//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "gvn"
-
#include "llvm/Transforms/Scalar.h"
#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/Compiler.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Target/TargetData.h"
using namespace llvm;
+STATISTIC(NumGVNInstr, "Number of instructions deleted");
+STATISTIC(NumGVNLoad, "Number of loads deleted");
+STATISTIC(NumMemSetInfer, "Number of memsets inferred");
+
+
//===----------------------------------------------------------------------===//
// ValueTable Class
//===----------------------------------------------------------------------===//
hash = e.secondVN + hash * 37;
hash = e.thirdVN + hash * 37;
- hash = (unsigned)((uintptr_t)e.type >> 4) ^
- (unsigned)((uintptr_t)e.type >> 9) +
- hash * 37;
+ hash = ((unsigned)((uintptr_t)e.type >> 4) ^
+ (unsigned)((uintptr_t)e.type >> 9)) +
+ hash * 37;
for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
E = e.varargs.end(); I != E; ++I)
hash = *I + hash * 37;
- hash = (unsigned)((uintptr_t)e.function >> 4) ^
- (unsigned)((uintptr_t)e.function >> 9) +
- hash * 37;
+ hash = ((unsigned)((uintptr_t)e.function >> 4) ^
+ (unsigned)((uintptr_t)e.function >> 9)) +
+ hash * 37;
return hash;
}
//===----------------------------------------------------------------------===//
// ValueTable Internal Functions
//===----------------------------------------------------------------------===//
-Expression::ExpressionOpcode
- ValueTable::getOpcode(BinaryOperator* BO) {
+Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
switch(BO->getOpcode()) {
- case Instruction::Add:
- return Expression::ADD;
- case Instruction::Sub:
- return Expression::SUB;
- case Instruction::Mul:
- return Expression::MUL;
- case Instruction::UDiv:
- return Expression::UDIV;
- case Instruction::SDiv:
- return Expression::SDIV;
- case Instruction::FDiv:
- return Expression::FDIV;
- case Instruction::URem:
- return Expression::UREM;
- case Instruction::SRem:
- return Expression::SREM;
- case Instruction::FRem:
- return Expression::FREM;
- case Instruction::Shl:
- return Expression::SHL;
- case Instruction::LShr:
- return Expression::LSHR;
- case Instruction::AShr:
- return Expression::ASHR;
- case Instruction::And:
- return Expression::AND;
- case Instruction::Or:
- return Expression::OR;
- case Instruction::Xor:
- return Expression::XOR;
-
- // THIS SHOULD NEVER HAPPEN
- default:
- assert(0 && "Binary operator with unknown opcode?");
- return Expression::ADD;
+ default: // THIS SHOULD NEVER HAPPEN
+ assert(0 && "Binary operator with unknown opcode?");
+ case Instruction::Add: return Expression::ADD;
+ case Instruction::Sub: return Expression::SUB;
+ case Instruction::Mul: return Expression::MUL;
+ case Instruction::UDiv: return Expression::UDIV;
+ case Instruction::SDiv: return Expression::SDIV;
+ case Instruction::FDiv: return Expression::FDIV;
+ case Instruction::URem: return Expression::UREM;
+ case Instruction::SRem: return Expression::SREM;
+ case Instruction::FRem: return Expression::FREM;
+ case Instruction::Shl: return Expression::SHL;
+ case Instruction::LShr: return Expression::LSHR;
+ case Instruction::AShr: return Expression::ASHR;
+ case Instruction::And: return Expression::AND;
+ case Instruction::Or: return Expression::OR;
+ case Instruction::Xor: return Expression::XOR;
}
}
Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
- if (C->getOpcode() == Instruction::ICmp) {
- switch (C->getPredicate()) {
- case ICmpInst::ICMP_EQ:
- return Expression::ICMPEQ;
- case ICmpInst::ICMP_NE:
- return Expression::ICMPNE;
- case ICmpInst::ICMP_UGT:
- return Expression::ICMPUGT;
- case ICmpInst::ICMP_UGE:
- return Expression::ICMPUGE;
- case ICmpInst::ICMP_ULT:
- return Expression::ICMPULT;
- case ICmpInst::ICMP_ULE:
- return Expression::ICMPULE;
- case ICmpInst::ICMP_SGT:
- return Expression::ICMPSGT;
- case ICmpInst::ICMP_SGE:
- return Expression::ICMPSGE;
- case ICmpInst::ICMP_SLT:
- return Expression::ICMPSLT;
- case ICmpInst::ICMP_SLE:
- return Expression::ICMPSLE;
-
- // THIS SHOULD NEVER HAPPEN
- default:
- assert(0 && "Comparison with unknown predicate?");
- return Expression::ICMPEQ;
- }
- } else {
+ if (isa<ICmpInst>(C)) {
switch (C->getPredicate()) {
- case FCmpInst::FCMP_OEQ:
- return Expression::FCMPOEQ;
- case FCmpInst::FCMP_OGT:
- return Expression::FCMPOGT;
- case FCmpInst::FCMP_OGE:
- return Expression::FCMPOGE;
- case FCmpInst::FCMP_OLT:
- return Expression::FCMPOLT;
- case FCmpInst::FCMP_OLE:
- return Expression::FCMPOLE;
- case FCmpInst::FCMP_ONE:
- return Expression::FCMPONE;
- case FCmpInst::FCMP_ORD:
- return Expression::FCMPORD;
- case FCmpInst::FCMP_UNO:
- return Expression::FCMPUNO;
- case FCmpInst::FCMP_UEQ:
- return Expression::FCMPUEQ;
- case FCmpInst::FCMP_UGT:
- return Expression::FCMPUGT;
- case FCmpInst::FCMP_UGE:
- return Expression::FCMPUGE;
- case FCmpInst::FCMP_ULT:
- return Expression::FCMPULT;
- case FCmpInst::FCMP_ULE:
- return Expression::FCMPULE;
- case FCmpInst::FCMP_UNE:
- return Expression::FCMPUNE;
-
- // THIS SHOULD NEVER HAPPEN
- default:
- assert(0 && "Comparison with unknown predicate?");
- return Expression::FCMPOEQ;
+ default: // THIS SHOULD NEVER HAPPEN
+ assert(0 && "Comparison with unknown predicate?");
+ case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
+ case ICmpInst::ICMP_NE: return Expression::ICMPNE;
+ case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
+ case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
+ case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
+ case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
+ case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
+ case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
+ case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
+ case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
}
}
+ assert(isa<FCmpInst>(C) && "Unknown compare");
+ switch (C->getPredicate()) {
+ default: // THIS SHOULD NEVER HAPPEN
+ assert(0 && "Comparison with unknown predicate?");
+ case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
+ case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
+ case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
+ case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
+ case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
+ case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
+ case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
+ case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
+ case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
+ case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
+ case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
+ case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
+ case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
+ case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
+ }
}
-Expression::ExpressionOpcode
- ValueTable::getOpcode(CastInst* C) {
+Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
switch(C->getOpcode()) {
- case Instruction::Trunc:
- return Expression::TRUNC;
- case Instruction::ZExt:
- return Expression::ZEXT;
- case Instruction::SExt:
- return Expression::SEXT;
- case Instruction::FPToUI:
- return Expression::FPTOUI;
- case Instruction::FPToSI:
- return Expression::FPTOSI;
- case Instruction::UIToFP:
- return Expression::UITOFP;
- case Instruction::SIToFP:
- return Expression::SITOFP;
- case Instruction::FPTrunc:
- return Expression::FPTRUNC;
- case Instruction::FPExt:
- return Expression::FPEXT;
- case Instruction::PtrToInt:
- return Expression::PTRTOINT;
- case Instruction::IntToPtr:
- return Expression::INTTOPTR;
- case Instruction::BitCast:
- return Expression::BITCAST;
-
- // THIS SHOULD NEVER HAPPEN
- default:
- assert(0 && "Cast operator with unknown opcode?");
- return Expression::BITCAST;
+ default: // THIS SHOULD NEVER HAPPEN
+ assert(0 && "Cast operator with unknown opcode?");
+ case Instruction::Trunc: return Expression::TRUNC;
+ case Instruction::ZExt: return Expression::ZEXT;
+ case Instruction::SExt: return Expression::SEXT;
+ case Instruction::FPToUI: return Expression::FPTOUI;
+ case Instruction::FPToSI: return Expression::FPTOSI;
+ case Instruction::UIToFP: return Expression::UITOFP;
+ case Instruction::SIToFP: return Expression::SITOFP;
+ case Instruction::FPTrunc: return Expression::FPTRUNC;
+ case Instruction::FPExt: return Expression::FPEXT;
+ case Instruction::PtrToInt: return Expression::PTRTOINT;
+ case Instruction::IntToPtr: return Expression::INTTOPTR;
+ case Instruction::BitCast: return Expression::BITCAST;
}
}
/// the value has not yet been numbered.
