#define DEBUG_TYPE "memcpyopt"
#include "llvm/Transforms/Scalar.h"
+#include "llvm/GlobalVariable.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Instructions.h"
-#include "llvm/LLVMContext.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
+#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/raw_ostream.h"
STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
STATISTIC(NumMemSetInfer, "Number of memsets inferred");
STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
-
-/// 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) {
- LLVMContext &Context = V->getContext();
-
- // All byte-wide stores are splatable, even of arbitrary variables.
- if (V->getType()->isIntegerTy(8)) 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()->isFloatTy())
- V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(Context));
- if (CFP->getType()->isDoubleTy())
- V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(Context));
- // 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(Context, 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;
-}
+STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx,
bool &VariableIdxFound, TargetData &TD) {
namespace {
class MemCpyOpt : public FunctionPass {
+ MemoryDependenceAnalysis *MD;
bool runOnFunction(Function &F);
public:
static char ID; // Pass identification, replacement for typeid
MemCpyOpt() : FunctionPass(ID) {
initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
+ MD = 0;
}
private:
bool processMemMove(MemMoveInst *M);
bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
uint64_t cpyLen, CallInst *C);
+ bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
+ uint64_t MSize);
+ bool processByValArgument(CallSite CS, unsigned ArgNo);
bool iterateOnFunction(Function &F);
};
// a memcpy.
if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
if (!LI->isVolatile() && LI->hasOneUse()) {
- MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
-
- MemDepResult dep = MD.getDependency(LI);
+ MemDepResult dep = MD->getDependency(LI);
CallInst *C = 0;
if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
C = dyn_cast<CallInst>(dep.getInst());
LI->getPointerOperand()->stripPointerCasts(),
TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
if (changed) {
- MD.removeInstruction(SI);
+ MD->removeInstruction(SI);
SI->eraseFromParent();
LI->eraseFromParent();
++NumMemCpyInstr;
// align
ConstantInt::get(Type::getInt32Ty(Context), Alignment),
// volatile
- ConstantInt::get(Type::getInt1Ty(Context), 0),
+ ConstantInt::getFalse(Context),
};
const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
Value *C = CallInst::Create(MemSetF, Ops, Ops+5, "", InsertPt);
DEBUG(dbgs() << "Replace stores:\n";
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
- dbgs() << *Range.TheStores[i];
- dbgs() << "With: " << *C); C=C;
+ dbgs() << *Range.TheStores[i] << '\n';
+ dbgs() << "With: " << *C << '\n'); (void)C;
// Don't invalidate the iterator
BBI = BI;
// Drop any cached information about the call, because we may have changed
// its dependence information by changing its parameter.
- MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
- MD.removeInstruction(C);
+ MD->removeInstruction(C);
- // Remove the memcpy
- MD.removeInstruction(cpy);
+ // Remove the memcpy.
+ MD->removeInstruction(cpy);
++NumMemCpyInstr;
return true;
}
-/// processMemCpy - perform simplification of memcpy's. If we have memcpy A
-/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
-/// B to be a memcpy from X to Z (or potentially a memmove, depending on
-/// circumstances). This allows later passes to remove the first memcpy
-/// altogether.
-bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
- MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
-
- // We can only optimize statically-sized memcpy's.
- ConstantInt *cpyLen = dyn_cast<ConstantInt>(M->getLength());
- if (!cpyLen) return false;
-
- // The are two possible optimizations we can do for memcpy:
- // a) memcpy-memcpy xform which exposes redundance for DSE.
- // b) call-memcpy xform for return slot optimization.
- MemDepResult dep = MD.getDependency(M);
- if (!dep.isClobber())
- return false;
- if (!isa<MemCpyInst>(dep.getInst())) {
- if (CallInst *C = dyn_cast<CallInst>(dep.getInst())) {
- bool changed = performCallSlotOptzn(M, M->getDest(), M->getSource(),
- cpyLen->getZExtValue(), C);
- if (changed) M->eraseFromParent();
- return changed;
- }
+/// processMemCpyMemCpyDependence - We've found that the (upward scanning)
+/// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
+/// copy from MDep's input if we can. MSize is the size of M's copy.
