//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "scalarrepl"
#include "llvm/Transforms/Scalar.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/DIBuilder.h"
-#include "llvm/DebugInfo.h"
#include "llvm/IR/CallSite.h"
#include "llvm/IR/Constants.h"
+#include "llvm/IR/DIBuilder.h"
#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DebugInfo.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/Transforms/Utils/SSAUpdater.h"
using namespace llvm;
+#define DEBUG_TYPE "scalarrepl"
+
STATISTIC(NumReplaced, "Number of allocas broken up");
STATISTIC(NumPromoted, "Number of allocas promoted");
STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
private:
bool HasDomTree;
- const DataLayout *DL;
/// DeadInsts - Keep track of instructions we have made dead, so that
/// we can remove them after we are done working.
void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
Type *MemOpType, bool isStore, AllocaInfo &Info,
Instruction *TheAccess, bool AllowWholeAccess);
- bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size);
- uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset,
- Type *&IdxTy);
+ bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size,
+ const DataLayout &DL);
+ uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, Type *&IdxTy,
+ const DataLayout &DL);
void DoScalarReplacement(AllocaInst *AI,
std::vector<AllocaInst*> &WorkList);
// getAnalysisUsage - This pass does not require any passes, but we know it
// will not alter the CFG, so say so.
void getAnalysisUsage(AnalysisUsage &AU) const override {
+ AU.addRequired<AssumptionCacheTracker>();
AU.addRequired<DominatorTreeWrapperPass>();
AU.setPreservesCFG();
}
// getAnalysisUsage - This pass does not require any passes, but we know it
// will not alter the CFG, so say so.
void getAnalysisUsage(AnalysisUsage &AU) const override {
+ AU.addRequired<AssumptionCacheTracker>();
AU.setPreservesCFG();
}
};
INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
"Scalar Replacement of Aggregates (DT)", false, false)
+INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
"Scalar Replacement of Aggregates (DT)", false, false)
INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
"Scalar Replacement of Aggregates (SSAUp)", false, false)
+INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
"Scalar Replacement of Aggregates (SSAUp)", false, false)
explicit ConvertToScalarInfo(unsigned Size, const DataLayout &DL,
unsigned SLT)
: AllocaSize(Size), DL(DL), ScalarLoadThreshold(SLT), IsNotTrivial(false),
- ScalarKind(Unknown), VectorTy(0), HadNonMemTransferAccess(false),
+ ScalarKind(Unknown), VectorTy(nullptr), HadNonMemTransferAccess(false),
HadDynamicAccess(false) { }
AllocaInst *TryConvert(AllocaInst *AI);
AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
// If we can't convert this scalar, or if mem2reg can trivially do it, bail
// out.
- if (!CanConvertToScalar(AI, 0, 0) || !IsNotTrivial)
- return 0;
+ if (!CanConvertToScalar(AI, 0, nullptr) || !IsNotTrivial)
+ return nullptr;
// If an alloca has only memset / memcpy uses, it may still have an Unknown
// ScalarKind. Treat it as an Integer below.
// Do not convert to scalar integer if the alloca size exceeds the
// scalar load threshold.
if (BitWidth > ScalarLoadThreshold)
- return 0;
+ return nullptr;
if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
!HadNonMemTransferAccess && !DL.fitsInLegalInteger(BitWidth))
- return 0;
+ return nullptr;
// Dynamic accesses on integers aren't yet supported. They need us to shift
// by a dynamic amount which could be difficult to work out as we might not
// know whether to use a left or right shift.
if (ScalarKind == Integer && HadDynamicAccess)
- return 0;
+ return nullptr;
DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
// Create and insert the integer alloca.
NewTy = IntegerType::get(AI->getContext(), BitWidth);
}
- AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
- ConvertUsesToScalar(AI, NewAI, 0, 0);
+ AllocaInst *NewAI = new AllocaInst(NewTy, nullptr, "",
+ AI->getParent()->begin());
+ ConvertUsesToScalar(AI, NewAI, 0, nullptr);
return NewAI;
}
/// SawVec flag.
bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset,
Value* NonConstantIdx) {
- for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
- Instruction *User = cast<Instruction>(*UI);
+ for (User *U : V->users()) {
+ Instruction *UI = cast<Instruction>(U);
- if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
+ if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
// Don't break volatile loads.
if (!LI->isSimple())
return false;
continue;
}
- if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
+ if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
// Storing the pointer, not into the value?
if (SI->getOperand(0) == V || !SI->isSimple()) return false;
// Don't touch MMX operations.
continue;
}
- if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(UI)) {
if (!onlyUsedByLifetimeMarkers(BCI))
IsNotTrivial = true; // Can't be mem2reg'd.
if (!CanConvertToScalar(BCI, Offset, NonConstantIdx))
continue;
}
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UI)) {
// If this is a GEP with a variable indices, we can't handle it.
PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType());
if (!PtrTy)
// Compute the offset that this GEP adds to the pointer.
SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
- Value *GEPNonConstantIdx = 0;
+ Value *GEPNonConstantIdx = nullptr;
if (!GEP->hasAllConstantIndices()) {
if (!isa<VectorType>(PtrTy->getElementType()))
return false;
// If this is a constant sized memset of a constant value (e.g. 0) we can
// handle it.
- if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
+ if (MemSetInst *MSI = dyn_cast<MemSetInst>(UI)) {
// Store to dynamic index.
if (NonConstantIdx)
return false;
// If this is a memcpy or memmove into or out of the whole allocation, we
// can handle it like a load or store of the scalar type.
- if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
+ if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(UI)) {
// Store to dynamic index.
if (NonConstantIdx)
return false;
ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
- if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
+ if (!Len || Len->getZExtValue() != AllocaSize || Offset != 0)
return false;
IsNotTrivial = true; // Can't be mem2reg'd.
}
// If this is a lifetime intrinsic, we can handle it.
- if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(UI)) {
if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
II->getIntrinsicID() == Intrinsic::lifetime_end) {
continue;
uint64_t Offset,
Value* NonConstantIdx) {
while (!Ptr->use_empty()) {
- Instruction *User = cast<Instruction>(Ptr->use_back());
+ Instruction *User = cast<Instruction>(Ptr->user_back());
if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx);
if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
// Compute the offset that this GEP adds to the pointer.
SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
- Value* GEPNonConstantIdx = 0;
+ Value* GEPNonConstantIdx = nullptr;
if (!GEP->hasAllConstantIndices()) {
assert(!NonConstantIdx &&
"Dynamic GEP reading from dynamic GEP unsupported");
Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
Value *New = ConvertScalar_InsertValue(
ConstantInt::get(User->getContext(), APVal),
- Old, Offset, 0, Builder);
+ Old, Offset, nullptr, Builder);
Builder.CreateStore(New, NewAI);
// If the load we just inserted is now dead, then the memset overwrote
// If the source and destination are both to the same alloca, then this is
// a noop copy-to-self, just delete it. Otherwise, emit a load and store
// as appropriate.
- AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &DL, 0));
+ AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, DL, 0));
- if (GetUnderlyingObject(MTI->getSource(), &DL, 0) != OrigAI) {
+ if (GetUnderlyingObject(MTI->getSource(), DL, 0) != OrigAI) {
// Dest must be OrigAI, change this to be a load from the original
// pointer (bitcasted), then a store to our new alloca.
assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
SrcVal->setAlignment(MTI->getAlignment());
Builder.CreateStore(SrcVal, NewAI);
- } else if (GetUnderlyingObject(MTI->getDest(), &DL, 0) != OrigAI) {
+ } else if (GetUnderlyingObject(MTI->getDest(), DL, 0) != OrigAI) {
// Src must be OrigAI, change this to be a load from NewAI then a store
// through the original dest pointer (bitcasted).
assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
Offset+Layout.getElementOffsetInBits(i),
- 0, Builder);
+ nullptr, Builder);
Res = Builder.CreateInsertValue(Res, Elt, i);
}
return Res;
Value *Res = UndefValue::get(AT);
for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
- Offset+i*EltSize, 0, Builder);
+ Offset+i*EltSize, nullptr,
+ Builder);
Res = Builder.CreateInsertValue(Res, Elt, i);
}
return Res;
Value *Elt = Builder.CreateExtractValue(SV, i);
Old = ConvertScalar_InsertValue(Elt, Old,
Offset+Layout.getElementOffsetInBits(i),
- 0, Builder);
+ nullptr, Builder);
}
return Old;
}
uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType());
for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
Value *Elt = Builder.CreateExtractValue(SV, i);
- Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, 0, Builder);
+ Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, nullptr,
+ Builder);
}
return Old;
}
if (skipOptnoneFunction(F))
return false;
- DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
- DL = DLP ? &DLP->getDataLayout() : 0;
-
bool Changed = performPromotion(F);
- // FIXME: ScalarRepl currently depends on DataLayout more than it
- // theoretically needs to. It should be refactored in order to support
- // target-independent IR. Until this is done, just skip the actual
- // scalar-replacement portion of this pass.
- if (!DL) return Changed;
-
while (1) {
bool LocalChange = performScalarRepl(F);
if (!LocalChange) break; // No need to repromote if no scalarrepl
public:
AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
DIBuilder *DB)
- : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {}
+ : LoadAndStorePromoter(Insts, S), AI(nullptr), DIB(DB) {}
void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
// Remember which alloca we're promoting (for isInstInList).
this->AI = AI;
- if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) {
- for (Value::use_iterator UI = DebugNode->use_begin(),
- E = DebugNode->use_end(); UI != E; ++UI)
- if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
- DDIs.push_back(DDI);
- else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
- DVIs.push_back(DVI);
+ if (auto *L = LocalAsMetadata::getIfExists(AI)) {
+ if (auto *DebugNode = MetadataAsValue::getIfExists(AI->getContext(), L)) {
+ for (User *U : DebugNode->users())
+ if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
+ DDIs.push_back(DDI);
+ else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
+ DVIs.push_back(DVI);
+ }
}
LoadAndStorePromoter::run(Insts);
for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
E = DVIs.end(); I != E; ++I) {
DbgValueInst *DVI = *I;
- Value *Arg = NULL;
+ Value *Arg = nullptr;
if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
// If an argument is zero extended then use argument directly. The ZExt
// may be zapped by an optimization pass in future.