uint32_t ValueTable::lookup(Value* V) const {
DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
- if (VI != valueNumbering.end())
- return VI->second;
- else
- assert(0 && "Value not numbered?");
-
- return 0;
+ assert(VI != valueNumbering.end() && "Value not numbered?");
+ return VI->second;
}
/// clear - Remove all entries from the ValueTable
// ValueNumberedSet Class
//===----------------------------------------------------------------------===//
namespace {
-class ValueNumberedSet {
+class VISIBILITY_HIDDEN ValueNumberedSet {
private:
SmallPtrSet<Value*, 8> contents;
BitVector numbers;
Value* find_leader(ValueNumberedSet& vals, uint32_t v) ;
void val_insert(ValueNumberedSet& s, Value* v);
bool processLoad(LoadInst* L,
- DenseMap<Value*, LoadInst*>& lastLoad,
- SmallVector<Instruction*, 4>& toErase);
+ DenseMap<Value*, LoadInst*> &lastLoad,
+ SmallVectorImpl<Instruction*> &toErase);
+ bool processStore(StoreInst *SI, SmallVectorImpl<Instruction*> &toErase);
bool processInstruction(Instruction* I,
ValueNumberedSet& currAvail,
DenseMap<Value*, LoadInst*>& lastSeenLoad,
- SmallVector<Instruction*, 4>& toErase);
+ SmallVectorImpl<Instruction*> &toErase);
bool processNonLocalLoad(LoadInst* L,
- SmallVector<Instruction*, 4>& toErase);
+ SmallVectorImpl<Instruction*> &toErase);
bool processMemCpy(MemCpyInst* M, MemCpyInst* MDep,
- SmallVector<Instruction*, 4>& toErase);
- bool performReturnSlotOptzn(MemCpyInst* cpy, CallInst* C,
- SmallVector<Instruction*, 4>& toErase);
+ SmallVectorImpl<Instruction*> &toErase);
+ bool performCallSlotOptzn(MemCpyInst* cpy, CallInst* C,
+ SmallVectorImpl<Instruction*> &toErase);
Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
DenseMap<BasicBlock*, Value*> &Phis,
bool top_level = false);
};
char GVN::ID = 0;
-
}
// createGVNPass - The public interface to this file...
static RegisterPass<GVN> X("gvn",
"Global Value Numbering");
-STATISTIC(NumGVNInstr, "Number of instructions deleted");
-STATISTIC(NumGVNLoad, "Number of loads deleted");
-
/// find_leader - Given a set and a value number, return the first
/// element of the set with that value number, or 0 if no such element
/// is present
DominatorTree &DT = getAnalysis<DominatorTree>();
Value* constVal = p->hasConstantValue();
- if (constVal) {
- if (Instruction* inst = dyn_cast<Instruction>(constVal)) {
- if (DT.dominates(inst, p))
- if (isSafeReplacement(p, inst))
- return inst;
- } else {
- return constVal;
- }
- }
+ if (!constVal) return 0;
+ Instruction* inst = dyn_cast<Instruction>(constVal);
+ if (!inst)
+ return constVal;
+
+ if (DT.dominates(inst, p))
+ if (isSafeReplacement(p, inst))
+ return inst;
return 0;
}
/// GetValueForBlock - Get the value to use within the specified basic block.
/// available values are in Phis.
Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
- DenseMap<BasicBlock*, Value*> &Phis,
- bool top_level) {
+ DenseMap<BasicBlock*, Value*> &Phis,
+ bool top_level) {
// If we have already computed this value, return the previously computed val.
DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
Phis[BB] = ret;
return ret;
}
+
// Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
// now, then get values to fill in the incoming values for the PHI.
PHINode *PN = new PHINode(orig->getType(), orig->getName()+".rle",
// Fill in the incoming values for the block.
for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
Value* val = GetValueForBlock(*PI, orig, Phis);
-
PN->addIncoming(val, *PI);
}
+
AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
AA.copyValue(orig, PN);
// Attempt to collapse PHI nodes that are trivially redundant
Value* v = CollapsePhi(PN);
- if (v) {
- MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
-
- MD.removeInstruction(PN);
- PN->replaceAllUsesWith(v);
-
- for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
- E = Phis.end(); I != E; ++I)
- if (I->second == PN)
- I->second = v;
+ if (!v) {
+ // Cache our phi construction results
+ phiMap[orig->getPointerOperand()].insert(PN);
+ return PN;
+ }
+
+ MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
- PN->eraseFromParent();
+ MD.removeInstruction(PN);
+ PN->replaceAllUsesWith(v);
- Phis[BB] = v;
+ for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
+ E = Phis.end(); I != E; ++I)
+ if (I->second == PN)
+ I->second = v;
- return v;
- }
+ PN->eraseFromParent();
- // Cache our phi construction results
- phiMap[orig->getPointerOperand()].insert(PN);
- return PN;
+ Phis[BB] = v;
+ return v;
}
/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
/// non-local by performing PHI construction.
