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.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#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"
|
2007-07-24 21:52:37 +00:00
|
|
|
#include "llvm/Support/CFG.h"
|
2007-07-06 23:14:35 +00:00
|
|
|
#include "llvm/Target/TargetData.h"
|
2007-09-09 21:43:49 +00:00
|
|
|
#include "llvm/ADT/Statistic.h"
|
|
|
|
|
|
|
|
#define DEBUG_TYPE "memdep"
|
2007-07-06 23:14:35 +00:00
|
|
|
|
|
|
|
using namespace llvm;
|
|
|
|
|
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;
|
|
|
|
|
2007-08-09 04:42:44 +00:00
|
|
|
Instruction* const MemoryDependenceAnalysis::NonLocal = (Instruction*)-3;
|
|
|
|
Instruction* const MemoryDependenceAnalysis::None = (Instruction*)-4;
|
2007-09-19 16:13:57 +00:00
|
|
|
Instruction* const MemoryDependenceAnalysis::Dirty = (Instruction*)-5;
|
2007-07-06 23:14:35 +00:00
|
|
|
|
|
|
|
// Register this pass...
|
2007-07-10 20:21:08 +00:00
|
|
|
static RegisterPass<MemoryDependenceAnalysis> X("memdep",
|
|
|
|
"Memory Dependence Analysis");
|
2007-07-06 23:14:35 +00:00
|
|
|
|
2007-12-08 01:37:09 +00:00
|
|
|
void MemoryDependenceAnalysis::ping(Instruction *D) {
|
|
|
|
for (depMapType::iterator I = depGraphLocal.begin(), E = depGraphLocal.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
assert(I->first != D);
|
|
|
|
assert(I->second.first != D);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (nonLocalDepMapType::iterator I = depGraphNonLocal.begin(), E = depGraphNonLocal.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
assert(I->first != D);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (reverseDepMapType::iterator I = reverseDep.begin(), E = reverseDep.end();
|
|
|
|
I != E; ++I)
|
|
|
|
for (SmallPtrSet<Instruction*, 4>::iterator II = I->second.begin(), EE = I->second.end();
|
|
|
|
II != EE; ++II)
|
|
|
|
assert(*II != D);
|
|
|
|
|
|
|
|
for (reverseDepMapType::iterator I = reverseDepNonLocal.begin(), E = reverseDepNonLocal.end();
|
|
|
|
I != E; ++I)
|
|
|
|
for (SmallPtrSet<Instruction*, 4>::iterator II = I->second.begin(), EE = I->second.end();
|
|
|
|
II != EE; ++II)
|
|
|
|
assert(*II != D);
|
|
|
|
}
|
|
|
|
|
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.
|
2007-08-09 04:42:44 +00:00
|
|
|
Instruction* MemoryDependenceAnalysis::getCallSiteDependency(CallSite C,
|
2007-08-08 22:01:54 +00:00
|
|
|
Instruction* start,
|
|
|
|
BasicBlock* block) {
|
2007-07-10 17:59:22 +00:00
|
|
|
|
2007-12-08 01:37:09 +00:00
|
|
|
std::pair<Instruction*, bool>& cachedResult =
|
|
|
|
depGraphLocal[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();
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
// If the starting point was specifiy, use it
|
2007-08-07 00:33:45 +00:00
|
|
|
if (start) {
|
|
|
|
QI = start;
|
|
|
|
blockBegin = start->getParent()->end();
|
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();
|
|
|
|
blockBegin = block->end();
|
|
|
|
}
|
|
|
|
|
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()))
|
2007-08-08 22:01:54 +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;
|
2007-12-01 07:51:45 +00:00
|
|
|
} else if (isa<CallInst>(QI)) {
|
2007-11-26 02:26:36 +00:00
|
|
|
AliasAnalysis::ModRefBehavior result =
|
2007-12-01 07:51:45 +00:00
|
|
|
AA.getModRefBehavior(CallSite::get(QI));
|
2007-11-26 02:26:36 +00:00
|
|
|
if (result != AliasAnalysis::DoesNotAccessMemory &&
|
|
|
|
result != AliasAnalysis::OnlyReadsMemory) {
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.first = QI;
|
|
|
|
cachedResult.second = true;
|
2007-08-07 00:33:45 +00:00
|
|
|
reverseDep[QI].insert(C.getInstruction());
|
|
|
|
}
|
2007-07-10 17:59:22 +00:00
|
|
|
return QI;
|
|
|
|
} 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) {
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.first = QI;
|
|
|
|
cachedResult.second = true;
|
2007-08-07 00:33:45 +00:00
|
|
|
reverseDep[QI].insert(C.getInstruction());
|
|
|
|
}
|
2007-07-10 17:59:22 +00:00
|
|
|
return QI;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// No dependence found
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.first = NonLocal;
|
|
|
|
cachedResult.second = true;
|
2007-08-07 00:33:45 +00:00
|
|
|
reverseDep[NonLocal].insert(C.getInstruction());
|
2007-07-10 17:59:22 +00:00
|
|
|
return NonLocal;
|
|
|
|
}
|
|
|
|
|
2007-08-08 22:26:03 +00:00
|
|
|
/// nonLocalHelper - Private helper used to calculate non-local dependencies
|
|
|
|
/// 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,
|
2007-08-08 22:01:54 +00:00
|
|
|
DenseMap<BasicBlock*, Value*>& 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.