} else {
continue;
}
- Instruction *DbgVal =
- DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
- Inst);
- DbgVal->setDebugLoc(DVI->getDebugLoc());
+ DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
+ DIExpression(DVI->getExpression()),
+ DVI->getDebugLoc(), Inst);
}
}
};
///
/// We can do this to a select if its only uses are loads and if the operand to
/// the select can be loaded unconditionally.
-static bool isSafeSelectToSpeculate(SelectInst *SI, const DataLayout *DL) {
- bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
- bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
+static bool isSafeSelectToSpeculate(SelectInst *SI) {
+ const DataLayout &DL = SI->getModule()->getDataLayout();
+ bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(DL);
+ bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(DL);
- for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
- UI != UE; ++UI) {
- LoadInst *LI = dyn_cast<LoadInst>(*UI);
- if (LI == 0 || !LI->isSimple()) return false;
+ for (User *U : SI->users()) {
+ LoadInst *LI = dyn_cast<LoadInst>(U);
+ if (!LI || !LI->isSimple()) return false;
// Both operands to the select need to be dereferencable, either absolutely
// (e.g. allocas) or at this point because we can see other accesses to it.
- if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
- LI->getAlignment(), DL))
+ if (!TDerefable &&
+ !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
+ LI->getAlignment()))
return false;
- if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
- LI->getAlignment(), DL))
+ if (!FDerefable &&
+ !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
+ LI->getAlignment()))
return false;
}
///
/// We can do this to a select if its only uses are loads and if the operand to
/// the select can be loaded unconditionally.
-static bool isSafePHIToSpeculate(PHINode *PN, const DataLayout *DL) {
+static bool isSafePHIToSpeculate(PHINode *PN) {
// For now, we can only do this promotion if the load is in the same block as
// the PHI, and if there are no stores between the phi and load.
// TODO: Allow recursive phi users.
// TODO: Allow stores.
BasicBlock *BB = PN->getParent();
unsigned MaxAlign = 0;
- for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
- UI != UE; ++UI) {
- LoadInst *LI = dyn_cast<LoadInst>(*UI);
- if (LI == 0 || !LI->isSimple()) return false;
+ for (User *U : PN->users()) {
+ LoadInst *LI = dyn_cast<LoadInst>(U);
+ if (!LI || !LI->isSimple()) return false;
// For now we only allow loads in the same block as the PHI. This is a
// common case that happens when instcombine merges two loads through a PHI.
MaxAlign = std::max(MaxAlign, LI->getAlignment());
}
+ const DataLayout &DL = PN->getModule()->getDataLayout();
+
// Okay, we know that we have one or more loads in the same block as the PHI.
// We can transform this if it is safe to push the loads into the predecessor
// blocks. The only thing to watch out for is that we can't put a possibly
// If this pointer is always safe to load, or if we can prove that there is
// already a load in the block, then we can move the load to the pred block.
- if (InVal->isDereferenceablePointer() ||
- isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, DL))
+ if (InVal->isDereferenceablePointer(DL) ||
+ isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign))
continue;
return false;
/// direct (non-volatile) loads and stores to it. If the alloca is close but
/// not quite there, this will transform the code to allow promotion. As such,
/// it is a non-pure predicate.
-static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout *DL) {
+static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout &DL) {
SetVector<Instruction*, SmallVector<Instruction*, 4>,
SmallPtrSet<Instruction*, 4> > InstsToRewrite;
-
- for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
- UI != UE; ++UI) {
- User *U = *UI;
+ for (User *U : AI->users()) {
if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
if (!LI->isSimple())
return false;
// If it is safe to turn "load (select c, AI, ptr)" into a select of two
// loads, then we can transform this by rewriting the select.
- if (!isSafeSelectToSpeculate(SI, DL))
+ if (!isSafeSelectToSpeculate(SI))
return false;
InstsToRewrite.insert(SI);
// If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
// in the pred blocks, then we can transform this by rewriting the PHI.
- if (!isSafePHIToSpeculate(PN, DL))
+ if (!isSafePHIToSpeculate(PN))
return false;
InstsToRewrite.insert(PN);
for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) {
// This could only be a bitcast used by nothing but lifetime intrinsics.
- for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end();
- I != E;) {
- Use &U = I.getUse();
- ++I;
- cast<Instruction>(U.getUser())->eraseFromParent();
- }
+ for (BitCastInst::user_iterator I = BCI->user_begin(), E = BCI->user_end();
+ I != E;)
+ cast<Instruction>(*I++)->eraseFromParent();
BCI->eraseFromParent();
continue;
}
// Selects in InstsToRewrite only have load uses. Rewrite each as two
// loads with a new select.
while (!SI->use_empty()) {
- LoadInst *LI = cast<LoadInst>(SI->use_back());
+ LoadInst *LI = cast<LoadInst>(SI->user_back());
IRBuilder<> Builder(LI);
LoadInst *TrueLoad =
LoadInst *FalseLoad =
Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
- // Transfer alignment and TBAA info if present.