+///
+bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
+ uint64_t MSize) {
+ // We can only transforms memcpy's where the dest of one is the source of the
+ // other.
+ if (M->getSource() != MDep->getDest() || MDep->isVolatile())
return false;
- }
-
- MemCpyInst *MDep = cast<MemCpyInst>(dep.getInst());
- // We can only transforms memcpy's where the dest of one is the source of the
- // other
- if (M->getSource() != MDep->getDest())
+ // If dep instruction is reading from our current input, then it is a noop
+ // transfer and substituting the input won't change this instruction. Just
+ // ignore the input and let someone else zap MDep. This handles cases like:
+ // memcpy(a <- a)
+ // memcpy(b <- a)
+ if (M->getSource() == MDep->getSource())
return false;
// Second, the length of the memcpy's must be the same, or the preceeding one
// must be larger than the following one.
ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
- ConstantInt *C2 = dyn_cast<ConstantInt>(M->getLength());
- if (!C1 || !C2)
- return false;
-
- uint64_t DepSize = C1->getValue().getZExtValue();
- uint64_t CpySize = C2->getValue().getZExtValue();
+ if (!C1) return false;
- if (DepSize < CpySize)
+ AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+
+ // Verify that the copied-from memory doesn't change in between the two
+ // transfers. For example, in:
+ // memcpy(a <- b)
+ // *b = 42;
+ // memcpy(c <- a)
+ // It would be invalid to transform the second memcpy into memcpy(c <- b).
+ //
+ // TODO: If the code between M and MDep is transparent to the destination "c",
+ // then we could still perform the xform by moving M up to the first memcpy.
+ //
+ // NOTE: This is conservative, it will stop on any read from the source loc,
+ // not just the defining memcpy.
+ MemDepResult SourceDep =
+ MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
+ false, M, M->getParent());
+ if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
return false;
- // Finally, we have to make sure that the dest of the second does not
- // alias the source of the first
- AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
- if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
+ // If the dest of the second might alias the source of the first, then the
+ // source and dest might overlap. We still want to eliminate the intermediate
+ // value, but we have to generate a memmove instead of memcpy.
+ Intrinsic::ID ResultFn = Intrinsic::memcpy;
+ if (AA.alias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)) !=
AliasAnalysis::NoAlias)
- return false;
- else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) !=
- AliasAnalysis::NoAlias)
- return false;
- else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize)
- != AliasAnalysis::NoAlias)
- return false;
+ ResultFn = Intrinsic::memmove;
+
+ // If all checks passed, then we can transform M.
+ const Type *ArgTys[3] = {
+ M->getRawDest()->getType(),
+ MDep->getRawSource()->getType(),
+ M->getLength()->getType()
+ };
+ Function *MemCpyFun =
+ Intrinsic::getDeclaration(MDep->getParent()->getParent()->getParent(),
+ ResultFn, ArgTys, 3);
- // If all checks passed, then we can transform these memcpy's
- const Type *ArgTys[3] = { M->getRawDest()->getType(),
- MDep->getRawSource()->getType(),
- M->getLength()->getType() };
- Function *MemCpyFun = Intrinsic::getDeclaration(
- M->getParent()->getParent()->getParent(),
- M->getIntrinsicID(), ArgTys, 3);
-
// Make sure to use the lesser of the alignment of the source and the dest
// since we're changing where we're reading from, but don't want to increase
// the alignment past what can be read from or written to.
// TODO: Is this worth it if we're creating a less aligned memcpy? For
// example we could be moving from movaps -> movq on x86.
