01:43:36 +00:00
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator II = I->second.begin(),
|
2008-06-01 20:51:41 +00:00
|
|
|
EE = I->second.end(); II != EE; ++II)
|
2008-11-29 01:43:36 +00:00
|
|
|
assert(II->second.getPointer() != D && "Inst occurs in data structures");
|
2007-12-08 01:37:09 +00:00
|
|
|
}
|
|
|
|
|
2008-11-28 21:42:09 +00:00
|
|
|
for (reverseDepMapType::const_iterator I = reverseDep.begin(),
|
|
|
|
E = reverseDep.end(); I != E; ++I)
|
|
|
|
for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
|
|
|
|
EE = I->second.end(); II != EE; ++II)
|
2008-11-28 21:45:17 +00:00
|
|
|
assert(*II != D && "Inst occurs in data structures");
|
2007-12-08 01:37:09 +00:00
|
|
|
|
2008-11-28 21:42:09 +00:00
|
|
|
for (reverseDepMapType::const_iterator I = reverseDepNonLocal.begin(),
|
|
|
|
E = reverseDepNonLocal.end();
|
2007-12-08 01:37:09 +00:00
|
|
|
I != E; ++I)
|
2008-11-28 21:42:09 +00:00
|
|
|
for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
|
|
|
|
EE = I->second.end(); II != EE; ++II)
|
2008-11-28 21:45:17 +00:00
|
|
|
assert(*II != D && "Inst occurs in data structures");
|
2007-12-08 01:37:09 +00:00
|
|
|
}
|
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
|
|
|
|
///
|
|
|
|
void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
|
|
|
|
AU.setPreservesAll();
|
|
|
|
AU.addRequiredTransitive<AliasAnalysis>();
|
|
|
|
AU.addRequiredTransitive<TargetData>();
|
|
|
|
}
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
/// getCallSiteDependency - Private helper for finding the local dependencies
|
|
|
|
/// of a call site.
|
2008-11-29 02:29:27 +00:00
|
|
|
MemDepResult MemoryDependenceAnalysis::
|
2008-11-29 01:43:36 +00:00
|
|
|
getCallSiteDependency(CallSite C, Instruction *start, BasicBlock *block) {
|
|
|
|
std::pair<DepResultTy, bool> &cachedResult = LocalDeps[C.getInstruction()];
|
2007-07-10 17:59:22 +00:00
|
|
|
AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
|
|
|
|
TargetData& TD = getAnalysis<TargetData>();
|
|
|
|
BasicBlock::iterator blockBegin = C.getInstruction()->getParent()->begin();
|
|
|
|
BasicBlock::iterator QI = C.getInstruction();
|
|
|
|
|
2008-09-11 19:41:10 +00:00
|
|
|
// If the starting point was specified, use it
|
2007-08-07 00:33:45 +00:00
|
|
|
if (start) {
|
|
|
|
QI = start;
|
2008-03-31 22:08:00 +00:00
|
|
|
blockBegin = start->getParent()->begin();
|
2007-08-08 22:26:03 +00:00
|
|
|
// If the starting point wasn't specified, but the block was, use it
|
2007-08-07 00:33:45 +00:00
|
|
|
} else if (!start && block) {
|
|
|
|
QI = block->end();
|
2008-03-31 22:08:00 +00:00
|
|
|
blockBegin = block->begin();
|
2007-08-07 00:33:45 +00:00
|
|
|
}
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// Walk backwards through the block, looking for dependencies
|
2007-07-10 17:59:22 +00:00
|
|
|
while (QI != blockBegin) {
|
|
|
|
--QI;
|
|
|
|
|
|
|
|
// If this inst is a memory op, get the pointer it accessed
|
|
|
|
Value* pointer = 0;
|
|
|
|
uint64_t pointerSize = 0;
|
|
|
|
if (StoreInst* S = dyn_cast<StoreInst>(QI)) {
|
|
|
|
pointer = S->getPointerOperand();
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
pointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
|
2007-07-10 17:59:22 +00:00
|
|
|
} else if (AllocationInst* AI = dyn_cast<AllocationInst>(QI)) {
|
|
|
|
pointer = AI;
|
|
|
|
if (ConstantInt* C = dyn_cast<ConstantInt>(AI->getArraySize()))
|
2008-11-28 21:16:44 +00:00
|
|
|
pointerSize = C->getZExtValue() *
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
TD.getABITypeSize(AI->getAllocatedType());
|
2007-07-10 17:59:22 +00:00
|
|
|
else
|
|
|
|
pointerSize = ~0UL;
|
2007-07-10 18:43:15 +00:00
|
|
|
} else if (VAArgInst* V = dyn_cast<VAArgInst>(QI)) {
|
|
|
|
pointer = V->getOperand(0);
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
pointerSize = TD.getTypeStoreSize(V->getType());
|
2007-07-10 17:59:22 +00:00
|
|
|
} else if (FreeInst* F = dyn_cast<FreeInst>(QI)) {
|
|
|
|