|
|
|
|
for (DenseMap<BasicBlock*, Value*>::iterator I = resp.begin(),
|
|
|
|
E = resp.end(); I != E; ++I)
|
|
|
|
if (I->second != Dirty)
|
|
|
|
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;
|
|
|
|
stack.push_back(block);
|
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);
|
|
|
|
|
2007-08-09 04:42:44 +00:00
|
|
|
Instruction* localDep = getDependency(query, 0, BB);
|
2007-08-01 22:01:54 +00:00
|
|
|
if (localDep != NonLocal) {
|
2007-08-09 04:42:44 +00:00
|
|
|
resp.insert(std::make_pair(BB, 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.
|
2007-08-02 17:56:05 +00:00
|
|
|
} else if (BB == block && stack.size() > 1) {
|
|
|
|
visited.insert(BB);
|
|
|
|
|
2007-08-09 04:42:44 +00:00
|
|
|
Instruction* localDep = getDependency(query, 0, BB);
|
2007-08-02 17:56:05 +00:00
|
|
|
if (localDep != query)
|
2007-08-09 04:42:44 +00:00
|
|
|
resp.insert(std::make_pair(BB, 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
|
2007-08-01 22:01:54 +00:00
|
|
|
bool predOnStack = false;
|
|
|
|
bool inserted = false;
|
|
|
|
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;
|
|
|
|
|
2007-08-08 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) {
|
2007-08-09 04:42:44 +00:00
|
|
|
resp.insert(std::make_pair(BB, 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){
|
2007-08-09 04:42:44 +00:00
|
|
|
resp.insert(std::make_pair(BB, 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,
|
2007-08-08 22:01:54 +00:00
|
|
|
DenseMap<BasicBlock*, Value*>& resp) {
|
2007-08-16 21:27:05 +00:00
|
|
|
if (depGraphNonLocal.count(query)) {
|
2007-09-21 03:53:52 +00:00
|
|
|
DenseMap<BasicBlock*, Value*>& cached = depGraphNonLocal[query];
|
2007-09-09 21:43:49 +00:00
|
|
|
NumCacheNonlocal++;
|
2007-09-21 03:53:52 +00:00
|
|
|
|
|
|
|
SmallVector<BasicBlock*, 4> dirtied;
|
|
|
|
for (DenseMap<BasicBlock*, Value*>::iterator I = cached.begin(),
|
|
|
|
E = cached.end(); I != E; ++I)
|
|
|
|
if (I->second == Dirty)
|
|
|
|
dirtied.push_back(I->first);
|
|
|
|
|
|
|
|
for (SmallVector<BasicBlock*, 4>::iterator I = dirtied.begin(),
|
|
|
|
E = dirtied.end(); I != E; ++I) {
|
|
|
|
Instruction* localDep = getDependency(query, 0, *I);
|
|
|
|
if (localDep != NonLocal)
|
|
|
|
cached[*I] = localDep;
|
|
|
|
else {
|
|
|
|
cached.erase(*I);
|
|
|
|
nonLocalHelper(query, *I, cached);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
resp = cached;
|
|
|
|
|
2007-08-16 21:27:05 +00:00
|
|
|
return;
|
2007-09-09 21:43:49 +00:00
|
|
|
} else
|
|
|
|
NumUncacheNonlocal++;
|
2007-07-24 21:52:37 +00:00
|
|
|
|
2007-09-09 21:43:49 +00:00
|
|
|
// If not, go ahead and search for non-local deps.
