llvm-6502/lib/Analysis/LoadValueNumbering.cpp
Duncan Sands 514ab348fd 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

530 lines
21 KiB
C++

//===- LoadValueNumbering.cpp - Load Value #'ing Implementation -*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements a value numbering pass that value numbers load and call
// instructions. To do this, it finds lexically identical load instructions,
// and uses alias analysis to determine which loads are guaranteed to produce
// the same value. To value number call instructions, it looks for calls to
// functions that do not write to memory which do not have intervening
// instructions that clobber the memory that is read from.
//
// This pass builds off of another value numbering pass to implement value
// numbering for non-load and non-call instructions. It uses Alias Analysis so
// that it can disambiguate the load instructions. The more powerful these base
// analyses are, the more powerful the resultant value numbering will be.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/LoadValueNumbering.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
#include "llvm/Pass.h"
#include "llvm/Type.h"
#include "llvm/Analysis/ValueNumbering.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Target/TargetData.h"
#include <set>
#include <algorithm>
using namespace llvm;
namespace {
// FIXME: This should not be a FunctionPass.
struct VISIBILITY_HIDDEN LoadVN : public FunctionPass, public ValueNumbering {
static char ID; // Class identification, replacement for typeinfo
LoadVN() : FunctionPass((intptr_t)&ID) {}
/// Pass Implementation stuff. This doesn't do any analysis.
///
bool runOnFunction(Function &) { return false; }
/// getAnalysisUsage - Does not modify anything. It uses Value Numbering
/// and Alias Analysis.
///
virtual void getAnalysisUsage(AnalysisUsage &AU) const;
/// getEqualNumberNodes - Return nodes with the same value number as the
/// specified Value. This fills in the argument vector with any equal
/// values.
///
virtual void getEqualNumberNodes(Value *V1,
std::vector<Value*> &RetVals) const;
/// deleteValue - This method should be called whenever an LLVM Value is
/// deleted from the program, for example when an instruction is found to be
/// redundant and is eliminated.
///
virtual void deleteValue(Value *V) {
getAnalysis<AliasAnalysis>().deleteValue(V);
}
/// copyValue - This method should be used whenever a preexisting value in
/// the program is copied or cloned, introducing a new value. Note that
/// analysis implementations should tolerate clients that use this method to
/// introduce the same value multiple times: if the analysis already knows
/// about a value, it should ignore the request.
///
virtual void copyValue(Value *From, Value *To) {
getAnalysis<AliasAnalysis>().copyValue(From, To);
}
/// getCallEqualNumberNodes - Given a call instruction, find other calls
/// that have the same value number.
void getCallEqualNumberNodes(CallInst *CI,
std::vector<Value*> &RetVals) const;
};
char LoadVN::ID = 0;
// Register this pass...
RegisterPass<LoadVN> X("load-vn", "Load Value Numbering");
// Declare that we implement the ValueNumbering interface
RegisterAnalysisGroup<ValueNumbering> Y(X);
}
FunctionPass *llvm::createLoadValueNumberingPass() { return new LoadVN(); }
/// getAnalysisUsage - Does not modify anything. It uses Value Numbering and
/// Alias Analysis.
///
void LoadVN::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequiredTransitive<AliasAnalysis>();
AU.addRequired<ValueNumbering>();
AU.addRequiredTransitive<DominatorTree>();
AU.addRequiredTransitive<TargetData>();
}
static bool isPathTransparentTo(BasicBlock *CurBlock, BasicBlock *Dom,
Value *Ptr, unsigned Size, AliasAnalysis &AA,
std::set<BasicBlock*> &Visited,
std::map<BasicBlock*, bool> &TransparentBlocks){
// If we have already checked out this path, or if we reached our destination,
// stop searching, returning success.
if (CurBlock == Dom || !Visited.insert(CurBlock).second)
return true;
// Check whether this block is known transparent or not.
std::map<BasicBlock*, bool>::iterator TBI =
TransparentBlocks.lower_bound(CurBlock);
if (TBI == TransparentBlocks.end() || TBI->first != CurBlock) {
// If this basic block can modify the memory location, then the path is not
// transparent!
if (AA.canBasicBlockModify(*CurBlock, Ptr, Size)) {
TransparentBlocks.insert(TBI, std::make_pair(CurBlock, false));
return false;
}
TransparentBlocks.insert(TBI, std::make_pair(CurBlock, true));
} else if (!TBI->second)
// This block is known non-transparent, so that path can't be either.
return false;
// The current block is known to be transparent. The entire path is
// transparent if all of the predecessors paths to the parent is also
// transparent to the memory location.
for (pred_iterator PI = pred_begin(CurBlock), E = pred_end(CurBlock);
PI != E; ++PI)
if (!isPathTransparentTo(*PI, Dom, Ptr, Size, AA, Visited,
TransparentBlocks))
return false;
return true;
}
/// getCallEqualNumberNodes - Given a call instruction, find other calls that
/// have the same value number.
void LoadVN::getCallEqualNumberNodes(CallInst *CI,
std::vector<Value*> &RetVals) const {
Function *CF = CI->getCalledFunction();
if (CF == 0) return; // Indirect call.
