2007-08-02 18:11:11 +00:00
|
|
|
//===- DeadStoreElimination.cpp - Fast Dead Store Elimination -------------===//
|
2007-07-11 00:46:18 +00:00
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
2007-07-11 19:03:09 +00:00
|
|
|
// This file was developed by Owen Anderson and is distributed under
|
2007-07-11 00:46:18 +00:00
|
|
|
// the University of Illinois Open Source License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This file implements a trivial dead store elimination that only considers
|
|
|
|
// basic-block local redundant stores.
|
|
|
|
//
|
|
|
|
// FIXME: This should eventually be extended to be a post-dominator tree
|
|
|
|
// traversal. Doing so would be pretty trivial.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2007-08-01 06:36:51 +00:00
|
|
|
#define DEBUG_TYPE "dse"
|
2007-07-11 00:46:18 +00:00
|
|
|
#include "llvm/Transforms/Scalar.h"
|
2007-07-12 21:41:30 +00:00
|
|
|
#include "llvm/Constants.h"
|
2007-07-11 00:46:18 +00:00
|
|
|
#include "llvm/Function.h"
|
|
|
|
#include "llvm/Instructions.h"
|
|
|
|
#include "llvm/Pass.h"
|
|
|
|
#include "llvm/ADT/SetVector.h"
|
|
|
|
#include "llvm/ADT/SmallPtrSet.h"
|
|
|
|
#include "llvm/ADT/Statistic.h"
|
2007-07-11 23:19:17 +00:00
|
|
|
#include "llvm/Analysis/AliasAnalysis.h"
|
2007-07-11 00:46:18 +00:00
|
|
|
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
|
2007-07-11 23:19:17 +00:00
|
|
|
#include "llvm/Target/TargetData.h"
|
2007-07-11 00:46:18 +00:00
|
|
|
#include "llvm/Transforms/Utils/Local.h"
|
|
|
|
#include "llvm/Support/Compiler.h"
|
|
|
|
using namespace llvm;
|
|
|
|
|
|
|
|
STATISTIC(NumFastStores, "Number of stores deleted");
|
|
|
|
STATISTIC(NumFastOther , "Number of other instrs removed");
|
|
|
|
|
|
|
|
namespace {
|
2007-08-01 06:36:51 +00:00
|
|
|
struct VISIBILITY_HIDDEN DSE : public FunctionPass {
|
2007-07-11 00:46:18 +00:00
|
|
|
static char ID; // Pass identification, replacement for typeid
|
2007-08-01 06:36:51 +00:00
|
|
|
DSE() : FunctionPass((intptr_t)&ID) {}
|
2007-07-11 00:46:18 +00:00
|
|
|
|
|
|
|
virtual bool runOnFunction(Function &F) {
|
|
|
|
bool Changed = false;
|
|
|
|
for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
|
|
|
|
Changed |= runOnBasicBlock(*I);
|
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool runOnBasicBlock(BasicBlock &BB);
|
2007-08-02 18:11:11 +00:00
|
|
|
bool handleFreeWithNonTrivialDependency(FreeInst* F,
|
|
|
|
Instruction* dependency,
|
|
|
|
SetVector<Instruction*>& possiblyDead);
|
2007-07-12 21:41:30 +00:00
|
|
|
bool handleEndBlock(BasicBlock& BB, SetVector<Instruction*>& possiblyDead);
|
2007-08-08 17:50:09 +00:00
|
|
|
bool RemoveUndeadPointers(Value* pointer,
|
2007-07-12 21:41:30 +00:00
|
|
|
BasicBlock::iterator& BBI,
|
2007-08-08 17:50:09 +00:00
|
|
|
SmallPtrSet<AllocaInst*, 64>& deadPointers,
|
2007-07-12 21:41:30 +00:00
|
|
|
SetVector<Instruction*>& possiblyDead);
|
2007-07-11 00:46:18 +00:00
|
|
|
void DeleteDeadInstructionChains(Instruction *I,
|
|
|
|
SetVector<Instruction*> &DeadInsts);
|
2007-08-08 04:52:29 +00:00
|
|
|
|
2007-08-08 06:06:02 +00:00
|
|
|
/// Find the base pointer that a pointer came from
|
|
|
|
/// Because this is used to find pointers that originate
|
|
|
|
/// from allocas, it is safe to ignore GEP indices, since
|
|
|
|
/// either the store will be in the alloca, and thus dead,
|
|
|
|
/// or beyond the end of the alloca, and thus undefined.