pointer = F->getPointerOperand();
|
|
|
|
|
|
|
|
// FreeInsts erase the entire structure
|
|
|
|
pointerSize = ~0UL;
|
2008-05-13 21:25:37 +00:00
|
|
|
} else if (CallSite::get(QI).getInstruction() != 0) {
|
2007-11-26 02:26:36 +00:00
|
|
|
AliasAnalysis::ModRefBehavior result =
|
2007-12-01 07:51:45 +00:00
|
|
|
AA.getModRefBehavior(CallSite::get(QI));
|
2008-04-17 05:36:50 +00:00
|
|
|
if (result != AliasAnalysis::DoesNotAccessMemory) {
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2008-11-29 01:43:36 +00:00
|
|
|
cachedResult.first = DepResultTy(QI, Normal);
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.second = true;
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[DepResultTy(QI, Normal)].insert(C.getInstruction());
|
2007-08-07 00:33:45 +00:00
|
|
|
}
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::get(QI);
|
2007-07-10 17:59:22 +00:00
|
|
|
} else {
|
|
|
|
continue;
|
|
|
|
}
|
2007-07-10 20:39:07 +00:00
|
|
|
} else
|
|
|
|
continue;
|
2007-07-10 17:59:22 +00:00
|
|
|
|
|
|
|
if (AA.getModRefInfo(C, pointer, pointerSize) != AliasAnalysis::NoModRef) {
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2008-11-29 01:43:36 +00:00
|
|
|
cachedResult.first = DepResultTy(QI, Normal);
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.second = true;
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[DepResultTy(QI, Normal)].insert(C.getInstruction());
|
2007-08-07 00:33:45 +00:00
|
|
|
}
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::get(QI);
|
2007-07-10 17:59:22 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// No dependence found
|
2008-11-29 01:43:36 +00:00
|
|
|
cachedResult.first = DepResultTy(0, NonLocal);
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.second = true;
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[DepResultTy(0, NonLocal)].insert(C.getInstruction());
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::getNonLocal();
|
2007-07-10 17:59:22 +00:00
|
|
|
}
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
/// nonLocalHelper - Private helper used to calculate non-local dependencies
|
2008-11-29 01:43:36 +00:00
|
|
|
/// by doing DFS on the predecessors of a block to find its dependencies.
|
2007-08-01 22:01:54 +00:00
|
|
|
void MemoryDependenceAnalysis::nonLocalHelper(Instruction* query,
|
2007-07-25 19:57:03 +00:00
|
|
|
BasicBlock* block,
|
2008-11-29 01:43:36 +00:00
|
|
|
DenseMap<BasicBlock*, DepResultTy> &resp) {
|
2007-08-08 22:26:03 +00:00
|
|
|
// Set of blocks that we've already visited in our DFS
|
2007-08-01 22:01:54 +00:00
|
|
|
SmallPtrSet<BasicBlock*, 4> visited;
|
2007-09-21 03:53:52 +00:00
|
|
|
// If we're updating a dirtied cache entry, we don't need to reprocess
|
|
|
|
// already computed entries.
|
2008-11-29 01:43:36 +00:00
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator I = resp.begin(),
|
2007-09-21 03:53:52 +00:00
|
|
|
E = resp.end(); I != E; ++I)
|
2008-11-29 01:43:36 +00:00
|
|
|
if (I->second.getInt() != Dirty)
|
2007-09-21 03:53:52 +00:00
|
|
|
visited.insert(I->first);
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// Current stack of the DFS
|
2007-08-01 22:01:54 +00:00
|
|
|
SmallVector<BasicBlock*, 4> stack;
|
2008-04-10 22:13:32 +00:00
|
|
|
for (pred_iterator PI = pred_begin(block), PE = pred_end(block);
|
|
|
|
PI != PE; ++PI)
|
|
|
|
stack.push_back(*PI);
|
2007-07-24 21:52:37 +00:00
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// Do a basic DFS
|
2007-08-01 22:01:54 +00:00
|
|
|
while (!stack.empty()) {
|
|
|
|
BasicBlock* BB = stack.back();
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// If we've already visited this block, no need to revist
|
2007-08-02 17:56:05 +00:00
|
|
|
if (visited.count(BB)) {
|
2007-08-01 22:01:54 +00:00
|
|
|
stack.pop_back();
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// If we find a new block with a local dependency for query,
|
|
|
|
// then we insert the new dependency and backtrack.