+ // Transfer alignment and AA info if present.
TrueLoad->setAlignment(LI->getAlignment());
FalseLoad->setAlignment(LI->getAlignment());
- if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
- TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
- FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
+
+ AAMDNodes Tags;
+ LI->getAAMetadata(Tags);
+ if (Tags) {
+ TrueLoad->setAAMetadata(Tags);
+ FalseLoad->setAAMetadata(Tags);
}
Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
PN->getName()+".ld", PN);
- // Get the TBAA tag and alignment to use from one of the loads. It doesn't
+ // Get the AA tags and alignment to use from one of the loads. It doesn't
// matter which one we get and if any differ, it doesn't matter.
- LoadInst *SomeLoad = cast<LoadInst>(PN->use_back());
- MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
+ LoadInst *SomeLoad = cast<LoadInst>(PN->user_back());
+
+ AAMDNodes AATags;
+ SomeLoad->getAAMetadata(AATags);
unsigned Align = SomeLoad->getAlignment();
// Rewrite all loads of the PN to use the new PHI.
while (!PN->use_empty()) {
- LoadInst *LI = cast<LoadInst>(PN->use_back());
+ LoadInst *LI = cast<LoadInst>(PN->user_back());
LI->replaceAllUsesWith(NewPN);
LI->eraseFromParent();
}
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
BasicBlock *Pred = PN->getIncomingBlock(i);
LoadInst *&Load = InsertedLoads[Pred];
- if (Load == 0) {
+ if (!Load) {
Load = new LoadInst(PN->getIncomingValue(i),
PN->getName() + "." + Pred->getName(),
Pred->getTerminator());
Load->setAlignment(Align);
- if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
+ if (AATags) Load->setAAMetadata(AATags);
}
NewPN->addIncoming(Load, Pred);
bool SROA::performPromotion(Function &F) {
std::vector<AllocaInst*> Allocas;
- DominatorTree *DT = 0;
+ const DataLayout &DL = F.getParent()->getDataLayout();
+ DominatorTree *DT = nullptr;
if (HasDomTree)
DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+ AssumptionCache &AC =
+ getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
- DIBuilder DIB(*F.getParent());
+ DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
bool Changed = false;
SmallVector<Instruction*, 64> Insts;
while (1) {
if (Allocas.empty()) break;
if (HasDomTree)
- PromoteMemToReg(Allocas, *DT);
+ PromoteMemToReg(Allocas, *DT, nullptr, &AC);
else {
SSAUpdater SSA;
for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
AllocaInst *AI = Allocas[i];
// Build list of instructions to promote.
- for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
- UI != E; ++UI)
- Insts.push_back(cast<Instruction>(*UI));
+ for (User *U : AI->users())
+ Insts.push_back(cast<Instruction>(U));
AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts);
Insts.clear();
}
//
bool SROA::performScalarRepl(Function &F) {
std::vector<AllocaInst*> WorkList;
+ const DataLayout &DL = F.getParent()->getDataLayout();
// Scan the entry basic block, adding allocas to the worklist.
BasicBlock &BB = F.getEntryBlock();
// transform the allocation instruction if it is an array allocation
// (allocations OF arrays are ok though), and an allocation of a scalar
// value cannot be decomposed at all.
- uint64_t AllocaSize = DL->getTypeAllocSize(AI->getAllocatedType());
+ uint64_t AllocaSize = DL.getTypeAllocSize(AI->getAllocatedType());
// Do not promote [0 x %struct].
if (AllocaSize == 0) continue;
// promoted itself. If so, we don't want to transform it needlessly. Note
// that we can't just check based on the type: the alloca may be of an i32
// but that has pointer arithmetic to set byte 3 of it or something.
- if (AllocaInst *NewAI = ConvertToScalarInfo(
- (unsigned)AllocaSize, *DL, ScalarLoadThreshold).TryConvert(AI)) {
+ if (AllocaInst *NewAI =
+ ConvertToScalarInfo((unsigned)AllocaSize, DL, ScalarLoadThreshold)
+ .TryConvert(AI)) {
NewAI->takeName(AI);
AI->eraseFromParent();
++NumConverted;
if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
ElementAllocas.reserve(ST->getNumContainedTypes());
for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
- AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
+ AllocaInst *NA = new AllocaInst(ST->getContainedType(i), nullptr,
AI->getAlignment(),
AI->getName() + "." + Twine(i), AI);
ElementAllocas.push_back(NA);
ElementAllocas.reserve(AT->getNumElements());
Type *ElTy = AT->getElementType();
for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
- AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
+ AllocaInst *NA = new AllocaInst(ElTy, nullptr, AI->getAlignment(),
AI->getName() + "." + Twine(i), AI);
ElementAllocas.push_back(NA);
WorkList.push_back(NA); // Add to worklist for recursive processing
// Zero out the operand and see if it becomes trivially dead.
// (But, don't add allocas to the dead instruction list -- they are
// already on the worklist and will be deleted separately.)