|
2007-11-01 05:29:16 +00:00
|
|
|
void TranslatePointerBitCasts(Value*& v, bool zeroGepsOnly = false) {
|
2007-08-02 18:11:11 +00:00
|
|
|
assert(isa<PointerType>(v->getType()) &&
|
|
|
|
"Translating a non-pointer type?");
|
2007-08-08 04:52:29 +00:00
|
|
|
while (true) {
|
2007-07-16 23:34:39 +00:00
|
|
|
if (BitCastInst* C = dyn_cast<BitCastInst>(v))
|
|
|
|
v = C->getOperand(0);
|
|
|
|
else if (GetElementPtrInst* G = dyn_cast<GetElementPtrInst>(v))
|
2007-11-01 05:29:16 +00:00
|
|
|
if (!zeroGepsOnly || G->hasAllZeroIndices()) {
|
|
|
|
v = G->getOperand(0);
|
|
|
|
} else {
|
|
|
|
break;
|
|
|
|
}
|
2007-08-08 04:52:29 +00:00
|
|
|
else
|
|
|
|
break;
|
|
|
|
}
|
2007-07-13 18:26:26 +00:00
|
|
|
}
|
2007-07-11 00:46:18 +00:00
|
|
|
|
|
|
|
// getAnalysisUsage - We require post dominance frontiers (aka Control
|
|
|
|
// Dependence Graph)
|
|
|
|
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
|
|
|
|
AU.setPreservesCFG();
|
2007-07-11 23:19:17 +00:00
|
|
|
AU.addRequired<TargetData>();
|
|
|
|
AU.addRequired<AliasAnalysis>();
|
2007-07-11 00:46:18 +00:00
|
|
|
AU.addRequired<MemoryDependenceAnalysis>();
|
2007-07-11 23:19:17 +00:00
|
|
|
AU.addPreserved<AliasAnalysis>();
|
2007-07-11 00:46:18 +00:00
|
|
|
AU.addPreserved<MemoryDependenceAnalysis>();
|
|
|
|
}
|
|
|
|
};
|
2007-08-01 06:36:51 +00:00
|
|
|
char DSE::ID = 0;
|
|
|
|
RegisterPass<DSE> X("dse", "Dead Store Elimination");
|
2007-07-11 00:46:18 +00:00
|
|
|
}
|
|
|
|
|
2007-08-01 06:36:51 +00:00
|
|
|
FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
|
2007-07-11 00:46:18 +00:00
|
|
|
|
2007-08-01 06:36:51 +00:00
|
|
|
bool DSE::runOnBasicBlock(BasicBlock &BB) {
|
2007-07-11 00:46:18 +00:00
|
|
|
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
|
2007-11-01 05:29:16 +00:00
|
|
|
TargetData &TD = getAnalysis<TargetData>();
|
|
|
|
|
2007-07-11 19:03:09 +00:00
|
|
|
// Record the last-seen store to this pointer
|
2007-07-11 00:46:18 +00:00
|
|
|
DenseMap<Value*, StoreInst*> lastStore;
|
2007-07-11 19:03:09 +00:00
|
|
|
// Record instructions possibly made dead by deleting a store
|
2007-07-11 00:46:18 +00:00
|
|
|
SetVector<Instruction*> possiblyDead;
|
|
|
|
|
|
|
|
bool MadeChange = false;
|
|
|
|
|
|
|
|
// Do a top-down walk on the BB
|
2007-08-02 18:11:11 +00:00
|
|
|
for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end();
|
|
|
|
BBI != BBE; ++BBI) {
|
2007-07-11 21:06:56 +00:00
|
|
|
// If we find a store or a free...