|
2007-08-01 22:01:54 +00:00
|
|
|
if (BB != block) {
|
2007-08-02 17:56:05 +00:00
|
|
|
visited.insert(BB);
|
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
MemDepResult localDep = getDependency(query, 0, BB);
|
|
|
|
if (!localDep.isNonLocal()) {
|
|
|
|
resp.insert(std::make_pair(BB, ConvFromResult(localDep)));
|
2007-08-02 17:56:05 +00:00
|
|
|
stack.pop_back();
|
2007-08-01 22:01:54 +00:00
|
|
|
continue;
|
|
|
|
}
|
2007-08-08 22:26:03 +00:00
|
|
|
// If we re-encounter the starting block, we still need to search it
|
|
|
|
// because there might be a dependency in the starting block AFTER
|
|
|
|
// the position of the query. This is necessary to get loops right.
|
2008-04-10 22:13:32 +00:00
|
|
|
} else if (BB == block) {
|
2007-08-02 17:56:05 +00:00
|
|
|
visited.insert(BB);
|
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
MemDepResult localDep = getDependency(query, 0, BB);
|
|
|
|
if (localDep.getInst() != query)
|
|
|
|
resp.insert(std::make_pair(BB, ConvFromResult(localDep)));
|
2007-08-02 17:56:05 +00:00
|
|
|
|
|
|
|
stack.pop_back();
|
|
|
|
continue;
|
2007-08-01 22:01:54 +00:00
|
|
|
}
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// If we didn't find anything, recurse on the precessors of this block
|
2008-02-06 00:54:55 +00:00
|
|
|
// Only do this for blocks with a small number of predecessors.
|
2007-08-01 22:01:54 +00:00
|
|
|
bool predOnStack = false;
|
|
|
|
bool inserted = false;
|
2008-02-06 00:54:55 +00:00
|
|
|
if (std::distance(pred_begin(BB), pred_end(BB)) <= PredLimit) {
|
|
|
|
for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
|
|
|
|
PI != PE; ++PI)
|
|
|
|
if (!visited.count(*PI)) {
|
|
|
|
stack.push_back(*PI);
|
|
|
|
inserted = true;
|
|
|
|
} else
|
|
|
|
predOnStack = true;
|
|
|
|
}
|
2007-08-01 22:01:54 +00:00
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// If we inserted a new predecessor, then we'll come back to this block
|
2007-08-01 22:01:54 +00:00
|
|
|
if (inserted)
|
|
|
|
continue;
|
2007-08-08 22:26:03 +00:00
|
|
|
// If we didn't insert because we have no predecessors, then this
|
|
|
|
// query has no dependency at all.
|
2007-08-01 22:01:54 +00:00
|
|
|
else if (!inserted && !predOnStack) {
|
2008-11-29 01:43:36 +00:00
|
|
|
resp.insert(std::make_pair(BB, DepResultTy(0, None)));
|
2007-08-08 22:26:03 +00:00
|
|
|
// If we didn't insert because our predecessors are already on the stack,
|
|
|
|
// then we might still have a dependency, but it will be discovered during
|
|
|
|
// backtracking.
|
2007-08-01 22:01:54 +00:00
|
|
|
} else if (!inserted && predOnStack){
|
2008-11-29 01:43:36 +00:00
|
|
|
resp.insert(std::make_pair(BB, DepResultTy(0, NonLocal)));
|
2007-08-01 22:01:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
stack.pop_back();
|
2007-07-24 21:52:37 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2007-08-08 22:01:54 +00:00
|
|
|
/// getNonLocalDependency - Fills the passed-in map with the non-local
|
|
|
|
/// dependencies of the queries. The map will contain NonLocal for
|
|
|
|
/// blocks between the query and its dependencies.
|
2007-08-01 22:01:54 +00:00
|
|
|
void MemoryDependenceAnalysis::getNonLocalDependency(Instruction* query,
|
2008-11-29 02:29:27 +00:00
|
|
|
DenseMap<BasicBlock*, MemDepResult> &resp) {
|
2007-08-16 21:27:05 +00:00
|
|
|
if (depGraphNonLocal.count(query)) {
|
2008-11-29 01:43:36 +00:00
|
|
|
DenseMap<BasicBlock*, DepResultTy> &cached = depGraphNonLocal[query];
|
2007-09-09 21:43:49 +00:00
|
|
|
NumCacheNonlocal++;
|
2007-09-21 03:53:52 +00:00
|
|
|
|
|
|
|
SmallVector<BasicBlock*, 4> dirtied;
|
2008-11-29 01:43:36 +00:00
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator I = cached.begin(),
|
2007-09-21 03:53:52 +00:00
|
|
|
E = cached.end(); I != E; ++I)
|
2008-11-29 01:43:36 +00:00
|
|
|
if (I->second.getInt() == Dirty)
|
2007-09-21 03:53:52 +00:00
|
|
|
dirtied.push_back(I->first);
|
|
|
|
|
|
|
|
for (SmallVector<BasicBlock*, 4>::iterator I = dirtied.begin(),
|
|
|
|
E = dirtied.end(); I != E; ++I) {
|
2008-11-29 02:29:27 +00:00
|
|
|
MemDepResult localDep = getDependency(query, 0, *I);
|
|
|
|
if (!localDep.isNonLocal())
|
|
|
|
cached[*I] = ConvFromResult(localDep);
|
2007-09-21 03:53:52 +00:00
|
|
|
else {
|
|
|
|
cached.erase(*I);
|
|
|
|
nonLocalHelper(query, *I, cached);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
// Update the reverse non-local dependency cache.