- unsigned Align = std::min(MDep->getAlignmentCst()->getZExtValue(),
- M->getAlignmentCst()->getZExtValue());
- LLVMContext &Context = M->getContext();
- ConstantInt *AlignCI = ConstantInt::get(Type::getInt32Ty(Context), Align);
+ unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
Value *Args[5] = {
- M->getRawDest(), MDep->getRawSource(), M->getLength(),
- AlignCI, M->getVolatileCst()
+ M->getRawDest(),
+ MDep->getRawSource(),
+ M->getLength(),
+ ConstantInt::get(Type::getInt32Ty(MemCpyFun->getContext()), Align),
+ M->getVolatileCst()
};
- CallInst *C = CallInst::Create(MemCpyFun, Args, Args+5, "", M);
-
- // If C and M don't interfere, then this is a valid transformation. If they
- // did, this would mean that the two sources overlap, which would be bad.
- if (MD.getDependency(C) == dep) {
- MD.removeInstruction(M);
+ CallInst::Create(MemCpyFun, Args, Args+5, "", M);
+
+ // Remove the instruction we're replacing.
+ MD->removeInstruction(M);
+ M->eraseFromParent();
+ ++NumMemCpyInstr;
+ return true;
+}
+
+
+/// processMemCpy - perform simplification of memcpy's. If we have memcpy A
+/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
+/// B to be a memcpy from X to Z (or potentially a memmove, depending on
+/// circumstances). This allows later passes to remove the first memcpy
+/// altogether.
+bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
+ // We can only optimize statically-sized memcpy's that are non-volatile.
+ ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
+ if (CopySize == 0 || M->isVolatile()) return false;
+
+ // If the source and destination of the memcpy are the same, then zap it.
+ if (M->getSource() == M->getDest()) {
+ MD->removeInstruction(M);
M->eraseFromParent();
- ++NumMemCpyInstr;
- return true;
+ return false;
}
+
+ // If copying from a constant, try to turn the memcpy into a memset.
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
+ if (GV->isConstant() && GV->hasDefinitiveInitializer())
+ if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
+ Value *Ops[] = {
+ M->getRawDest(), ByteVal, // Start, value
+ CopySize, // Size
+ M->getAlignmentCst(), // Alignment
+ ConstantInt::getFalse(M->getContext()), // volatile
+ };
+ const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
+ Module *Mod = M->getParent()->getParent()->getParent();
+ Function *MemSetF = Intrinsic::getDeclaration(Mod, Intrinsic::memset,
+ Tys, 2);
+ CallInst::Create(MemSetF, Ops, Ops+5, "", M);
+ MD->removeInstruction(M);
+ M->eraseFromParent();
+ ++NumCpyToSet;
+ return true;
+ }
+
+ // The are two possible optimizations we can do for memcpy:
+ // a) memcpy-memcpy xform which exposes redundance for DSE.
+ // b) call-memcpy xform for return slot optimization.
+ MemDepResult DepInfo = MD->getDependency(M);
+ if (!DepInfo.isClobber())
+ return false;
- // Otherwise, there was no point in doing this, so we remove the call we
- // inserted and act like nothing happened.
- MD.removeInstruction(C);
- C->eraseFromParent();
+ if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
+ return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
+
+ if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
+ if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
+ CopySize->getZExtValue(), C)) {
+ M->eraseFromParent();
+ return true;
+ }
+ }
return false;
}
bool MemCpyOpt::processMemMove(MemMoveInst *M) {
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
- // If the memmove is a constant size, use it for the alias query, this allows
- // us to optimize things like: memmove(P, P+64, 64);
- unsigned MemMoveSize = AliasAnalysis::UnknownSize;
- if (ConstantInt *Len = dyn_cast<ConstantInt>(M->getLength()))
- MemMoveSize = Len->getZExtValue();
-
// See if the pointers alias.
- if (AA.alias(M->getRawDest(), MemMoveSize, M->getRawSource(), MemMoveSize) !=
+ if (AA.alias(AA.getLocationForDest(M),
+ AA.getLocationForSource(M)) !=
AliasAnalysis::NoAlias)
return false;
// MemDep may have over conservative information about this instruction, just
// conservatively flush it from the cache.