bool GVN::processNonLocalLoad(LoadInst* L,
- SmallVector<Instruction*, 4>& toErase) {
+ SmallVectorImpl<Instruction*> &toErase) {
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
// Find the non-local dependencies of the load
// Filter out useless results (non-locals, etc)
for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), E = deps.end();
- I != E; ++I)
- if (I->second == MemoryDependenceAnalysis::None) {
+ I != E; ++I) {
+ if (I->second == MemoryDependenceAnalysis::None)
return false;
- } else if (I->second == MemoryDependenceAnalysis::NonLocal) {
+
+ if (I->second == MemoryDependenceAnalysis::NonLocal)
continue;
- } else if (StoreInst* S = dyn_cast<StoreInst>(I->second)) {
- if (S->getPointerOperand() == L->getPointerOperand())
- repl[I->first] = S->getOperand(0);
- else
+
+ if (StoreInst* S = dyn_cast<StoreInst>(I->second)) {
+ if (S->getPointerOperand() != L->getPointerOperand())
return false;
+ repl[I->first] = S->getOperand(0);
} else if (LoadInst* LD = dyn_cast<LoadInst>(I->second)) {
- if (LD->getPointerOperand() == L->getPointerOperand())
- repl[I->first] = LD;
- else
+ if (LD->getPointerOperand() != L->getPointerOperand())
return false;
+ repl[I->first] = LD;
} else {
return false;
}
+ }
// Use cached PHI construction information from previous runs
SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()];
L->replaceAllUsesWith(*I);
toErase.push_back(L);
NumGVNLoad++;
-
return true;
- } else {
- repl.insert(std::make_pair((*I)->getParent(), *I));
}
+
+ repl.insert(std::make_pair((*I)->getParent(), *I));
}
// Perform PHI construction
/// processLoad - Attempt to eliminate a load, first by eliminating it
/// locally, and then attempting non-local elimination if that fails.
-bool GVN::processLoad(LoadInst* L,
- DenseMap<Value*, LoadInst*>& lastLoad,
- SmallVector<Instruction*, 4>& toErase) {
+bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad,
+ SmallVectorImpl<Instruction*> &toErase) {
if (L->isVolatile()) {
lastLoad[L->getPointerOperand()] = L;
return false;
return deletedLoad;
}
-/// isReturnSlotOptznProfitable - Determine if performing a return slot
-/// fusion with the slot dest is profitable
-static bool isReturnSlotOptznProfitable(Value* dest, MemCpyInst* cpy) {
- // We currently consider it profitable if dest is otherwise dead.
- SmallVector<User*, 8> useList(dest->use_begin(), dest->use_end());
- while (!useList.empty()) {
- User* UI = useList.back();
+/// isBytewiseValue - If the specified value can be set by repeating the same
+/// byte in memory, return the i8 value that it is represented with. This is
+/// true for all i8 values obviously, but is also true for i32 0, i32 -1,
+/// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
+/// byte store (e.g. i16 0x1234), return null.
+static Value *isBytewiseValue(Value *V) {
+ // All byte-wide stores are splatable, even of arbitrary variables.
+ if (V->getType() == Type::Int8Ty) return V;
+
+ // Constant float and double values can be handled as integer values if the
+ // corresponding integer value is "byteable". An important case is 0.0.
+ if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
+ if (CFP->getType() == Type::FloatTy)
+ V = ConstantExpr::getBitCast(CFP, Type::Int32Ty);
+ if (CFP->getType() == Type::DoubleTy)
+ V = ConstantExpr::getBitCast(CFP, Type::Int64Ty);
+ // Don't handle long double formats, which have strange constraints.
+ }
+
+ // We can handle constant integers that are power of two in size and a
+ // multiple of 8 bits.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
+ unsigned Width = CI->getBitWidth();
+ if (isPowerOf2_32(Width) && Width > 8) {
+ // We can handle this value if the recursive binary decomposition is the
+ // same at all levels.
+ APInt Val = CI->getValue();
+ APInt Val2;
+ while (Val.getBitWidth() != 8) {
+ unsigned NextWidth = Val.getBitWidth()/2;
+ Val2 = Val.lshr(NextWidth);
+ Val2.trunc(Val.getBitWidth()/2);
+ Val.trunc(Val.getBitWidth()/2);
+
+ // If the top/bottom halves aren't the same, reject it.