|
2007-08-08 04:52:29 +00:00
|
|
|
if (!isa<StoreInst>(BBI) && !isa<FreeInst>(BBI))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
Value* pointer = 0;
|
2007-08-26 21:14:47 +00:00
|
|
|
if (StoreInst* S = dyn_cast<StoreInst>(BBI)) {
|
|
|
|
if (!S->isVolatile())
|
|
|
|
pointer = S->getPointerOperand();
|
|
|
|
else
|
|
|
|
continue;
|
|
|
|
} else
|
2007-08-08 06:06:02 +00:00
|
|
|
pointer = cast<FreeInst>(BBI)->getPointerOperand();
|
2007-07-11 20:38:34 +00:00
|
|
|
|
2007-11-01 05:29:16 +00:00
|
|
|
TranslatePointerBitCasts(pointer, true);
|
2007-08-08 04:52:29 +00:00
|
|
|
StoreInst*& last = lastStore[pointer];
|
|
|
|
bool deletedStore = false;
|
2007-07-11 00:46:18 +00:00
|
|
|
|
2007-08-08 04:52:29 +00:00
|
|
|
// ... to a pointer that has been stored to before...
|
|
|
|
if (last) {
|
2007-08-09 04:42:44 +00:00
|
|
|
Instruction* dep = MD.getDependency(BBI);
|
2007-07-11 00:46:18 +00:00
|
|
|
|
2007-08-08 04:52:29 +00:00
|
|
|
// ... and no other memory dependencies are between them....
|
|
|
|
while (dep != MemoryDependenceAnalysis::None &&
|
|
|
|
dep != MemoryDependenceAnalysis::NonLocal &&
|
|
|
|
isa<StoreInst>(dep)) {
|
2007-11-01 05:29:16 +00:00
|
|
|
if (dep != last ||
|
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.getTypeStoreSize(last->getOperand(0)->getType()) >
|
|
|
|
TD.getTypeStoreSize(BBI->getOperand(0)->getType())) {
|
2007-08-09 04:42:44 +00:00
|
|
|
dep = MD.getDependency(BBI, dep);
|
2007-08-08 04:52:29 +00:00
|
|
|
continue;
|
|
|
|
}
|
2007-07-16 21:52:50 +00:00
|
|
|
|
2007-08-08 04:52:29 +00:00
|
|
|
// Remove it!
|
|
|
|
MD.removeInstruction(last);
|
2007-07-11 00:46:18 +00:00
|
|
|
|
2007-08-08 04:52:29 +00:00
|
|
|
// DCE instructions only used to calculate that store
|
|
|
|
if (Instruction* D = dyn_cast<Instruction>(last->getOperand(0)))
|
|
|
|
possiblyDead.insert(D);
|
|
|
|
if (Instruction* D = dyn_cast<Instruction>(last->getOperand(1)))
|
|
|
|
possiblyDead.insert(D);
|
2007-07-11 00:46:18 +00:00
|
|
|
|
2007-08-08 04:52:29 +00:00
|
|
|
last->eraseFromParent();
|
|
|
|
NumFastStores++;
|
|
|
|
deletedStore = true;
|
|
|
|
MadeChange = true;
|
|
|
|
|
|
|
|
break;
|
2007-07-11 00:46:18 +00:00
|
|
|
}
|
2007-08-08 04:52:29 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Handle frees whose dependencies are non-trivial.
|
|
|
|
if (FreeInst* F = dyn_cast<FreeInst>(BBI)) {
|
|
|
|
if (!deletedStore)
|
|
|
|
MadeChange |= handleFreeWithNonTrivialDependency(F,
|
2007-08-09 04:42:44 +00:00
|
|
|
MD.getDependency(F),
|
2007-08-08 04:52:29 +00:00
|
|
|
possiblyDead);
|
|
|
|
// No known stores after the free
|
|
|
|
last = 0;
|
|
|
|
} else {
|
|
|
|
// Update our most-recent-store map.
|
|
|
|
last = cast<StoreInst>(BBI);
|
2007-07-11 00:46:18 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2007-07-12 21:41:30 +00:00
|
|
|
// If this block ends in a return, unwind, unreachable, and eventually
|
|
|
|
// tailcall, then all allocas are dead at its end.