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
AliasAnalysis::ModRefBehavior MRB = AA.getModRefBehavior(CF, CI);
if (MRB != AliasAnalysis::DoesNotAccessMemory &&
MRB != AliasAnalysis::OnlyReadsMemory)
return; // Nothing we can do for now.
// Scan all of the arguments of the function, looking for one that is not
// global. In particular, we would prefer to have an argument or instruction
// operand to chase the def-use chains of.
Value *Op = CF;
for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
if (isa<Argument>(CI->getOperand(i)) ||
isa<Instruction>(CI->getOperand(i))) {
Op = CI->getOperand(i);
break;
}
// Identify all lexically identical calls in this function.
std::vector<CallInst*> IdenticalCalls;
Function *CIFunc = CI->getParent()->getParent();
for (Value::use_iterator UI = Op->use_begin(), E = Op->use_end(); UI != E;
++UI)
if (CallInst *C = dyn_cast<CallInst>(*UI))
if (C->getNumOperands() == CI->getNumOperands() &&
C->getOperand(0) == CI->getOperand(0) &&
C->getParent()->getParent() == CIFunc && C != CI) {
bool AllOperandsEqual = true;
for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
if (C->getOperand(i) != CI->getOperand(i)) {
AllOperandsEqual = false;
break;
}
if (AllOperandsEqual)
IdenticalCalls.push_back(C);
}
if (IdenticalCalls.empty()) return;
// Eliminate duplicates, which could occur if we chose a value that is passed
// into a call site multiple times.
std::sort(IdenticalCalls.begin(), IdenticalCalls.end());
IdenticalCalls.erase(std::unique(IdenticalCalls.begin(),IdenticalCalls.end()),
IdenticalCalls.end());
// If the call reads memory, we must make sure that there are no stores
// between the calls in question.
//
// FIXME: This should use mod/ref information. What we really care about it
// whether an intervening instruction could modify memory that is read, not
// ANY memory.
//
if (MRB == AliasAnalysis::OnlyReadsMemory) {
DominatorTree &DT = getAnalysis<DominatorTree>();
BasicBlock *CIBB = CI->getParent();
for (unsigned i = 0; i != IdenticalCalls.size(); ++i) {
CallInst *C = IdenticalCalls[i];
bool CantEqual = false;
if (DT.dominates(CIBB, C->getParent())) {
// FIXME: we currently only handle the case where both calls are in the
// same basic block.
if (CIBB != C->getParent()) {
CantEqual = true;
} else {
Instruction *First = CI, *Second = C;
if (!DT.dominates(CI, C))
std::swap(First, Second);
// Scan the instructions between the calls, checking for stores or
// calls to dangerous functions.
BasicBlock::iterator I = First;
for (++First; I != BasicBlock::iterator(Second); ++I) {
if (isa<StoreInst>(I)) {
// FIXME: We could use mod/ref information to make this much
// better!
CantEqual = true;
break;
} else if (CallInst *CI = dyn_cast<CallInst>(I)) {
if (CI->getCalledFunction() == 0 ||
!AA.onlyReadsMemory(CI->getCalledFunction())) {
CantEqual = true;
break;
}
} else if (I->mayWriteToMemory()) {
CantEqual = true;
break;
}
}
}
} else if (DT.dominates(C->getParent(), CIBB)) {
// FIXME: We could implement this, but we don't for now.
CantEqual = true;
} else {
// FIXME: if one doesn't dominate the other, we can't tell yet.
CantEqual = true;
}
if (CantEqual) {
// This call does not produce the same value as the one in the query.
std::swap(IdenticalCalls[i--], IdenticalCalls.back());
IdenticalCalls.pop_back();
}
}
}
// Any calls that are identical and not destroyed will produce equal values!
for (unsigned i = 0, e = IdenticalCalls.size(); i != e; ++i)
RetVals.push_back(IdenticalCalls[i]);
}
// getEqualNumberNodes - Return nodes with the same value number as the
// specified Value. This fills in the argument vector with any equal values.