|
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator I = cached.begin(),
|
|
|
|
E = cached.end(); I != E; ++I) {
|
2008-06-01 21:03:52 +00:00
|
|
|
reverseDepNonLocal[I->second].insert(query);
|
2008-11-29 02:29:27 +00:00
|
|
|
resp[I->first] = ConvToResult(I->second);
|
|
|
|
}
|
2008-06-01 21:03:52 +00:00
|
|
|
|
2007-08-16 21:27:05 +00:00
|
|
|
return;
|
2008-11-29 02:29:27 +00:00
|
|
|
}
|
2007-07-24 21:52:37 +00:00
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
NumUncacheNonlocal++;
|
2007-08-16 21:27:05 +00:00
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
// If not, go ahead and search for non-local deps.
|
|
|
|
DenseMap<BasicBlock*, DepResultTy> &cached = depGraphNonLocal[query];
|
|
|
|
nonLocalHelper(query, query->getParent(), cached);
|
|
|
|
|
2007-08-16 21:27:05 +00:00
|
|
|
// Update the non-local dependency cache
|
2008-11-29 02:29:27 +00:00
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator I = cached.begin(),
|
|
|
|
E = cached.end(); I != E; ++I) {
|
|
|
|
// FIXME: Merge with the code above!
|
2007-08-16 21:27:05 +00:00
|
|
|
reverseDepNonLocal[I->second].insert(query);
|
2008-11-29 02:29:27 +00:00
|
|
|
resp[I->first] = ConvToResult(I->second);
|
2007-08-16 21:27:05 +00:00
|
|
|
}
|
2007-07-24 21:52:37 +00:00
|
|
|
}
|
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
/// getDependency - Return the instruction on which a memory operation
|
2008-04-10 23:02:38 +00:00
|
|
|
/// depends. The local parameter indicates if the query should only
|
2007-07-10 17:08:11 +00:00
|
|
|
/// evaluate dependencies within the same basic block.
|
2008-11-29 02:29:27 +00:00
|
|
|
MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *query,
|
|
|
|
Instruction *start,
|
|
|
|
BasicBlock *block) {
|
2007-07-06 23:14:35 +00:00
|
|
|
// Start looking for dependencies with the queried inst
|
|
|
|
BasicBlock::iterator QI = query;
|
|
|
|
|
|
|
|
// Check for a cached result
|
2008-11-29 01:43:36 +00:00
|
|
|
std::pair<DepResultTy, bool>& cachedResult = LocalDeps[query];
|
2007-07-06 23:14:35 +00:00
|
|
|
// If we have a _confirmed_ cached entry, return it
|
2007-12-08 01:37:09 +00:00
|
|
|
if (!block && !start) {
|
|
|
|
if (cachedResult.second)
|
2008-11-29 02:29:27 +00:00
|
|
|
return ConvToResult(cachedResult.first);
|
2008-11-29 01:43:36 +00:00
|
|
|
else if (cachedResult.first.getInt() == Normal &&
|
|
|
|
cachedResult.first.getPointer())
|
|
|
|
// If we have an unconfirmed cached entry, we can start our search from
|
|
|
|
// it.