|
|
|
|
if (BB.getTerminator()->getNumSuccessors() == 0)
|
|
|
|
MadeChange |= handleEndBlock(BB, possiblyDead);
|
|
|
|
|
2007-07-11 00:46:18 +00:00
|
|
|
// Do a trivial DCE
|
|
|
|
while (!possiblyDead.empty()) {
|
|
|
|
Instruction *I = possiblyDead.back();
|
|
|
|
possiblyDead.pop_back();
|
|
|
|
DeleteDeadInstructionChains(I, possiblyDead);
|
|
|
|
}
|
|
|
|
|
|
|
|
return MadeChange;
|
|
|
|
}
|
|
|
|
|
2007-07-11 23:19:17 +00:00
|
|
|
/// handleFreeWithNonTrivialDependency - Handle frees of entire structures whose
|
|
|
|
/// dependency is a store to a field of that structure
|
2007-08-01 06:36:51 +00:00
|
|
|
bool DSE::handleFreeWithNonTrivialDependency(FreeInst* F, Instruction* dep,
|
2007-08-02 18:11:11 +00:00
|
|
|
SetVector<Instruction*>& possiblyDead) {
|
2007-07-11 23:19:17 +00:00
|
|
|
TargetData &TD = getAnalysis<TargetData>();
|
|
|
|
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
|
|
|
|
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
|
|
|
|
|
2007-07-12 18:08:51 +00:00
|
|
|
if (dep == MemoryDependenceAnalysis::None ||
|
|
|
|
dep == MemoryDependenceAnalysis::NonLocal)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
StoreInst* dependency = dyn_cast<StoreInst>(dep);
|
|
|
|
if (!dependency)
|
|
|
|
return false;
|
2007-08-26 21:14:47 +00:00
|
|
|
else if (dependency->isVolatile())
|
|
|
|
return false;
|
2007-07-12 18:08:51 +00:00
|
|
|
|
2007-07-11 23:19:17 +00:00
|
|
|
Value* depPointer = dependency->getPointerOperand();
|
2007-08-02 18:11:11 +00:00
|
|
|
const Type* depType = dependency->getOperand(0)->getType();
|
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
|
|
|
unsigned depPointerSize = TD.getTypeStoreSize(depType);
|
2007-07-13 18:26:26 +00:00
|
|
|
|
2007-07-11 23:19:17 +00:00
|
|
|
// Check for aliasing
|
|
|
|
AliasAnalysis::AliasResult A = AA.alias(F->getPointerOperand(), ~0UL,
|
|
|
|
depPointer, depPointerSize);
|
|
|
|
|
|
|
|
if (A == AliasAnalysis::MustAlias) {
|
|
|
|
// Remove it!
|
|
|
|
MD.removeInstruction(dependency);
|
|
|
|
|
|
|
|
// DCE instructions only used to calculate that store
|
|
|
|
if (Instruction* D = dyn_cast<Instruction>(dependency->getOperand(0)))
|
|
|
|
possiblyDead.insert(D);
|
2007-07-13 18:26:26 +00:00
|
|
|
if (Instruction* D = dyn_cast<Instruction>(dependency->getOperand(1)))
|
|
|
|
possiblyDead.insert(D);
|
2007-07-11 23:19:17 +00:00
|
|
|
|
|
|
|
dependency->eraseFromParent();
|
|
|
|
NumFastStores++;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2007-08-02 18:11:11 +00:00
|
|
|
/// handleEndBlock - Remove dead stores to stack-allocated locations in the
|
2007-08-08 17:50:09 +00:00
|
|
|
/// function end block. Ex:
|
|
|
|
/// %A = alloca i32
|
|
|
|
/// ...
|
|
|
|
/// store i32 1, i32* %A
|
|
|
|
/// ret void
|
2007-08-02 18:11:11 +00:00
|
|
|
bool DSE::handleEndBlock(BasicBlock& BB,
|
|
|
|
SetVector<Instruction*>& possiblyDead) {
|
2007-07-12 21:41:30 +00:00
|
|
|
TargetData &TD = getAnalysis<TargetData>();
|
|
|
|
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
|
|
|
|
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
|
|
|
|
|
|
|
|
bool MadeChange = false;
|
|
|
|
|
|
|
|
// Pointers alloca'd in this function are dead in the end block
|
2007-08-08 17:50:09 +00:00
|
|
|
SmallPtrSet<AllocaInst*, 64> deadPointers;
|
2007-07-12 21:41:30 +00:00
|
|
|
|
|
|
|
// Find all of the alloca'd pointers in the entry block
|
|
|
|
BasicBlock *Entry = BB.getParent()->begin();
|
|
|
|
for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
|
|
|
|
if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
|
|
|
|
deadPointers.insert(AI);
|
|
|
|
|
|
|
|
// Scan the basic block backwards
|
|
|
|
for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
|
|
|
|
--BBI;
|
|
|
|
|
|
|
|
if (deadPointers.empty())
|
|
|
|
break;
|
|
|
|
|
|
|
|
// If we find a store whose pointer is dead...