- *OI = 0;
+ *OI = nullptr;
if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
DeadInsts.push_back(U);
}
/// referenced by this instruction.
void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
AllocaInfo &Info) {
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
- Instruction *User = cast<Instruction>(*UI);
+ const DataLayout &DL = I->getModule()->getDataLayout();
+ for (Use &U : I->uses()) {
+ Instruction *User = cast<Instruction>(U.getUser());
if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
isSafeForScalarRepl(BC, Offset, Info);
isSafeForScalarRepl(GEPI, GEPOffset, Info);
} else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
- if (Length == 0)
- return MarkUnsafe(Info, User);
- if (Length->isNegative())
+ if (!Length || Length->isNegative())
return MarkUnsafe(Info, User);
- isSafeMemAccess(Offset, Length->getZExtValue(), 0,
- UI.getOperandNo() == 0, Info, MI,
+ isSafeMemAccess(Offset, Length->getZExtValue(), nullptr,
+ U.getOperandNo() == 0, Info, MI,
true /*AllowWholeAccess*/);
} else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
if (!LI->isSimple())
return MarkUnsafe(Info, User);
Type *LIType = LI->getType();
- isSafeMemAccess(Offset, DL->getTypeAllocSize(LIType),
- LIType, false, Info, LI, true /*AllowWholeAccess*/);
+ isSafeMemAccess(Offset, DL.getTypeAllocSize(LIType), LIType, false, Info,
+ LI, true /*AllowWholeAccess*/);
Info.hasALoadOrStore = true;
} else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
return MarkUnsafe(Info, User);
Type *SIType = SI->getOperand(0)->getType();
- isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType),
- SIType, true, Info, SI, true /*AllowWholeAccess*/);
+ isSafeMemAccess(Offset, DL.getTypeAllocSize(SIType), SIType, true, Info,
+ SI, true /*AllowWholeAccess*/);
Info.hasALoadOrStore = true;
} else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
AllocaInfo &Info) {
// If we've already checked this PHI, don't do it again.
if (PHINode *PN = dyn_cast<PHINode>(I))
- if (!Info.CheckedPHIs.insert(PN))
+ if (!Info.CheckedPHIs.insert(PN).second)
return;
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
- Instruction *User = cast<Instruction>(*UI);
+ const DataLayout &DL = I->getModule()->getDataLayout();
+ for (User *U : I->users()) {
+ Instruction *UI = cast<Instruction>(U);
- if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
+ if (BitCastInst *BC = dyn_cast<BitCastInst>(UI)) {
isSafePHISelectUseForScalarRepl(BC, Offset, Info);
- } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
+ } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
// Only allow "bitcast" GEPs for simplicity. We could generalize this,
// but would have to prove that we're staying inside of an element being
// promoted.
if (!GEPI->hasAllZeroIndices())
- return MarkUnsafe(Info, User);
+ return MarkUnsafe(Info, UI);
isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
- } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
+ } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
if (!LI->isSimple())
- return MarkUnsafe(Info, User);
+ return MarkUnsafe(Info, UI);
Type *LIType = LI->getType();
- isSafeMemAccess(Offset, DL->getTypeAllocSize(LIType),
- LIType, false, Info, LI, false /*AllowWholeAccess*/);
+ isSafeMemAccess(Offset, DL.getTypeAllocSize(LIType), LIType, false, Info,
+ LI, false /*AllowWholeAccess*/);
Info.hasALoadOrStore = true;
- } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
// Store is ok if storing INTO the pointer, not storing the pointer
if (!SI->isSimple() || SI->getOperand(0) == I)
- return MarkUnsafe(Info, User);
+ return MarkUnsafe(Info, UI);
Type *SIType = SI->getOperand(0)->getType();
- isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType),
- SIType, true, Info, SI, false /*AllowWholeAccess*/);
+ isSafeMemAccess(Offset, DL.getTypeAllocSize(SIType), SIType, true, Info,
+ SI, false /*AllowWholeAccess*/);
Info.hasALoadOrStore = true;
- } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
- isSafePHISelectUseForScalarRepl(User, Offset, Info);
+ } else if (isa<PHINode>(UI) || isa<SelectInst>(UI)) {
+ isSafePHISelectUseForScalarRepl(UI, Offset, Info);
} else {
- return MarkUnsafe(Info, User);
+ return MarkUnsafe(Info, UI);
}
if (Info.isUnsafe) return;
}
// constant part of the offset.
if (NonConstant)
Indices.pop_back();
- Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
- if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset,
- NonConstantIdxSize))
+
+ const DataLayout &DL = GEPI->getModule()->getDataLayout();
+ Offset += DL.getIndexedOffset(GEPI->getPointerOperandType(), Indices);
+ if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, NonConstantIdxSize,
+ DL))
MarkUnsafe(Info, GEPI);
}
Type *&EltTy) {
if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
NumElts = AT->getNumElements();
- EltTy = (NumElts == 0 ? 0 : AT->getElementType());
+ EltTy = (NumElts == 0 ? nullptr : AT->getElementType());
return true;
}
if (StructType *ST = dyn_cast<StructType>(T)) {
NumElts = ST->getNumContainedTypes();
- EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
+ EltTy = (NumElts == 0 ? nullptr : ST->getContainedType(0));
for (unsigned n = 1; n < NumElts; ++n) {
if (ST->getContainedType(n) != EltTy)
return false;
Type *MemOpType, bool isStore,
AllocaInfo &Info, Instruction *TheAccess,
bool AllowWholeAccess) {
+ const DataLayout &DL = TheAccess->getModule()->getDataLayout();
// Check if this is a load/store of the entire alloca.