|
|
|
|
QI = cachedResult.first.getPointer();
|
2007-12-08 01:37:09 +00:00
|
|
|
}
|
2007-07-06 23:14:35 +00:00
|
|
|
|
2007-07-16 21:52:50 +00:00
|
|
|
if (start)
|
|
|
|
QI = start;
|
2007-07-25 19:57:03 +00:00
|
|
|
else if (!start && block)
|
|
|
|
QI = block->end();
|
2007-07-16 21:52:50 +00:00
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
|
2007-07-10 17:08:11 +00:00
|
|
|
TargetData& TD = getAnalysis<TargetData>();
|
2007-07-06 23:14:35 +00:00
|
|
|
|
|
|
|
// Get the pointer value for which dependence will be determined
|
|
|
|
Value* dependee = 0;
|
2007-07-10 17:08:11 +00:00
|
|
|
uint64_t dependeeSize = 0;
|
2007-07-10 18:11:42 +00:00
|
|
|
bool queryIsVolatile = false;
|
2007-07-16 21:52:50 +00:00
|
|
|
if (StoreInst* S = dyn_cast<StoreInst>(query)) {
|
2007-07-06 23:14:35 +00:00
|
|
|
dependee = S->getPointerOperand();
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
dependeeSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
|
2007-07-10 18:11:42 +00:00
|
|
|
queryIsVolatile = S->isVolatile();
|
2007-07-16 21:52:50 +00:00
|
|
|
} else if (LoadInst* L = dyn_cast<LoadInst>(query)) {
|
2007-07-06 23:14:35 +00:00
|
|
|
dependee = L->getPointerOperand();
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
dependeeSize = TD.getTypeStoreSize(L->getType());
|
2007-07-10 18:11:42 +00:00
|
|
|
queryIsVolatile = L->isVolatile();
|
2007-07-16 21:52:50 +00:00
|
|
|
} else if (VAArgInst* V = dyn_cast<VAArgInst>(query)) {
|
2007-07-10 18:43:15 +00:00
|
|
|
dependee = V->getOperand(0);
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
dependeeSize = TD.getTypeStoreSize(V->getType());
|
2007-07-16 21:52:50 +00:00
|
|
|
} else if (FreeInst* F = dyn_cast<FreeInst>(query)) {
|
2007-07-06 23:14:35 +00:00
|
|
|
dependee = F->getPointerOperand();
|
2007-07-10 17:08:11 +00:00
|
|
|
|
|
|
|
// FreeInsts erase the entire structure, not just a field
|
|
|
|
dependeeSize = ~0UL;
|
2007-07-16 21:52:50 +00:00
|
|
|
} else if (CallSite::get(query).getInstruction() != 0)
|
2007-08-07 00:33:45 +00:00
|
|
|
return getCallSiteDependency(CallSite::get(query), start, block);
|
2007-07-10 17:59:22 +00:00
|
|
|
else if (isa<AllocationInst>(query))
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::getNone();
|
2007-07-10 17:25:03 +00:00
|
|
|
else
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::getNone();
|
2007-07-06 23:14:35 +00:00
|
|
|
|
2007-07-24 21:52:37 +00:00
|
|
|
BasicBlock::iterator blockBegin = block ? block->begin()
|
|
|
|
: query->getParent()->begin();
|
2007-07-06 23:14:35 +00:00
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// Walk backwards through the basic block, looking for dependencies
|
2007-07-06 23:14:35 +00:00
|
|
|
while (QI != blockBegin) {
|
2007-07-10 17:08:11 +00:00
|
|
|
--QI;
|
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
// If this inst is a memory op, get the pointer it accessed
|
|
|
|
Value* pointer = 0;
|
2007-07-10 17:08:11 +00:00
|
|
|
uint64_t pointerSize = 0;
|
|
|
|
if (StoreInst* S = dyn_cast<StoreInst>(QI)) {
|
2007-07-10 18:11:42 +00:00
|
|
|
// All volatile loads/stores depend on each other
|
|
|
|
if (queryIsVolatile && S->isVolatile()) {
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2008-11-29 01:43:36 +00:00
|
|
|
cachedResult.first = DepResultTy(S, Normal);
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.second = true;
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[DepResultTy(S, Normal)].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::get(S);
|
2007-07-10 18:11:42 +00:00
|
|
|
}
|
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
pointer = S->getPointerOperand();
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
pointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
|
2007-07-10 17:08:11 +00:00
|
|
|
} else if (LoadInst* L = dyn_cast<LoadInst>(QI)) {
|
2007-07-10 18:11:42 +00:00
|
|
|
// All volatile loads/stores depend on each other
|
|
|
|
if (queryIsVolatile && L->isVolatile()) {
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2008-11-29 01:43:36 +00:00
|
|
|
cachedResult.first = DepResultTy(L, Normal);
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.second = true;
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[DepResultTy(L, Normal)].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::get(L);
|
2007-07-10 18:11:42 +00:00
|
|
|
}
|
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
pointer = L->getPointerOperand();
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
pointerSize = TD.getTypeStoreSize(L->getType());
|
2007-07-10 17:08:11 +00:00
|
|
|
} else if (AllocationInst* AI = dyn_cast<AllocationInst>(QI)) {
|
|
|
|
pointer = AI;
|
2007-07-10 17:25:03 +00:00
|
|
|
if (ConstantInt* C = dyn_cast<ConstantInt>(AI->getArraySize()))
|
2008-11-28 21:16:44 +00:00
|
|
|
pointerSize = C->getZExtValue() *
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
TD.getABITypeSize(AI->getAllocatedType());
|
2007-07-10 17:08:11 +00:00
|
|
|
else
|
|
|
|
pointerSize = ~0UL;
|
2007-07-10 18:43:15 +00:00
|
|
|
} else if (VAArgInst* V = dyn_cast<VAArgInst>(QI)) {
|
|
|
|
pointer = V->getOperand(0);
|
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-01 20:53:16 +00:00
|
|
|
pointerSize = TD.getTypeStoreSize(V->getType());
|
2007-07-10 17:08:11 +00:00
|
|
|