|
|
|
|
if (StoreInst* S = dyn_cast<StoreInst>(BBI)) {
|
2007-08-26 21:14:47 +00:00
|
|
|
if (!S->isVolatile()) {
|
|
|
|
Value* pointerOperand = S->getPointerOperand();
|
|
|
|
// See through pointer-to-pointer bitcasts
|
|
|
|
TranslatePointerBitCasts(pointerOperand);
|
2007-07-13 18:26:26 +00:00
|
|
|
|
2007-08-26 21:14:47 +00:00
|
|
|
if (deadPointers.count(pointerOperand)){
|
|
|
|
// Remove it!
|
|
|
|
MD.removeInstruction(S);
|
2007-07-12 21:41:30 +00:00
|
|
|
|
2007-08-26 21:14:47 +00:00
|
|
|
// DCE instructions only used to calculate that store
|
|
|
|
if (Instruction* D = dyn_cast<Instruction>(S->getOperand(0)))
|
|
|
|
possiblyDead.insert(D);
|
|
|
|
if (Instruction* D = dyn_cast<Instruction>(S->getOperand(1)))
|
|
|
|
possiblyDead.insert(D);
|
2007-07-12 21:41:30 +00:00
|
|
|
|
2007-08-26 21:14:47 +00:00
|
|
|
BBI++;
|
|
|
|
S->eraseFromParent();
|
|
|
|
NumFastStores++;
|
|
|
|
MadeChange = true;
|
|
|
|
}
|
2007-07-12 21:41:30 +00:00
|
|
|
}
|
2007-08-08 17:50:09 +00:00
|
|
|
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
Value* killPointer = 0;
|
2007-07-12 21:41:30 +00:00
|
|
|
|
|
|
|
// If we encounter a use of the pointer, it is no longer considered dead
|
2007-08-08 17:50:09 +00:00
|
|
|
if (LoadInst* L = dyn_cast<LoadInst>(BBI)) {
|
2007-07-12 21:41:30 +00:00
|
|
|
killPointer = L->getPointerOperand();
|
|
|
|
} else if (VAArgInst* V = dyn_cast<VAArgInst>(BBI)) {
|
|
|
|
killPointer = V->getOperand(0);
|
|
|
|
} else if (AllocaInst* A = dyn_cast<AllocaInst>(BBI)) {
|
|
|
|
deadPointers.erase(A);
|
|
|
|
continue;
|
|
|
|
} else if (CallSite::get(BBI).getInstruction() != 0) {
|
2007-08-08 17:58:56 +00:00
|
|
|
// If this call does not access memory, it can't
|
|
|
|
// be undeadifying any of our pointers.
|
|
|
|
CallSite CS = CallSite::get(BBI);
|
|
|
|
if (CS.getCalledFunction() &&
|
|
|
|
AA.doesNotAccessMemory(CS.getCalledFunction()))
|
|
|
|
continue;
|
|
|
|
|
2007-08-08 18:38:28 +00:00
|
|
|
unsigned modRef = 0;
|
|
|
|
unsigned other = 0;
|
|
|
|
|
2007-07-12 21:41:30 +00:00
|
|
|
// Remove any pointers made undead by the call from the dead set
|
|
|
|
std::vector<Instruction*> dead;
|
2007-08-08 17:50:09 +00:00
|
|
|
for (SmallPtrSet<AllocaInst*, 64>::iterator I = deadPointers.begin(),
|
2007-07-12 21:41:30 +00:00
|
|
|
E = deadPointers.end(); I != E; ++I) {
|
2007-08-08 18:38:28 +00:00
|
|
|
// HACK: if we detect that our AA is imprecise, it's not
|
|
|
|
// worth it to scan the rest of the deadPointers set. Just
|
|
|
|
// assume that the AA will return ModRef for everything, and
|
|
|
|
// go ahead and bail.