|
2007-08-01 22:01:54 +00:00
|
|
|
nonLocalHelper(query, query->getParent(), resp);
|
2007-08-16 21:27:05 +00:00
|
|
|
|
|
|
|
// Update the non-local dependency cache
|
|
|
|
for (DenseMap<BasicBlock*, Value*>::iterator I = resp.begin(), E = resp.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
depGraphNonLocal[query].insert(*I);
|
|
|
|
reverseDepNonLocal[I->second].insert(query);
|
|
|
|
}
|
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
|
2007-07-10 17:08:11 +00:00
|
|
|
/// depends. The local paramter indicates if the query should only
|
|
|
|
/// evaluate dependencies within the same basic block.
|
2007-08-09 04:42:44 +00:00
|
|
|
Instruction* MemoryDependenceAnalysis::getDependency(Instruction* query,
|
2007-07-16 21:52:50 +00:00
|
|
|
Instruction* start,
|
2007-07-24 21:52:37 +00:00
|
|
|
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
|
2007-12-08 01:37:09 +00:00
|
|
|
std::pair<Instruction*, bool>& cachedResult = depGraphLocal[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)
|
|
|
|
return cachedResult.first;
|
|
|
|
else if (cachedResult.first && cachedResult.first != NonLocal)
|
|
|
|
// If we have an unconfirmed cached entry, we can start our search from there
|
|
|
|
QI = cachedResult.first;
|
|
|
|
}
|
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))
|
2007-07-06 23:14:35 +00:00
|
|
|
return None;
|
2007-07-10 17:25:03 +00:00
|
|
|
else
|
2007-07-10 17:08:11 +00:00
|
|
|
return None;
|
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) {
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.first = S;
|
|
|
|
cachedResult.second = true;
|
2007-08-07 00:33:45 +00:00
|
|
|
reverseDep[S].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
|
|
|
|
2007-07-10 18:11:42 +00:00
|
|
|
return S;
|
|
|
|
}
|
|
|
|
|
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) {
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.first = L;
|
|
|
|
cachedResult.second = true;
|
2007-08-07 00:33:45 +00:00
|
|
|
reverseDep[L].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
|
|
|
|
2007-07-10 18:11:42 +00:00
|
|
|
return L;
|
|
|
|
}
|
|
|
|
|
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()))
|
2007-08-08 22:01:54 +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),
|
|
|
|
dependee, dependeeSize);
|
|
|
|
|
|
|
|
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) {
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.first = QI;
|
|
|
|
cachedResult.second = true;
|
2007-08-07 00:33:45 +00:00
|
|
|
reverseDep[QI].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
|
|
|
|
2007-07-10 17:59:22 +00:00
|
|
|
return 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) {
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.first = QI;
|
|
|
|
cachedResult.second = true;
|
2007-08-07 00:33:45 +00:00
|
|
|
reverseDep[QI].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
|
|
|
|
2007-07-06 23:14:35 +00:00
|
|
|
return QI;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If we found nothing, return the non-local flag
|
2007-08-07 00:33:45 +00:00
|
|
|
if (!start && !block) {
|
2007-12-08 01:37:09 +00:00
|
|
|
cachedResult.first = NonLocal;
|
|
|
|
cachedResult.second = true;
|
2007-08-07 00:33:45 +00:00
|
|
|
reverseDep[NonLocal].insert(query);
|
2007-07-16 21:52:50 +00:00
|
|
|
}
|
2007-07-06 23:14:35 +00:00
|
|
|
|
|
|
|
return NonLocal;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// 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.