} else if (FreeInst* F = dyn_cast<FreeInst>(QI)) {
|
2007-07-06 23:14:35 +00:00
|
|
|
pointer = F->getPointerOperand();
|
2007-07-10 17:08:11 +00:00
|
|
|
|
|
|
|
// FreeInsts erase the entire structure
|
|
|
|
pointerSize = ~0UL;
|
2007-07-10 17:59:22 +00:00
|
|
|
} else if (CallSite::get(QI).getInstruction() != 0) {
|
2007-08-08 22:01:54 +00:00
|
|
|
// Call insts need special handling. Check if they can modify our pointer
|
2007-08-06 23:26:03 +00:00
|
|
|
AliasAnalysis::ModRefResult MR = AA.getModRefInfo(CallSite::get(QI),
|
2008-11-29 01:43:36 +00:00
|
|
|
dependee, dependeeSize);
|
2007-08-06 23:26:03 +00:00
|
|
|
|
|
|
|
if (MR != AliasAnalysis::NoModRef) {
|
|
|
|
// Loads don't depend on read-only calls
|
|
|
|
if (isa<LoadInst>(query) && MR == AliasAnalysis::Ref)
|
|
|
|
continue;
|
|
|
|
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2008-11-29 01:43:36 +00:00
|
|
|
cachedResult.first = DepResultTy(QI, Normal);
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.second = true;
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[DepResultTy(QI, Normal)].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::get(QI);
|
2007-07-10 17:25:03 +00:00
|
|
|
} else {
|
|
|
|
continue;
|
|
|
|
}
|
2007-07-06 23:14:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// If we found a pointer, check if it could be the same as our pointer
|
|
|
|
if (pointer) {
|
2007-07-10 17:08:11 +00:00
|
|
|
AliasAnalysis::AliasResult R = AA.alias(pointer, pointerSize,
|
|
|
|
dependee, dependeeSize);
|
2007-07-06 23:14:35 +00:00
|
|
|
|
|
|
|
if (R != AliasAnalysis::NoAlias) {
|
2007-08-06 23:26:03 +00:00
|
|
|
// May-alias loads don't depend on each other
|
|
|
|
if (isa<LoadInst>(query) && isa<LoadInst>(QI) &&
|
|
|
|
R == AliasAnalysis::MayAlias)
|
|
|
|
continue;
|
|
|
|
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2008-11-29 01:43:36 +00:00
|
|
|
cachedResult.first = DepResultTy(QI, Normal);
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.second = true;
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[DepResultTy(QI, Normal)].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::get(QI);
|
2007-07-06 23:14:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If we found nothing, return the non-local flag
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2008-11-29 01:43:36 +00:00
|
|
|
cachedResult.first = DepResultTy(0, NonLocal);
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.second = true;
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[DepResultTy(0, NonLocal)].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
2007-07-06 23:14:35 +00:00
|
|
|
|
2008-11-29 02:29:27 +00:00
|
|
|
return MemDepResult::getNonLocal();
|
2007-07-06 23:14:35 +00:00
|
|
|
}
|
|
|
|
|
2008-02-12 21:15:18 +00:00
|
|
|
/// dropInstruction - Remove an instruction from the analysis, making
|
|
|
|
/// absolutely conservative assumptions when updating the cache. This is
|
|
|
|
/// useful, for example when an instruction is changed rather than removed.
|
|
|
|
void MemoryDependenceAnalysis::dropInstruction(Instruction* drop) {
|
2008-11-29 01:43:36 +00:00
|
|
|
LocalDepMapType::iterator depGraphEntry = LocalDeps.find(drop);
|
|
|
|
if (depGraphEntry != LocalDeps.end())
|
2008-02-12 21:15:18 +00:00
|
|
|
reverseDep[depGraphEntry->second.first].erase(drop);
|
|
|
|
|
|
|
|
// Drop dependency information for things that depended on this instr
|
2008-11-29 01:43:36 +00:00
|
|
|
SmallPtrSet<Instruction*, 4>& set = reverseDep[DepResultTy(drop, Normal)];
|
2008-02-12 21:15:18 +00:00
|
|
|
for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
|
|
|
|
I != E; ++I)
|
2008-11-29 01:43:36 +00:00
|
|
|
LocalDeps.erase(*I);
|
2008-02-12 21:15:18 +00:00
|
|
|
|
2008-11-29 01:43:36 +00:00
|
|
|
LocalDeps.erase(drop);
|
|
|
|
reverseDep.erase(DepResultTy(drop, Normal));
|
2008-02-12 21:15:18 +00:00
|
|
|
|
2008-11-29 01:43:36 +00:00
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
|
|
|
|
depGraphNonLocal[drop].begin(), DE = depGraphNonLocal[drop].end();
|
2008-02-12 21:15:18 +00:00
|
|
|
DI != DE; ++DI)
|
2008-11-29 01:43:36 +00:00
|
|
|
if (DI->second.getInt() != None)
|
2008-02-12 21:15:18 +00:00
|
|
|
reverseDepNonLocal[DI->second].erase(drop);
|
|
|
|
|
2008-11-29 01:43:36 +00:00
|
|
|
if (reverseDepNonLocal.count(DepResultTy(drop, Normal))) {
|
|
|
|
SmallPtrSet<Instruction*, 4>& set =
|
|
|
|
reverseDepNonLocal[DepResultTy(drop, Normal)];
|
2008-02-12 21:15:18 +00:00
|
|
|
for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
|
|
|
|
I != E; ++I)
|
2008-11-29 01:43:36 +00:00
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
|
2008-02-12 21:15:18 +00:00
|
|
|
depGraphNonLocal[*I].begin(), DE = depGraphNonLocal[*I].end();
|
|
|
|
DI != DE; ++DI)
|
2008-11-29 01:43:36 +00:00
|
|
|
if (DI->second == DepResultTy(drop, Normal))
|
|
|
|
DI->second = DepResultTy(0, Dirty);
|
2008-02-12 21:15:18 +00:00
|
|
|
}
|
|
|
|
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDepNonLocal.erase(DepResultTy(drop, Normal));
|
|
|
|
depGraphNonLocal.erase(drop);
|
2008-02-12 21:15:18 +00:00
|
|
|
}
|
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
/// removeInstruction - Remove an instruction from the dependence analysis,
|
|
|
|
/// updating the dependence of instructions that previously depended on it.