|
|
|
|
if (modRef >= 16 && other == 0) {
|
|
|
|
deadPointers.clear();
|
|
|
|
return MadeChange;
|
|
|
|
}
|
|
|
|
|
2007-07-12 21:41:30 +00:00
|
|
|
// Get size information for the alloca
|
|
|
|
unsigned pointerSize = ~0UL;
|
|
|
|
if (ConstantInt* C = dyn_cast<ConstantInt>((*I)->getArraySize()))
|
2007-08-02 18:11:11 +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((*I)->getAllocatedType());
|
2007-07-12 21:41:30 +00:00
|
|
|
|
|
|
|
// See if the call site touches it
|
2007-08-08 17:58:56 +00:00
|
|
|
AliasAnalysis::ModRefResult A = AA.getModRefInfo(CS, *I, pointerSize);
|
2007-08-08 18:38:28 +00:00
|
|
|
|
|
|
|
if (A == AliasAnalysis::ModRef)
|
|
|
|
modRef++;
|
|
|
|
else
|
|
|
|
other++;
|
|
|
|
|
2007-07-13 18:26:26 +00:00
|
|
|
if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
|
2007-07-12 21:41:30 +00:00
|
|
|
dead.push_back(*I);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (std::vector<Instruction*>::iterator I = dead.begin(), E = dead.end();
|
|
|
|
I != E; ++I)
|
|
|
|
deadPointers.erase(*I);
|
|
|
|
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!killPointer)
|
|
|
|
continue;
|
|
|
|
|
2007-08-08 18:38:28 +00:00
|
|
|
TranslatePointerBitCasts(killPointer);
|
|
|
|
|
2007-07-12 21:41:30 +00:00
|
|
|
// Deal with undead pointers
|
2007-08-08 17:50:09 +00:00
|
|
|
MadeChange |= RemoveUndeadPointers(killPointer, BBI,
|
2007-07-12 21:41:30 +00:00
|
|
|
deadPointers, possiblyDead);
|
|
|
|
}
|
|
|
|
|
|
|
|
return MadeChange;
|
|
|
|
}
|
|
|
|
|
2007-08-08 18:38:28 +00:00
|
|
|
/// RemoveUndeadPointers - check for uses of a pointer that make it
|
|
|
|
/// undead when scanning for dead stores to alloca's.
|
2007-08-08 17:50:09 +00:00
|
|
|
bool DSE::RemoveUndeadPointers(Value* killPointer,
|
2007-07-12 21:41:30 +00:00
|
|
|
BasicBlock::iterator& BBI,
|
2007-08-08 17:50:09 +00:00
|
|
|
SmallPtrSet<AllocaInst*, 64>& deadPointers,
|
2007-07-12 21:41:30 +00:00
|
|
|
SetVector<Instruction*>& possiblyDead) {
|
|
|
|
TargetData &TD = getAnalysis<TargetData>();
|
|
|
|
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
|
|
|
|
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
|
|
|
|
|
2007-08-08 18:38:28 +00:00
|
|
|
// If the kill pointer can be easily reduced to an alloca,
|
|
|
|
// don't bother doing extraneous AA queries
|
|
|
|
if (AllocaInst* A = dyn_cast<AllocaInst>(killPointer)) {
|
|
|
|
if (deadPointers.count(A))
|
|
|
|
deadPointers.erase(A);
|
|
|
|
return false;
|
2007-08-08 19:12:31 +00:00
|
|
|
} else if (isa<GlobalValue>(killPointer)) {
|
|
|
|
// A global can't be in the dead pointer set
|
|
|
|
return false;
|
2007-08-08 18:38:28 +00:00
|
|
|
}
|
|
|
|
|
2007-07-12 21:41:30 +00:00
|
|
|
bool MadeChange = false;
|
|
|
|
|
|
|
|
std::vector<Instruction*> undead;
|
|
|
|
|
2007-08-08 17:50:09 +00:00
|
|
|
for (SmallPtrSet<AllocaInst*, 64>::iterator I = deadPointers.begin(),
|
2007-07-12 21:41:30 +00:00
|
|
|