|
2007-07-06 23:14:35 +00:00
|
|
|
void MemoryDependenceAnalysis::removeInstruction(Instruction* rem) {
|
|
|
|
// Figure out the new dep for things that currently depend on rem
|
2007-08-09 04:42:44 +00:00
|
|
|
Instruction* newDep = NonLocal;
|
2007-07-06 23:14:35 +00:00
|
|
|
|
2007-12-08 01:37:09 +00:00
|
|
|
reverseDep[depGraphLocal[rem].first].erase(rem);
|
|
|
|
|
|
|
|
for (DenseMap<BasicBlock*, Value*>::iterator DI =
|
|
|
|
depGraphNonLocal[rem].begin(), DE = depGraphNonLocal[rem].end();
|
|
|
|
DI != DE; ++DI)
|
|
|
|
if (DI->second != None)
|
|
|
|
reverseDepNonLocal[DI->second].erase(rem);
|
|
|
|
|
2007-07-31 20:01:27 +00:00
|
|
|
depMapType::iterator depGraphEntry = depGraphLocal.find(rem);
|
|
|
|
|
|
|
|
if (depGraphEntry != depGraphLocal.end()) {
|
|
|
|
if (depGraphEntry->second.first != NonLocal &&
|
|
|
|
depGraphEntry->second.second) {
|
|
|
|
// If we have dep info for rem, set them to it
|
2007-08-09 04:42:44 +00:00
|
|
|
BasicBlock::iterator RI = depGraphEntry->second.first;
|
2007-07-31 20:01:27 +00:00
|
|
|
RI++;
|
|
|
|
newDep = RI;
|
|
|
|
} else if (depGraphEntry->second.first == NonLocal &&
|
|
|
|
depGraphEntry->second.second ) {
|
|
|
|
// If we have a confirmed non-local flag, use it
|
|
|
|
newDep = NonLocal;
|
|
|
|
} else {
|
|
|
|
// Otherwise, use the immediate successor of rem
|
2007-08-08 22:01:54 +00:00
|
|
|
// NOTE: This is because, when getDependence is called, it will first
|
|
|
|
// check the immediate predecessor of what is in the cache.
|
2007-07-31 20:01:27 +00:00
|
|
|
BasicBlock::iterator RI = rem;
|
|
|
|
RI++;
|
|
|
|
newDep = RI;
|
|
|
|
}
|
|
|
|
|
2007-08-07 00:33:45 +00:00
|
|
|
SmallPtrSet<Instruction*, 4>& set = reverseDep[rem];
|
|
|
|
for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
|
|
|
|
I != E; ++I) {
|
2007-07-31 20:01:27 +00:00
|
|
|
// Insert the new dependencies
|
|
|
|
// Mark it as unconfirmed as long as it is not the non-local flag
|
2007-08-07 00:33:45 +00:00
|
|
|
depGraphLocal[*I] = std::make_pair(newDep, !newDep);
|
2007-07-31 20:01:27 +00:00
|
|
|
}
|
2007-07-06 23:14:35 +00:00
|
|
|
}
|
2007-08-16 21:27:05 +00:00
|
|
|
|
2007-12-08 01:37:09 +00:00
|
|
|
depGraphLocal.erase(rem);
|
|
|
|
reverseDep.erase(rem);
|
|
|
|
|
2007-09-11 04:31:00 +00:00
|
|
|
if (reverseDepNonLocal.count(rem)) {
|
2007-08-16 21:27:05 +00:00
|
|
|
SmallPtrSet<Instruction*, 4>& set = reverseDepNonLocal[rem];
|
|
|
|
for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
|
|
|
|
I != E; ++I)
|
2007-09-21 03:53:52 +00:00
|
|
|
for (DenseMap<BasicBlock*, Value*>::iterator DI =
|
|
|
|
depGraphNonLocal[*I].begin(), DE = depGraphNonLocal[*I].end();
|
|
|
|
DI != DE; ++DI)
|
|
|
|
if (DI->second == rem)
|
|
|
|
DI->second = Dirty;
|
2007-08-16 21:27:05 +00:00
|
|
|
|
|
|
|
}
|
2007-12-08 01:37:09 +00:00
|
|
|
|
|
|
|
reverseDepNonLocal.erase(rem);
|
|
|
|
nonLocalDepMapType::iterator I = depGraphNonLocal.find(rem);
|
|
|
|
if (I != depGraphNonLocal.end())
|
|
|
|
depGraphNonLocal.erase(I);
|
2007-07-31 20:01:27 +00:00
|
|
|
|
2007-07-12 00:06:21 +00:00
|
|
|
getAnalysis<AliasAnalysis>().deleteValue(rem);
|
2007-07-06 23:14:35 +00:00
|
|
|
}
|