|
2007-08-08 22:26:03 +00:00
|
|
|
/// This method attempts to keep the cache coherent using the reverse map.
|
2008-11-28 22:04:47 +00:00
|
|
|
void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
|
|
|
|
// Walk through the Non-local dependencies, removing this one as the value
|
|
|
|
// for any cached queries.
|
2008-11-29 01:43:36 +00:00
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
|
2008-11-28 22:04:47 +00:00
|
|
|
depGraphNonLocal[RemInst].begin(), DE = depGraphNonLocal[RemInst].end();
|
2007-12-08 01:37:09 +00:00
|
|
|
DI != DE; ++DI)
|
2008-11-29 01:43:36 +00:00
|
|
|
if (DI->second.getInt() != None)
|
2008-11-28 22:04:47 +00:00
|
|
|
reverseDepNonLocal[DI->second].erase(RemInst);
|
2007-07-31 20:01:27 +00:00
|
|
|
|
2008-11-28 22:28:27 +00:00
|
|
|
// Shortly after this, we will look for things that depend on RemInst. In
|
|
|
|
// order to update these, we'll need a new dependency to base them on. We
|
|
|
|
// could completely delete any entries that depend on this, but it is better
|
|
|
|
// to make a more accurate approximation where possible. Compute that better
|
|
|
|
// approximation if we can.
|
2008-11-29 01:43:36 +00:00
|
|
|
DepResultTy NewDependency;
|
2008-11-28 22:28:27 +00:00
|
|
|
bool NewDependencyConfirmed = false;
|
|
|
|
|
2008-11-28 22:04:47 +00:00
|
|
|
// If we have a cached local dependence query for this instruction, remove it.
|
2008-11-28 22:28:27 +00:00
|
|
|
//
|
2008-11-29 01:43:36 +00:00
|
|
|
LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
|
|
|
|
if (LocalDepEntry != LocalDeps.end()) {
|
|
|
|
DepResultTy LocalDep = LocalDepEntry->second.first;
|
2008-11-28 22:28:27 +00:00
|
|
|
bool IsConfirmed = LocalDepEntry->second.second;
|
2008-11-28 22:04:47 +00:00
|
|
|
|
2008-11-28 22:28:27 +00:00
|
|
|
// Remove this local dependency info.
|
2008-11-29 01:43:36 +00:00
|
|
|
LocalDeps.erase(LocalDepEntry);
|
2008-01-30 01:24:05 +00:00
|
|
|
|
2008-11-28 22:28:27 +00:00
|
|
|
// Remove us from DepInst's reverse set now that the local dep info is gone.
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep[LocalDep].erase(RemInst);
|
2008-11-28 22:28:27 +00:00
|
|
|
|
|
|
|
// If we have unconfirmed info, don't trust it.
|
|
|
|
if (IsConfirmed) {
|
|
|
|
// If we have a confirmed non-local flag, use it.
|
2008-11-29 01:43:36 +00:00
|
|
|
if (LocalDep.getInt() == NonLocal || LocalDep.getInt() == None) {
|
2008-11-28 22:51:08 +00:00
|
|
|
// The only time this dependency is confirmed is if it is non-local.
|
2008-11-29 01:43:36 +00:00
|
|
|
NewDependency = LocalDep;
|
2008-11-28 22:28:27 +00:00
|
|
|
NewDependencyConfirmed = true;
|
|
|
|
} else {
|
|
|
|
// If we have dep info for RemInst, set them to it.