//
void LoadVN::getEqualNumberNodes(Value *V,
std::vector<Value*> &RetVals) const {
// If the alias analysis has any must alias information to share with us, we
// can definitely use it.
if (isa<PointerType>(V->getType()))
getAnalysis<AliasAnalysis>().getMustAliases(V, RetVals);
if (!isa<LoadInst>(V)) {
if (CallInst *CI = dyn_cast<CallInst>(V))
getCallEqualNumberNodes(CI, RetVals);
// Not a load instruction? Just chain to the base value numbering
// implementation to satisfy the request...
assert(&getAnalysis<ValueNumbering>() != (ValueNumbering*)this &&
"getAnalysis() returned this!");
return getAnalysis<ValueNumbering>().getEqualNumberNodes(V, RetVals);
}
// Volatile loads cannot be replaced with the value of other loads.
LoadInst *LI = cast<LoadInst>(V);
if (LI->isVolatile())
return getAnalysis<ValueNumbering>().getEqualNumberNodes(V, RetVals);
Value *LoadPtr = LI->getOperand(0);
BasicBlock *LoadBB = LI->getParent();
Function *F = LoadBB->getParent();
// Find out how many bytes of memory are loaded by the load instruction...
unsigned LoadSize = getAnalysis<TargetData>().getTypeStoreSize(LI->getType());
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
// Figure out if the load is invalidated from the entry of the block it is in
// until the actual instruction. This scans the block backwards from LI. If
// we see any candidate load or store instructions, then we know that the
// candidates have the same value # as LI.
bool LoadInvalidatedInBBBefore = false;
for (BasicBlock::iterator I = LI; I != LoadBB->begin(); ) {
--I;
if (I == LoadPtr) {
// If we run into an allocation of the value being loaded, then the
// contents are not initialized.
if (isa<AllocationInst>(I))
RetVals.push_back(UndefValue::get(LI->getType()));
// Otherwise, since this is the definition of what we are loading, this
// loaded value cannot occur before this block.
LoadInvalidatedInBBBefore = true;
break;
} else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
// If this instruction is a candidate load before LI, we know there are no
// invalidating instructions between it and LI, so they have the same
// value number.
if (LI->getOperand(0) == LoadPtr && !LI->isVolatile())
RetVals.push_back(I);
}
if (AA.getModRefInfo(I, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
// If the invalidating instruction is a store, and its in our candidate
// set, then we can do store-load forwarding: the load has the same value
// # as the stored value.
if (StoreInst *SI = dyn_cast<StoreInst>(I))
if (SI->getOperand(1) == LoadPtr)
RetVals.push_back(I->getOperand(0));
LoadInvalidatedInBBBefore = true;
break;
}
}
// Figure out if the load is invalidated between the load and the exit of the
// block it is defined in. While we are scanning the current basic block, if
// we see any candidate loads, then we know they have the same value # as LI.
//
bool LoadInvalidatedInBBAfter = false;
{
BasicBlock::iterator I = LI;
for (++I; I != LoadBB->end(); ++I) {
// If this instruction is a load, then this instruction returns the same
// value as LI.
if (isa<LoadInst>(I) && cast<LoadInst>(I)->getOperand(0) == LoadPtr)
RetVals.push_back(I);
if (AA.getModRefInfo(I, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
LoadInvalidatedInBBAfter = true;
break;
}
}
}
// If the pointer is clobbered on entry and on exit to the function, there is
// no need to do any global analysis at all.
if (LoadInvalidatedInBBBefore && LoadInvalidatedInBBAfter)
return;
// Now that we know the value is not neccesarily killed on entry or exit to
// the BB, find out how many load and store instructions (to this location)
// live in each BB in the function.
//
std::map<BasicBlock*, unsigned> CandidateLoads;
std::set<BasicBlock*> CandidateStores;
for (Value::use_iterator UI = LoadPtr->use_begin(), UE = LoadPtr->use_end();
UI != UE; ++UI)
if (LoadInst *Cand = dyn_cast<LoadInst>(*UI)) {// Is a load of source?
if (Cand->getParent()->getParent() == F && // In the same function?
// Not in LI's block?
Cand->getParent() != LoadBB && !Cand->isVolatile())
++CandidateLoads[Cand->getParent()]; // Got one.
} else if (StoreInst *Cand = dyn_cast<StoreInst>(*UI)) {
if (Cand->getParent()->getParent() == F && !Cand->isVolatile() &&
Cand->getOperand(1) == LoadPtr) // It's a store THROUGH the ptr.
CandidateStores.insert(Cand->getParent());
}
// Get dominators.
DominatorTree &DT = getAnalysis<DominatorTree>();
// TransparentBlocks - For each basic block the load/store is alive across,
// figure out if the pointer is invalidated or not. If it is invalidated, the
// boolean is set to false, if it's not it is set to true. If we don't know
// yet, the entry is not in the map.
std::map<BasicBlock*, bool> TransparentBlocks;
// Loop over all of the basic blocks that also load the value. If the value
// is live across the CFG from the source to destination blocks, and if the
// value is not invalidated in either the source or destination blocks, add it
// to the equivalence sets.
for (std::map<BasicBlock*, unsigned>::iterator
I = CandidateLoads.begin(), E = CandidateLoads.end(); I != E; ++I) {
bool CantEqual = false;
// Right now we only can handle cases where one load dominates the other.