if (Offset == 0 && AllowWholeAccess &&
- MemSize == DL->getTypeAllocSize(Info.AI->getAllocatedType())) {
+ MemSize == DL.getTypeAllocSize(Info.AI->getAllocatedType())) {
// This can be safe for MemIntrinsics (where MemOpType is 0) and integer
// loads/stores (which are essentially the same as the MemIntrinsics with
// regard to copying padding between elements). But, if an alloca is
}
// Check if the offset/size correspond to a component within the alloca type.
Type *T = Info.AI->getAllocatedType();
- if (TypeHasComponent(T, Offset, MemSize)) {
+ if (TypeHasComponent(T, Offset, MemSize, DL)) {
Info.hasSubelementAccess = true;
return;
}
/// TypeHasComponent - Return true if T has a component type with the
/// specified offset and size. If Size is zero, do not check the size.
-bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) {
+bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size,
+ const DataLayout &DL) {
Type *EltTy;
uint64_t EltSize;
if (StructType *ST = dyn_cast<StructType>(T)) {
- const StructLayout *Layout = DL->getStructLayout(ST);
+ const StructLayout *Layout = DL.getStructLayout(ST);
unsigned EltIdx = Layout->getElementContainingOffset(Offset);
EltTy = ST->getContainedType(EltIdx);
- EltSize = DL->getTypeAllocSize(EltTy);
+ EltSize = DL.getTypeAllocSize(EltTy);
Offset -= Layout->getElementOffset(EltIdx);
} else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
EltTy = AT->getElementType();
- EltSize = DL->getTypeAllocSize(EltTy);
+ EltSize = DL.getTypeAllocSize(EltTy);
if (Offset >= AT->getNumElements() * EltSize)
return false;
Offset %= EltSize;
} else if (VectorType *VT = dyn_cast<VectorType>(T)) {
EltTy = VT->getElementType();
- EltSize = DL->getTypeAllocSize(EltTy);
+ EltSize = DL.getTypeAllocSize(EltTy);
if (Offset >= VT->getNumElements() * EltSize)
return false;
Offset %= EltSize;
// Check if the component spans multiple elements.
if (Offset + Size > EltSize)
return false;
- return TypeHasComponent(EltTy, Offset, Size);
+ return TypeHasComponent(EltTy, Offset, Size, DL);
}
/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
/// instruction.
void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
SmallVectorImpl<AllocaInst *> &NewElts) {
+ const DataLayout &DL = I->getModule()->getDataLayout();
for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
- Use &TheUse = UI.getUse();
- Instruction *User = cast<Instruction>(*UI++);
+ Use &TheUse = *UI++;
+ Instruction *User = cast<Instruction>(TheUse.getUser());
if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
RewriteBitCast(BC, AI, Offset, NewElts);
if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
uint64_t MemSize = Length->getZExtValue();
- if (Offset == 0 &&
- MemSize == DL->getTypeAllocSize(AI->getAllocatedType()))
+ if (Offset == 0 && MemSize == DL.getTypeAllocSize(AI->getAllocatedType()))
RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
// Otherwise the intrinsic can only touch a single element and the
// address operand will be updated, so nothing else needs to be done.
LI->replaceAllUsesWith(Insert);
DeadInsts.push_back(LI);
} else if (LIType->isIntegerTy() &&
- DL->getTypeAllocSize(LIType) ==
- DL->getTypeAllocSize(AI->getAllocatedType())) {
+ DL.getTypeAllocSize(LIType) ==
+ DL.getTypeAllocSize(AI->getAllocatedType())) {
// If this is a load of the entire alloca to an integer, rewrite it.
RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
}
}
DeadInsts.push_back(SI);
} else if (SIType->isIntegerTy() &&
- DL->getTypeAllocSize(SIType) ==
- DL->getTypeAllocSize(AI->getAllocatedType())) {
+ DL.getTypeAllocSize(SIType) ==
+ DL.getTypeAllocSize(AI->getAllocatedType())) {
// If this is a store of the entire alloca from an integer, rewrite it.
RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
}
Type *T = AI->getAllocatedType();
uint64_t EltOffset = 0;
Type *IdxTy;
- uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
+ uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy,
+ BC->getModule()->getDataLayout());
Instruction *Val = NewElts[Idx];
if (Val->getType() != BC->getDestTy()) {
Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
/// Sets T to the type of the element and Offset to the offset within that
/// element. IdxTy is set to the type of the index result to be used in a
/// GEP instruction.
-uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset,
- Type *&IdxTy) {
+uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, Type *&IdxTy,
+ const DataLayout &DL) {
uint64_t Idx = 0;
+
if (StructType *ST = dyn_cast<StructType>(T)) {
- const StructLayout *Layout = DL->getStructLayout(ST);
+ const StructLayout *Layout = DL.getStructLayout(ST);
Idx = Layout->getElementContainingOffset(Offset);
T = ST->getContainedType(Idx);
Offset -= Layout->getElementOffset(Idx);
return Idx;
} else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
T = AT->getElementType();
- uint64_t EltSize = DL->getTypeAllocSize(T);
+ uint64_t EltSize = DL.getTypeAllocSize(T);
Idx = Offset / EltSize;
Offset -= Idx * EltSize;
IdxTy = Type::getInt64Ty(T->getContext());
}
VectorType *VT = cast<VectorType>(T);
T = VT->getElementType();
- uint64_t EltSize = DL->getTypeAllocSize(T);
+ uint64_t EltSize = DL.getTypeAllocSize(T);
Idx = Offset / EltSize;
Offset -= Idx * EltSize;
IdxTy = Type::getInt64Ty(T->getContext());
void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
SmallVectorImpl<AllocaInst *> &NewElts) {
uint64_t OldOffset = Offset;
+ const DataLayout &DL = GEPI->getModule()->getDataLayout();
SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
// If the GEP was dynamic then it must have been a dynamic vector lookup.
// In this case, it must be the last GEP operand which is dynamic so keep that
// aside until we've found the constant GEP offset then add it back in at the
// end.
- Value* NonConstantIdx = 0;
+ Value* NonConstantIdx = nullptr;
if (!GEPI->hasAllConstantIndices())
NonConstantIdx = Indices.pop_back_val();
- Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
+ Offset += DL.getIndexedOffset(GEPI->getPointerOperandType(), Indices);
RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
Type *T = AI->getAllocatedType();
Type *IdxTy;
- uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
+ uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy, DL);
if (GEPI->getOperand(0) == AI)
OldIdx = ~0ULL; // Force the GEP to be rewritten.
T = AI->getAllocatedType();
uint64_t EltOffset = Offset;
- uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
+ uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy, DL);
// If this GEP does not move the pointer across elements of the alloca
// being split, then it does not needs to be rewritten.
SmallVector<Value*, 8> NewArgs;
NewArgs.push_back(Constant::getNullValue(i32Ty));
while (EltOffset != 0) {
- uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
+ uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy, DL);
NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
}
if (NonConstantIdx) {
// Put matching lifetime markers on everything from Offset up to
// Offset+OldSize.
Type *AIType = AI->getAllocatedType();
+ const DataLayout &DL = II->getModule()->getDataLayout();
uint64_t NewOffset = Offset;
Type *IdxTy;
- uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy);
+ uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy, DL);
IRBuilder<> Builder(II);
uint64_t Size = OldSize->getLimitedValue();
if (NewOffset) {
// Splice the first element and index 'NewOffset' bytes in. SROA will
// split the alloca again later.
- Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy());
- V = Builder.CreateGEP(V, Builder.getInt64(NewOffset));
+ unsigned AS = AI->getType()->getAddressSpace();
+ Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy(AS));
+ V = Builder.CreateGEP(Builder.getInt8Ty(), V, Builder.getInt64(NewOffset));
IdxTy = NewElts[Idx]->getAllocatedType();
- uint64_t EltSize = DL->getTypeAllocSize(IdxTy) - NewOffset;
+ uint64_t EltSize = DL.getTypeAllocSize(IdxTy) - NewOffset;
if (EltSize > Size) {
EltSize = Size;
Size = 0;
for (; Idx != NewElts.size() && Size; ++Idx) {
IdxTy = NewElts[Idx]->getAllocatedType();
- uint64_t EltSize = DL->getTypeAllocSize(IdxTy);
+ uint64_t EltSize = DL.getTypeAllocSize(IdxTy);
if (EltSize > Size) {
EltSize = Size;
Size = 0;
// appropriate type. The "Other" pointer is the pointer that goes to memory
// that doesn't have anything to do with the alloca that we are promoting. For
// memset, this Value* stays null.
- Value *OtherPtr = 0;
+ Value *OtherPtr = nullptr;
unsigned MemAlignment = MI->getAlignment();
if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
if (Inst == MTI->getRawDest())
bool SROADest = MI->getRawDest() == Inst;
Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
+ const DataLayout &DL = MI->getModule()->getDataLayout();
for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
// If this is a memcpy/memmove, emit a GEP of the other element address.
- Value *OtherElt = 0;
+ Value *OtherElt = nullptr;
unsigned OtherEltAlign = MemAlignment;
if (OtherPtr) {
PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
Type *OtherTy = OtherPtrTy->getElementType();
if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
- EltOffset = DL->getStructLayout(ST)->getElementOffset(i);
+ EltOffset = DL.getStructLayout(ST)->getElementOffset(i);
} else {
Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
- EltOffset = DL->getTypeAllocSize(EltTy)*i;
+ EltOffset = DL.getTypeAllocSize(EltTy) * i;
}
// The alignment of the other pointer is the guaranteed alignment of the
Type *ValTy = EltTy->getScalarType();
// Construct an integer with the right value.