+ if (Val != Val2)
+ return 0;
+ }
+ return ConstantInt::get(Val);
+ }
+ }
+
+ // Conceptually, we could handle things like:
+ // %a = zext i8 %X to i16
+ // %b = shl i16 %a, 8
+ // %c = or i16 %a, %b
+ // but until there is an example that actually needs this, it doesn't seem
+ // worth worrying about.
+ return 0;
+}
+
+static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx,
+ bool &VariableIdxFound, TargetData &TD) {
+ // Skip over the first indices.
+ gep_type_iterator GTI = gep_type_begin(GEP);
+ for (unsigned i = 1; i != Idx; ++i, ++GTI)
+ /*skip along*/;
+
+ // Compute the offset implied by the rest of the indices.
+ int64_t Offset = 0;
+ for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
+ ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
+ if (OpC == 0)
+ return VariableIdxFound = true;
+ if (OpC->isZero()) continue; // No offset.
+
+ // Handle struct indices, which add their field offset to the pointer.
+ if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
+ Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
+ continue;
+ }
- if (isa<GetElementPtrInst>(UI) || isa<BitCastInst>(UI)) {
- useList.pop_back();
- for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
- I != E; ++I)
- useList.push_back(*I);
- } else if (UI == cpy)
- useList.pop_back();
- else
- return false;
+ // Otherwise, we have a sequential type like an array or vector. Multiply
+ // the index by the ElementSize.
+ uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
+ Offset += Size*OpC->getSExtValue();
+ }
+
+ return Offset;
+}
+
+/// IsPointerAtOffset - Return true if Ptr1 is exactly provably equal to Ptr2
+/// plus the specified constant offset. For example, Ptr1 might be &A[42], and
+/// Ptr2 might be &A[40] and Offset might be 8.
+static bool IsPointerAtOffset(Value *Ptr1, Value *Ptr2, uint64_t Offset,
+ TargetData &TD) {
+ // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
+ // base. After that base, they may have some number of common (and
+ // potentially variable) indices. After that they handle some constant
+ // offset, which determines their offset from each other. At this point, we
+ // handle no other case.
+ GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
+ GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
+ if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
+ return false;
+
+ // Skip any common indices and track the GEP types.
+ unsigned Idx = 1;
+ for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
+ if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
+ break;
+
+ bool VariableIdxFound = false;
+ int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
+ int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
+ if (VariableIdxFound) return false;
+
+ return Offset1 == Offset2+(int64_t)Offset;
+}
+
+
+/// processStore - When GVN is scanning forward over instructions, we look for
+/// some other patterns to fold away. In particular, this looks for stores to
+/// neighboring locations of memory. If it sees enough consequtive ones
+/// (currently 4) it attempts to merge them together into a memcpy/memset.
+bool GVN::processStore(StoreInst *SI, SmallVectorImpl<Instruction*> &toErase) {
+ if (SI->isVolatile()) return false;
+
+ // There are two cases that are interesting for this code to handle: memcpy
+ // and memset. Right now we only handle memset.
+
+ // Ensure that the value being stored is something that can be memset'able a
+ // byte at a time like "0" or "-1" or any width, as well as things like
+ // 0xA0A0A0A0 and 0.0.
+ Value *ByteVal = isBytewiseValue(SI->getOperand(0));
+ if (!ByteVal)
+ return false;
+
+ TargetData &TD = getAnalysis<TargetData>();
+ AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+
+ // Okay, so we now have a single store that can be splatable. Try to 'grow'
+ // this store by looking for neighboring stores to the immediate left or right
+ // of the store we have so far. While we could in theory handle stores in
+ // this order: A[0], A[2], A[1]
+ // in practice, right now we only worry about cases where stores are
+ // consequtive in increasing or decreasing address order.
+ uint64_t BytesSoFar = TD.getTypeStoreSize(SI->getOperand(0)->getType());
+ uint64_t BytesFromSI = 0;
+ unsigned StartAlign = SI->getAlignment();
+ Value *StartPtr = SI->getPointerOperand();
+ SmallVector<StoreInst*, 16> Stores;
+ Stores.push_back(SI);
+
+ BasicBlock::iterator BI = SI;
+ for (++BI; !isa<TerminatorInst>(BI); ++BI) {
+ if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
+ // If the call is readnone, ignore it, otherwise bail out. We don't even
+ // allow readonly here because we don't want something like:
+ // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
+ if (AA.getModRefBehavior(CallSite::get(BI)) ==
+ AliasAnalysis::DoesNotAccessMemory)
+ continue;
+
+ // TODO: If this is a memset, try to join it in.
+
+ break;
+ } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI))
+ break;
+
+ // If this is a non-store instruction it is fine, ignore it.