// FIXME: generalize this!
BasicBlock *BB1 = I->first, *BB2 = LoadBB;
if (DT.dominates(BB1, BB2)) {
// The other load dominates LI. If the loaded value is killed entering
// the LoadBB block, we know the load is not live.
if (LoadInvalidatedInBBBefore)
CantEqual = true;
} else if (DT.dominates(BB2, BB1)) {
std::swap(BB1, BB2); // Canonicalize
// LI dominates the other load. If the loaded value is killed exiting
// the LoadBB block, we know the load is not live.
if (LoadInvalidatedInBBAfter)
CantEqual = true;
} else {
// None of these loads can VN the same.
CantEqual = true;
}
if (!CantEqual) {
// Ok, at this point, we know that BB1 dominates BB2, and that there is
// nothing in the LI block that kills the loaded value. Check to see if
// the value is live across the CFG.
std::set<BasicBlock*> Visited;
for (pred_iterator PI = pred_begin(BB2), E = pred_end(BB2); PI!=E; ++PI)
if (!isPathTransparentTo(*PI, BB1, LoadPtr, LoadSize, AA,
Visited, TransparentBlocks)) {
// None of these loads can VN the same.
CantEqual = true;
break;
}
}
// If the loads can equal so far, scan the basic block that contains the
// loads under consideration to see if they are invalidated in the block.
// For any loads that are not invalidated, add them to the equivalence
// set!
if (!CantEqual) {
unsigned NumLoads = I->second;
if (BB1 == LoadBB) {
// If LI dominates the block in question, check to see if any of the
// loads in this block are invalidated before they are reached.
for (BasicBlock::iterator BBI = I->first->begin(); ; ++BBI) {
if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
if (LI->getOperand(0) == LoadPtr && !LI->isVolatile()) {
// The load is in the set!
RetVals.push_back(BBI);
if (--NumLoads == 0) break; // Found last load to check.
}
} else if (AA.getModRefInfo(BBI, LoadPtr, LoadSize)
& AliasAnalysis::Mod) {
// If there is a modifying instruction, nothing below it will value
// # the same.
break;
}
}
} else {
// If the block dominates LI, make sure that the loads in the block are
// not invalidated before the block ends.
BasicBlock::iterator BBI = I->first->end();
while (1) {
--BBI;
if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
if (LI->getOperand(0) == LoadPtr && !LI->isVolatile()) {
// The load is the same as this load!
RetVals.push_back(BBI);
if (--NumLoads == 0) break; // Found all of the laods.
}
} else if (AA.getModRefInfo(BBI, LoadPtr, LoadSize)
& AliasAnalysis::Mod) {
// If there is a modifying instruction, nothing above it will value
// # the same.
break;
}
}
}
}
}
// Handle candidate stores. If the loaded location is clobbered on entrance
// to the LoadBB, no store outside of the LoadBB can value number equal, so
// quick exit.
if (LoadInvalidatedInBBBefore)
return;
// Stores in the load-bb are handled above.
CandidateStores.erase(LoadBB);
for (std::set<BasicBlock*>::iterator I = CandidateStores.begin(),
E = CandidateStores.end(); I != E; ++I)
if (DT.dominates(*I, LoadBB)) {
BasicBlock *StoreBB = *I;
// Check to see if the path from the store to the load is transparent
// w.r.t. the memory location.
bool CantEqual = false;
std::set<BasicBlock*> Visited;
for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
PI != E; ++PI)
if (!isPathTransparentTo(*PI, StoreBB, LoadPtr, LoadSize, AA,
Visited, TransparentBlocks)) {
// None of these stores can VN the same.
CantEqual = true;
break;
}
Visited.clear();
if (!CantEqual) {
// Okay, the path from the store block to the load block is clear, and
// we know that there are no invalidating instructions from the start
// of the load block to the load itself. Now we just scan the store
// block.
BasicBlock::iterator BBI = StoreBB->end();
while (1) {
assert(BBI != StoreBB->begin() &&
"There is a store in this block of the pointer, but the store"
" doesn't mod the address being stored to?? Must be a bug in"
" the alias analysis implementation!");
--BBI;
if (AA.getModRefInfo(BBI, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
// If the invalidating instruction is one of the candidates,
// then it provides the value the load loads.
if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
if (SI->getOperand(1) == LoadPtr)
RetVals.push_back(SI->getOperand(0));
break;
}
}
}
}
}