E = deadPointers.end(); I != E; ++I) {
|
|
|
|
// Get size information for the alloca
|
|
|
|
unsigned pointerSize = ~0UL;
|
|
|
|
if (ConstantInt* C = dyn_cast<ConstantInt>((*I)->getArraySize()))
|
2007-08-02 18:11:11 +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((*I)->getAllocatedType());
|
2007-07-12 21:41:30 +00:00
|
|
|
|
|
|
|
// See if this pointer could alias it
|
2007-08-02 18:11:11 +00:00
|
|
|
AliasAnalysis::AliasResult A = AA.alias(*I, pointerSize,
|
2007-08-08 17:50:09 +00:00
|
|
|
killPointer, ~0UL);
|
2007-07-12 21:41:30 +00:00
|
|
|
|
|
|
|
// If it must-alias and a store, we can delete it
|
|
|
|
if (isa<StoreInst>(BBI) && A == AliasAnalysis::MustAlias) {
|
|
|
|
StoreInst* S = cast<StoreInst>(BBI);
|
|
|
|
|
|
|
|
// Remove it!
|
|
|
|
MD.removeInstruction(S);
|
|
|
|
|
|
|
|
// DCE instructions only used to calculate that store
|
|
|
|
if (Instruction* D = dyn_cast<Instruction>(S->getOperand(0)))
|
|
|
|
possiblyDead.insert(D);
|
2007-07-13 18:26:26 +00:00
|
|
|
if (Instruction* D = dyn_cast<Instruction>(S->getOperand(1)))
|
|
|
|
possiblyDead.insert(D);
|
2007-07-12 21:41:30 +00:00
|
|
|
|
|
|
|
BBI++;
|
|
|
|
S->eraseFromParent();
|
|
|
|
NumFastStores++;
|
|
|
|
MadeChange = true;
|
|
|
|
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Otherwise, it is undead
|
|
|
|
} else if (A != AliasAnalysis::NoAlias)
|
|
|
|
undead.push_back(*I);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (std::vector<Instruction*>::iterator I = undead.begin(), E = undead.end();
|
|
|
|
I != E; ++I)
|
|
|
|
deadPointers.erase(*I);
|
|
|
|
|
|
|
|
return MadeChange;
|
|
|
|
}
|
|
|
|
|
2007-08-08 17:50:09 +00:00
|
|
|
/// DeleteDeadInstructionChains - takes an instruction and a setvector of
|
|
|
|
/// dead instructions. If I is dead, it is erased, and its operands are
|
|
|
|
/// checked for deadness. If they are dead, they are added to the dead
|
|
|
|
/// setvector.
|
2007-08-01 06:36:51 +00:00
|
|
|
void DSE::DeleteDeadInstructionChains(Instruction *I,
|
2007-07-11 00:46:18 +00:00
|
|
|
SetVector<Instruction*> &DeadInsts) {
|
|
|
|
// Instruction must be dead.
|
|
|
|
if (!I->use_empty() || !isInstructionTriviallyDead(I)) return;
|
|
|
|
|
|
|
|
// Let the memory dependence know
|
|
|
|
getAnalysis<MemoryDependenceAnalysis>().removeInstruction(I);
|
|
|
|
|
|
|
|
// See if this made any operands dead. We do it this way in case the
|
|
|
|
// instruction uses the same operand twice. We don't want to delete a
|
|
|
|
// value then reference it.
|
|
|
|
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
|
2007-07-11 19:03:09 +00:00
|
|
|
if (I->getOperand(i)->hasOneUse())
|
|
|
|
if (Instruction* Op = dyn_cast<Instruction>(I->getOperand(i)))
|
|
|
|
DeadInsts.insert(Op); // Attempt to nuke it later.
|
|
|
|
|
2007-07-11 00:46:18 +00:00
|
|
|
I->setOperand(i, 0); // Drop from the operand list.
|
|
|
|
}
|
|
|
|
|
|
|
|
I->eraseFromParent();
|
|
|
|
++NumFastOther;
|
|
|
|
}
|