+ StoreInst *NextStore = dyn_cast<StoreInst>(BI);
+ if (NextStore == 0) continue;
+
+ // If this is a store, see if we can merge it in.
+ if (NextStore->isVolatile()) break;
+
+ // Check to see if this stored value is of the same byte-splattable value.
+ if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
+ break;
+
+ Value *ThisPointer = NextStore->getPointerOperand();
+ unsigned AccessSize = TD.getTypeStoreSize(SI->getOperand(0)->getType());
+
+ // If so, check to see if the store is before the current range or after it
+ // in either case, extend the range, otherwise reject it.
+ if (IsPointerAtOffset(ThisPointer, StartPtr, BytesSoFar, TD)) {
+ // Okay, this extends the stored area on the end, just add to the bytes
+ // so far and remember this store.
+ BytesSoFar += AccessSize;
+ Stores.push_back(NextStore);
+ continue;
+ }
+
+ if (IsPointerAtOffset(StartPtr, ThisPointer, AccessSize, TD)) {
+ // Okay, the store is before the current range. Reset our start pointer
+ // and get new alignment info etc.
+ BytesSoFar += AccessSize;
+ BytesFromSI += AccessSize;
+ Stores.push_back(NextStore);
+ StartPtr = ThisPointer;
+ StartAlign = NextStore->getAlignment();
+ continue;
+ }
+
+ // Otherwise, this store wasn't contiguous with our current range, bail out.
+ break;
}
+ // If we found less than 4 stores to merge, bail out, it isn't worth losing
+ // type information in llvm IR to do the transformation.
+ if (Stores.size() < 4)
+ return false;
+
+ // Otherwise, we do want to transform this! Create a new memset. We put the
+ // memset right after the first store that we found in this block. This
+ // ensures that the caller will increment the iterator to the memset before
+ // it deletes all the stores.
+ BasicBlock::iterator InsertPt = SI; ++InsertPt;
+
+ Function *F = Intrinsic::getDeclaration(SI->getParent()->getParent()
+ ->getParent(), Intrinsic::memset_i64);
+
+ // StartPtr may not dominate the starting point. Instead of using it, base
+ // the destination pointer off the input to the first store in the block.
+ StartPtr = SI->getPointerOperand();
+
+ // Cast the start ptr to be i8* as memset requires.
+ const Type *i8Ptr = PointerType::getUnqual(Type::Int8Ty);
+ if (StartPtr->getType() != i8Ptr)
+ StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getNameStart(),
+ InsertPt);
+
+ // Offset the pointer if needed.
+ if (BytesFromSI)
+ StartPtr = new GetElementPtrInst(StartPtr, ConstantInt::get(Type::Int64Ty,
+ -BytesFromSI),
+ "ptroffset", InsertPt);
+
+ Value *Ops[] = {
+ StartPtr, ByteVal, // Start, value
+ ConstantInt::get(Type::Int64Ty, BytesSoFar), // size
+ ConstantInt::get(Type::Int32Ty, StartAlign) // align
+ };
+ new CallInst(F, Ops, Ops+4, "", InsertPt);
+
+ // Zap all the stores.
+ toErase.append(Stores.begin(), Stores.end());
+
+ ++NumMemSetInfer;
return true;
}
-/// performReturnSlotOptzn - takes a memcpy and a call that it depends on,
-/// and checks for the possibility of a return slot optimization by having
-/// the call write its result directly into the callees return parameter
-/// rather than using memcpy
-bool GVN::performReturnSlotOptzn(MemCpyInst* cpy, CallInst* C,
- SmallVector<Instruction*, 4>& toErase) {
+
+/// performCallSlotOptzn - takes a memcpy and a call that it depends on,
+/// and checks for the possibility of a call slot optimization by having
+/// the call write its result directly into the destination of the memcpy.
+bool GVN::performCallSlotOptzn(MemCpyInst *cpy, CallInst *C,
+ SmallVectorImpl<Instruction*> &toErase) {
+ // The general transformation to keep in mind is
+ //
+ // call @func(..., src, ...)
+ // memcpy(dest, src, ...)
+ //
+ // ->
+ //
+ // memcpy(dest, src, ...)
+ // call @func(..., dest, ...)
+ //
+ // Since moving the memcpy is technically awkward, we additionally check that
+ // src only holds uninitialized values at the moment of the call, meaning that
+ // the memcpy can be discarded rather than moved.
+
// Deliberately get the source and destination with bitcasts stripped away,
// because we'll need to do type comparisons based on the underlying type.