|
2008-11-29 01:43:36 +00:00
|
|
|
Instruction *NDI = next(BasicBlock::iterator(LocalDep.getPointer()));
|
|
|
|
if (NDI != RemInst) // Don't use RemInst for the new dependency!
|
|
|
|
NewDependency = DepResultTy(NDI, Normal);
|
2008-11-28 22:28:27 +00:00
|
|
|
}
|
2007-07-31 20:01:27 +00:00
|
|
|
}
|
2008-02-05 04:34:03 +00:00
|
|
|
}
|
|
|
|
|
2008-11-28 22:28:27 +00:00
|
|
|
// If we don't already have a local dependency answer for this instruction,
|
|
|
|
// use the immediate successor of RemInst. We use the successor because
|
|
|
|
// getDependence starts by checking the immediate predecessor of what is in
|
|
|
|
// the cache.
|
2008-11-29 01:43:36 +00:00
|
|
|
if (NewDependency == DepResultTy(0, Normal))
|
|
|
|
NewDependency = DepResultTy(next(BasicBlock::iterator(RemInst)), Normal);
|
2008-11-28 22:28:27 +00:00
|
|
|
|
2008-11-28 22:51:08 +00:00
|
|
|
// Loop over all of the things that depend on the instruction we're removing.
|
|
|
|
//
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDepMapType::iterator ReverseDepIt =
|
|
|
|
reverseDep.find(DepResultTy(RemInst, Normal));
|
2008-11-28 22:51:08 +00:00
|
|
|
if (ReverseDepIt != reverseDep.end()) {
|
|
|
|
SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
|
|
|
|
for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
|
|
|
|
E = ReverseDeps.end(); I != E; ++I) {
|
|
|
|
Instruction *InstDependingOnRemInst = *I;
|
|
|
|
|
|
|
|
// If we thought the instruction depended on itself (possible for
|
|
|
|
// unconfirmed dependencies) ignore the update.
|
|
|
|
if (InstDependingOnRemInst == RemInst) continue;
|
|
|
|
|
|
|
|
// Insert the new dependencies.
|
2008-11-29 01:43:36 +00:00
|
|
|
LocalDeps[InstDependingOnRemInst] =
|
2008-11-28 22:51:08 +00:00
|
|
|
std::make_pair(NewDependency, NewDependencyConfirmed);
|
|
|
|
|
|
|
|
// If our NewDependency is an instruction, make sure to remember that new
|
|
|
|
// things depend on it.
|
2008-11-29 01:43:36 +00:00
|
|
|
// FIXME: Just insert all deps!
|
|
|
|
if (NewDependency.getInt() != NonLocal && NewDependency.getInt() != None)
|
2008-11-28 22:51:08 +00:00
|
|
|
reverseDep[NewDependency].insert(InstDependingOnRemInst);
|
|
|
|
}
|
2008-11-29 01:43:36 +00:00
|
|
|
reverseDep.erase(DepResultTy(RemInst, Normal));
|
2007-07-06 23:14:35 +00:00
|
|
|
}
|
2007-08-16 21:27:05 +00:00
|
|
|
|
2008-11-29 01:43:36 +00:00
|
|
|
ReverseDepIt = reverseDepNonLocal.find(DepResultTy(RemInst, Normal));
|
2008-11-28 22:51:08 +00:00
|
|
|
if (ReverseDepIt != reverseDepNonLocal.end()) {
|
|
|
|
SmallPtrSet<Instruction*, 4>& set = ReverseDepIt->second;
|
2007-08-16 21:27:05 +00:00
|
|
|
for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
|
|
|
|
I != E; ++I)
|
2008-11-29 01:43:36 +00:00
|
|
|
for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
|
2007-09-21 03:53:52 +00:00
|
|
|
depGraphNonLocal[*I].begin(), DE = depGraphNonLocal[*I].end();
|
|
|
|
DI != DE; ++DI)
|
2008-11-29 01:43:36 +00:00
|
|
|
if (DI->second == DepResultTy(RemInst, Normal))
|
|
|
|
DI->second = DepResultTy(0, Dirty);
|
|
|
|
reverseDepNonLocal.erase(ReverseDepIt);
|
2007-08-16 21:27:05 +00:00
|
|
|
}
|
2007-12-08 01:37:09 +00:00
|
|
|
|
2008-11-28 22:04:47 +00:00
|
|
|
depGraphNonLocal.erase(RemInst);
|
2007-07-31 20:01:27 +00:00
|
|
|
|
2008-11-28 22:04:47 +00:00
|
|
|
getAnalysis<AliasAnalysis>().deleteValue(RemInst);
|
2008-11-28 21:45:17 +00:00
|
|
|
|
2008-11-28 22:04:47 +00:00
|
|
|
DEBUG(verifyRemoved(RemInst));
|
2007-07-06 23:14:35 +00:00
|
|
|
}
|