- getAnalysis<MemoryDependenceAnalysis>().removeInstruction(M);
+ MD->removeInstruction(M);
++NumMoveToCpy;
return true;
}
+/// processByValArgument - This is called on every byval argument in call sites.
+bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
+ TargetData *TD = getAnalysisIfAvailable<TargetData>();
+ if (!TD) return false;
+
+ // Find out what feeds this byval argument.
+ Value *ByValArg = CS.getArgument(ArgNo);
+ const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType();
+ uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
+ MemDepResult DepInfo =
+ MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
+ true, CS.getInstruction(),
+ CS.getInstruction()->getParent());
+ if (!DepInfo.isClobber())
+ return false;
+
+ // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
+ // a memcpy, see if we can byval from the source of the memcpy instead of the
+ // result.
+ MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
+ if (MDep == 0 || MDep->isVolatile() ||
+ ByValArg->stripPointerCasts() != MDep->getDest())
+ return false;
+
+ // The length of the memcpy must be larger or equal to the size of the byval.
+ ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
+ if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
+ return false;
-// MemCpyOpt::iterateOnFunction - Executes one iteration of GVN.
+ // Get the alignment of the byval. If it is greater than the memcpy, then we
+ // can't do the substitution. If the call doesn't specify the alignment, then
+ // it is some target specific value that we can't know.
+ unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
+ if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign)
+ return false;
+
+ // Verify that the copied-from memory doesn't change in between the memcpy and
+ // the byval call.
+ // memcpy(a <- b)
+ // *b = 42;
+ // foo(*a)
+ // It would be invalid to transform the second memcpy into foo(*b).
+ //
+ // NOTE: This is conservative, it will stop on any read from the source loc,
+ // not just the defining memcpy.
+ MemDepResult SourceDep =
+ MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
+ false, CS.getInstruction(), MDep->getParent());
+ if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
+ return false;
+
+ Value *TmpCast = MDep->getSource();
+ if (MDep->getSource()->getType() != ByValArg->getType())
+ TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
+ "tmpcast", CS.getInstruction());
+
+ DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
+ << " " << *MDep << "\n"
+ << " " << *CS.getInstruction() << "\n");
+
+ // Otherwise we're good! Update the byval argument.
+ CS.setArgument(ArgNo, TmpCast);
+ ++NumMemCpyInstr;
+ return true;
+}
+
+/// iterateOnFunction - Executes one iteration of MemCpyOpt.
bool MemCpyOpt::iterateOnFunction(Function &F) {
bool MadeChange = false;
// Walk all instruction in the function.
for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
- for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
- BI != BE;) {
+ for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
// Avoid invalidating the iterator.
Instruction *I = BI++;
+ bool RepeatInstruction = false;
+
if (StoreInst *SI = dyn_cast<StoreInst>(I))
MadeChange |= processStore(SI, BI);
- else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
- MadeChange |= processMemCpy(M);
- else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) {
- if (processMemMove(M)) {
- --BI; // Reprocess the new memcpy.
- MadeChange = true;
- }
+ else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) {
+ RepeatInstruction = processMemCpy(M);
+ } else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) {
+ RepeatInstruction = processMemMove(M);
+ } else if (CallSite CS = (Value*)I) {
+ for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
+ if (CS.paramHasAttr(i+1, Attribute::ByVal))
+ MadeChange |= processByValArgument(CS, i);
+ }
+
+ // Reprocess the instruction if desired.
+ if (RepeatInstruction) {
+ --BI;
+ MadeChange = true;
}
}
}
//
bool MemCpyOpt::runOnFunction(Function &F) {
bool MadeChange = false;
+ MD = &getAnalysis<MemoryDependenceAnalysis>();
while (1) {
if (!iterateOnFunction(F))
break;
MadeChange = true;
}
+ MD = 0;
return MadeChange;
}
-
-
-
projects / oota-llvm.git / blobdiff