Value* cpyDest = cpy->getDest();
Value* cpySrc = cpy->getSource();
CallSite CS = CallSite::get(C);
-
- // Since this is a return slot optimization, we need to make sure that
- // the value being copied is, in fact, in a return slot. We also need to
- // check that the return slot parameter is marked noalias, so that we can
- // be sure that changing it will not cause unexpected behavior changes due
- // to it being accessed through a global or another parameter.
- if (CS.arg_size() == 0 ||
- cpySrc != CS.getArgument(0) ||
- !CS.paramHasAttr(1, ParamAttr::NoAlias | ParamAttr::StructRet))
- return false;
-
- // Check that something sneaky is not happening involving casting
- // return slot types around.
- if (CS.getArgument(0)->getType() != cpyDest->getType())
- return false;
- // sret --> pointer
- const PointerType* PT = cast<PointerType>(cpyDest->getType());
-
- // We can only perform the transformation if the size of the memcpy
- // is constant and equal to the size of the structure.
+
+ // We need to be able to reason about the size of the memcpy, so we require
+ // that it be a constant.
ConstantInt* cpyLength = dyn_cast<ConstantInt>(cpy->getLength());
if (!cpyLength)
return false;
-
+
+ // Require that src be an alloca. This simplifies the reasoning considerably.
+ AllocaInst* srcAlloca = dyn_cast<AllocaInst>(cpySrc);
+ if (!srcAlloca)
+ return false;
+
+ // Check that all of src is copied to dest.
TargetData& TD = getAnalysis<TargetData>();
- if (TD.getTypeStoreSize(PT->getElementType()) != cpyLength->getZExtValue())
+
+ ConstantInt* srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
+ if (!srcArraySize)
return false;
-
- // We only perform the transformation if it will be profitable.
- if (!isReturnSlotOptznProfitable(cpyDest, cpy))
+
+ uint64_t srcSize = TD.getABITypeSize(srcAlloca->getAllocatedType()) *
+ srcArraySize->getZExtValue();
+
+ if (cpyLength->getZExtValue() < srcSize)
return false;
-
- // In addition to knowing that the call does not access the return slot
- // in some unexpected manner, which we derive from the noalias attribute,
- // we also need to know that it does not sneakily modify the destination
- // slot in the caller. We don't have parameter attributes to go by
- // for this one, so we just rely on AA to figure it out for us.
+
+ // Check that accessing the first srcSize bytes of dest will not cause a
+ // trap. Otherwise the transform is invalid since it might cause a trap
+ // to occur earlier than it otherwise would.
+ if (AllocaInst* A = dyn_cast<AllocaInst>(cpyDest)) {
+ // The destination is an alloca. Check it is larger than srcSize.
+ ConstantInt* destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
+ if (!destArraySize)
+ return false;
+
+ uint64_t destSize = TD.getABITypeSize(A->getAllocatedType()) *
+ destArraySize->getZExtValue();
+
+ if (destSize < srcSize)
+ return false;
+ } else if (Argument* A = dyn_cast<Argument>(cpyDest)) {
+ // If the destination is an sret parameter then only accesses that are
+ // outside of the returned struct type can trap.
+ if (!A->hasStructRetAttr())
+ return false;
+
+ const Type* StructTy = cast<PointerType>(A->getType())->getElementType();
+ uint64_t destSize = TD.getABITypeSize(StructTy);
+
+ if (destSize < srcSize)
+ return false;
+ } else {
+ return false;
+ }
+
+ // Check that src is not accessed except via the call and the memcpy. This
+ // guarantees that it holds only undefined values when passed in (so the final
+ // memcpy can be dropped), that it is not read or written between the call and
+ // the memcpy, and that writing beyond the end of it is undefined.
+ SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
+ srcAlloca->use_end());
+ while (!srcUseList.empty()) {
+ User* UI = srcUseList.back();
+ srcUseList.pop_back();
+
+ if (isa<GetElementPtrInst>(UI) || isa<BitCastInst>(UI)) {
+ for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
+ I != E; ++I)
+ srcUseList.push_back(*I);
+ } else if (UI != C && UI != cpy) {
+ return false;
+ }
+ }
+
+ // Since we're changing the parameter to the callsite, we need to make sure
+ // that what would be the new parameter dominates the callsite.
+ DominatorTree& DT = getAnalysis<DominatorTree>();
+ if (Instruction* cpyDestInst = dyn_cast<Instruction>(cpyDest))
+ if (!DT.dominates(cpyDestInst, C))
+ return false;
+
+ // In addition to knowing that the call does not access src in some
+ // unexpected manner, for example via a global, which we deduce from
+ // the use analysis, we also need to know that it does not sneakily
+ // access dest. We rely on AA to figure this out for us.
AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
- if (AA.getModRefInfo(C, cpy->getRawDest(), cpyLength->getZExtValue()) !=
+ if (AA.getModRefInfo(C, cpy->getRawDest(), srcSize) !=
AliasAnalysis::NoModRef)
return false;
-
- // If all the checks have passed, then we're alright to do the transformation.
- CS.setArgument(0, cpyDest);
-
+
+ // All the checks have passed, so do the transformation.
+ for (unsigned i = 0; i < CS.arg_size(); ++i)
+ if (CS.getArgument(i) == cpySrc) {
+ if (cpySrc->getType() != cpyDest->getType())
+ cpyDest = CastInst::createPointerCast(cpyDest, cpySrc->getType(),
+ cpyDest->getName(), C);
+ CS.setArgument(i, cpyDest);
+ }
+
// Drop any cached information about the call, because we may have changed
// its dependence information by changing its parameter.
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
MD.dropInstruction(C);
-
+
// Remove the memcpy
MD.removeInstruction(cpy);
toErase.push_back(cpy);
-
+
return true;
}
/// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
/// This allows later passes to remove the first memcpy altogether.
bool GVN::processMemCpy(MemCpyInst* M, MemCpyInst* MDep,
- SmallVector<Instruction*, 4>& toErase) {
+ SmallVectorImpl<Instruction*> &toErase) {
// We can only transforms memcpy's where the dest of one is the source of the
// other
if (M->getSource() != MDep->getDest())
if (!C1 || !C2)
return false;
- uint64_t CpySize = C1->getValue().getZExtValue();
- uint64_t DepSize = C2->getValue().getZExtValue();
+ uint64_t DepSize = C1->getValue().getZExtValue();
+ uint64_t CpySize = C2->getValue().getZExtValue();
if (DepSize < CpySize)
return false;
MD.dropInstruction(M);
toErase.push_back(M);
return true;
- } else {
- MD.removeInstruction(C);
- toErase.push_back(C);
- return false;
}
+
+ MD.removeInstruction(C);
+ toErase.push_back(C);
+ return false;
}
/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
-bool GVN::processInstruction(Instruction* I,
- ValueNumberedSet& currAvail,
- DenseMap<Value*, LoadInst*>& lastSeenLoad,
- SmallVector<Instruction*, 4>& toErase) {
- if (LoadInst* L = dyn_cast<LoadInst>(I)) {
+bool GVN::processInstruction(Instruction *I, ValueNumberedSet &currAvail,
+ DenseMap<Value*, LoadInst*> &lastSeenLoad,
+ SmallVectorImpl<Instruction*> &toErase) {
+ if (LoadInst* L = dyn_cast<LoadInst>(I))
return processLoad(L, lastSeenLoad, toErase);
- } else if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) {
+
+ if (StoreInst *SI = dyn_cast<StoreInst>(I))
+ return processStore(SI, toErase);
+
+ if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) {
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
// The are two possible optimizations we can do for memcpy:
// a) memcpy-memcpy xform which exposes redundance for DSE
- // b) call-memcpy xform for sret return slot optimization
+ // b) call-memcpy xform for return slot optimization
Instruction* dep = MD.getDependency(M);
if (dep == MemoryDependenceAnalysis::None ||
dep == MemoryDependenceAnalysis::NonLocal)
if (MemCpyInst *MemCpy = dyn_cast<MemCpyInst>(dep))
return processMemCpy(M, MemCpy, toErase);
if (CallInst* C = dyn_cast<CallInst>(dep))
- return performReturnSlotOptzn(M, C, toErase);
+ return performCallSlotOptzn(M, C, toErase);
return false;
}
DominatorTree &DT = getAnalysis<DominatorTree>();
SmallVector<Instruction*, 4> toErase;
-
+ DenseMap<Value*, LoadInst*> lastSeenLoad;
+
// Top-down walk of the dominator tree
for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
E = df_end(DT.getRootNode()); DI != E; ++DI) {
// Get the set to update for this block
ValueNumberedSet& currAvail = availableOut[DI->getBlock()];
- DenseMap<Value*, LoadInst*> lastSeenLoad;
-
+ lastSeenLoad.clear();
+
BasicBlock* BB = DI->getBlock();
// A block inherits AVAIL_OUT from its dominator
++BI;
for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
- E = toErase.end(); I != E; ++I) {
+ E = toErase.end(); I != E; ++I)
(*I)->eraseFromParent();
- }
toErase.clear();
}
projects / oota-llvm.git / blobdiff