- unsigned EltSize = DL->getTypeSizeInBits(ValTy);
+ unsigned EltSize = DL.getTypeSizeInBits(ValTy);
APInt OneVal(EltSize, CI->getZExtValue());
APInt TotalVal(OneVal);
// Set each byte.
// this element.
}
- unsigned EltSize = DL->getTypeAllocSize(EltTy);
+ unsigned EltSize = DL.getTypeAllocSize(EltTy);
if (!EltSize)
continue;
// and store the element value to the individual alloca.
Value *SrcVal = SI->getOperand(0);
Type *AllocaEltTy = AI->getAllocatedType();
- uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy);
+ const DataLayout &DL = SI->getModule()->getDataLayout();
+ uint64_t AllocaSizeBits = DL.getTypeAllocSizeInBits(AllocaEltTy);
IRBuilder<> Builder(SI);
// Handle tail padding by extending the operand
- if (DL->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
+ if (DL.getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
SrcVal = Builder.CreateZExt(SrcVal,
IntegerType::get(SI->getContext(), AllocaSizeBits));
// There are two forms here: AI could be an array or struct. Both cases
// have different ways to compute the element offset.
if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
- const StructLayout *Layout = DL->getStructLayout(EltSTy);
+ const StructLayout *Layout = DL.getStructLayout(EltSTy);
for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
// Get the number of bits to shift SrcVal to get the value.
Type *FieldTy = EltSTy->getElementType(i);
uint64_t Shift = Layout->getElementOffsetInBits(i);
- if (DL->isBigEndian())
- Shift = AllocaSizeBits-Shift-DL->getTypeAllocSizeInBits(FieldTy);
+ if (DL.isBigEndian())
+ Shift = AllocaSizeBits - Shift - DL.getTypeAllocSizeInBits(FieldTy);
Value *EltVal = SrcVal;
if (Shift) {
}
// Truncate down to an integer of the right size.
- uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy);
+ uint64_t FieldSizeBits = DL.getTypeSizeInBits(FieldTy);
// Ignore zero sized fields like {}, they obviously contain no data.
if (FieldSizeBits == 0) continue;
} else {
ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
Type *ArrayEltTy = ATy->getElementType();
- uint64_t ElementOffset = DL->getTypeAllocSizeInBits(ArrayEltTy);
- uint64_t ElementSizeBits = DL->getTypeSizeInBits(ArrayEltTy);
+ uint64_t ElementOffset = DL.getTypeAllocSizeInBits(ArrayEltTy);
+ uint64_t ElementSizeBits = DL.getTypeSizeInBits(ArrayEltTy);
uint64_t Shift;
- if (DL->isBigEndian())
+ if (DL.isBigEndian())
Shift = AllocaSizeBits-ElementOffset;
else
Shift = 0;
}
new StoreInst(EltVal, DestField, SI);
- if (DL->isBigEndian())
+ if (DL.isBigEndian())
Shift -= ElementOffset;
else
Shift += ElementOffset;
// Extract each element out of the NewElts according to its structure offset
// and form the result value.
Type *AllocaEltTy = AI->getAllocatedType();
- uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy);
+ const DataLayout &DL = LI->getModule()->getDataLayout();
+ uint64_t AllocaSizeBits = DL.getTypeAllocSizeInBits(AllocaEltTy);
DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
<< '\n');
// There are two forms here: AI could be an array or struct. Both cases
// have different ways to compute the element offset.
- const StructLayout *Layout = 0;
+ const StructLayout *Layout = nullptr;
uint64_t ArrayEltBitOffset = 0;
if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
- Layout = DL->getStructLayout(EltSTy);
+ Layout = DL.getStructLayout(EltSTy);
} else {
Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
- ArrayEltBitOffset = DL->getTypeAllocSizeInBits(ArrayEltTy);
+ ArrayEltBitOffset = DL.getTypeAllocSizeInBits(ArrayEltTy);
}
Value *ResultVal =
Value *SrcField = NewElts[i];
Type *FieldTy =
cast<PointerType>(SrcField->getType())->getElementType();
- uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy);
+ uint64_t FieldSizeBits = DL.getTypeSizeInBits(FieldTy);
// Ignore zero sized fields like {}, they obviously contain no data.
if (FieldSizeBits == 0) continue;
else // Array case.
Shift = i*ArrayEltBitOffset;
- if (DL->isBigEndian())
+ if (DL.isBigEndian())
Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
if (Shift) {
}
// Handle tail padding by truncating the result
- if (DL->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
+ if (DL.getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
LI->replaceAllUsesWith(ResultVal);
return false;
}
+ const DataLayout &DL = AI->getModule()->getDataLayout();
+
// Okay, we know all the users are promotable. If the aggregate is a memcpy
// source and destination, we have to be careful. In particular, the memcpy
// could be moving around elements that live in structure padding of the LLVM
// types, but may actually be used. In these cases, we refuse to promote the
// struct.
if (Info.isMemCpySrc && Info.isMemCpyDst &&
- HasPadding(AI->getAllocatedType(), *DL))
+ HasPadding(AI->getAllocatedType(), DL))
return false;
// If the alloca never has an access to just *part* of it, but is accessed