index d88554189d17753e41b94f25d138525e35106943..73861866956df2ce1c37a0f3ea88f4a8717fd06b 100644 (file)
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/CallGraph.h"
#include "llvm/Analysis/CaptureTracking.h"
#include "llvm/Analysis/InstructionSimplify.h"
using namespace llvm;
static cl::opt<bool>
-EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(false),
+EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
cl::Hidden,
cl::desc("Convert noalias attributes to metadata during inlining."));
+static cl::opt<bool>
+PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
+ cl::init(true), cl::Hidden,
+ cl::desc("Convert align attributes to assumptions during inlining."));
+
bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
bool InsertLifetime) {
return InlineFunction(CallSite(CI), IFI, InsertLifetime);
// Walk the existing metadata, adding the complete (perhaps cyclic) chain to
// the set.
- SmallVector<const Value *, 16> Queue(MD.begin(), MD.end());
+ SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
while (!Queue.empty()) {
const MDNode *M = cast<MDNode>(Queue.pop_back_val());
for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
// Now we have a complete set of all metadata in the chains used to specify
// the noalias scopes and the lists of those scopes.
- SmallVector<MDNode *, 16> DummyNodes;
- DenseMap<const MDNode *, TrackingVH<MDNode> > MDMap;
+ SmallVector<TempMDTuple, 16> DummyNodes;
+ DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
I != IE; ++I) {
- MDNode *Dummy = MDNode::getTemporary(CalledFunc->getContext(), None);
- DummyNodes.push_back(Dummy);
- MDMap[*I] = Dummy;
+ DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
+ MDMap[*I].reset(DummyNodes.back().get());
}
// Create new metadata nodes to replace the dummy nodes, replacing old
// node.
for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
I != IE; ++I) {
- SmallVector<Value *, 4> NewOps;
+ SmallVector<Metadata *, 4> NewOps;
for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
- const Value *V = (*I)->getOperand(i);
+ const Metadata *V = (*I)->getOperand(i);
if (const MDNode *M = dyn_cast<MDNode>(V))
NewOps.push_back(MDMap[M]);
else
- NewOps.push_back(const_cast<Value *>(V));
+ NewOps.push_back(const_cast<Metadata *>(V));
}
- MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps),
- *TempM = MDMap[*I];
+ MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
+ MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
+ assert(TempM->isTemporary() && "Expected temporary node");
TempM->replaceAllUsesWith(NewM);
}
// which instructions inside it might belong), propagate those scopes to
// the inlined instructions.
if (MDNode *CSM =
- CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
+ CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
NewMD = MDNode::concatenate(NewMD, CSM);
NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
} else if (NI->mayReadOrWriteMemory()) {
if (MDNode *M =
- CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
+ CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
NI->setMetadata(LLVMContext::MD_alias_scope, M);
}
// which instructions inside it don't alias), propagate those scopes to
// the inlined instructions.
if (MDNode *CSM =
- CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
+ CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
NewMD = MDNode::concatenate(NewMD, CSM);
NI->setMetadata(LLVMContext::MD_noalias, NewMD);
} else if (NI->mayReadOrWriteMemory()) {
- if (MDNode *M =
- CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
+ if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
NI->setMetadata(LLVMContext::MD_noalias, M);
}
}
-
- // Now that everything has been replaced, delete the dummy nodes.
- for (unsigned i = 0, ie = DummyNodes.size(); i != ie; ++i)
- MDNode::deleteTemporary(DummyNodes[i]);
}
/// AddAliasScopeMetadata - If the inlined function has noalias arguments, then
/// parameters with noalias metadata specifying the new scope, and tag all
/// non-derived loads, stores and memory intrinsics with the new alias scopes.
static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
- const DataLayout *DL) {
+ const DataLayout *DL, AliasAnalysis *AA) {
if (!EnableNoAliasConversion)
return;
if (!NI)
continue;
+ bool IsArgMemOnlyCall = false, IsFuncCall = false;
SmallVector<const Value *, 2> PtrArgs;
if (const LoadInst *LI = dyn_cast<LoadInst>(I))
if (ICS.doesNotAccessMemory())
continue;
+ IsFuncCall = true;
+ if (AA) {
+ AliasAnalysis::ModRefBehavior MRB = AA->getModRefBehavior(ICS);
+ if (MRB == AliasAnalysis::OnlyAccessesArgumentPointees ||
+ MRB == AliasAnalysis::OnlyReadsArgumentPointees)
+ IsArgMemOnlyCall = true;
+ }
+
for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
- AE = ICS.arg_end(); AI != AE; ++AI)
+ AE = ICS.arg_end(); AI != AE; ++AI) {
// We need to check the underlying objects of all arguments, not just
// the pointer arguments, because we might be passing pointers as
// integers, etc.
- // FIXME: If we know that the call only accesses pointer arguments,
+ // However, if we know that the call only accesses pointer arguments,
// then we only need to check the pointer arguments.
+ if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
+ continue;
+
PtrArgs.push_back(*AI);
+ }
}
- bool IsFuncCall = isa<CallInst>(I) || isa<InvokeInst>(I);
- // FIXME: We should have a way to access the
- // IntrReadArgMem/IntrReadWriteArgMem properties of intrinsics, and we
- // should have a way to determine that for regular functions too. For
- // now, just do this for the memory intrinsics we understand.
- bool IsArgMemOnlyCall = isa<MemIntrinsic>(I);
-
// If we found no pointers, then this instruction is not suitable for
// pairing with an instruction to receive aliasing metadata.
// However, if this is a call, this we might just alias with none of the
// need to go through several PHIs to see it, and thus could be
// repeated in the Objects list.
SmallPtrSet<const Value *, 4> ObjSet;
- SmallVector<Value *, 4> Scopes, NoAliases;
+ SmallVector<Metadata *, 4> Scopes, NoAliases;
SmallSetVector<const Argument *, 4> NAPtrArgs;
for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
}
if (!NoAliases.empty())
- NI->setMetadata(LLVMContext::MD_noalias, MDNode::concatenate(
- NI->getMetadata(LLVMContext::MD_noalias),
- MDNode::get(CalledFunc->getContext(), NoAliases)));
+ NI->setMetadata(LLVMContext::MD_noalias,
+ MDNode::concatenate(
+ NI->getMetadata(LLVMContext::MD_noalias),
+ MDNode::get(CalledFunc->getContext(), NoAliases)));
// Next, we want to figure out all of the sets to which we might belong.
// We might belong to a set if the noalias argument is in the set of
}
if (!Scopes.empty())
- NI->setMetadata(LLVMContext::MD_alias_scope, MDNode::concatenate(
- NI->getMetadata(LLVMContext::MD_alias_scope),
- MDNode::get(CalledFunc->getContext(), Scopes)));
+ NI->setMetadata(
+ LLVMContext::MD_alias_scope,
+ MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
+ MDNode::get(CalledFunc->getContext(), Scopes)));
+ }
+ }
+}
+
+/// If the inlined function has non-byval align arguments, then
+/// add @llvm.assume-based alignment assumptions to preserve this information.
+static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
+ if (!PreserveAlignmentAssumptions || !IFI.DL)
+ return;
+
+ // To avoid inserting redundant assumptions, we should check for assumptions
+ // already in the caller. To do this, we might need a DT of the caller.
+ DominatorTree DT;
+ bool DTCalculated = false;
+
+ Function *CalledFunc = CS.getCalledFunction();
+ for (Function::arg_iterator I = CalledFunc->arg_begin(),
+ E = CalledFunc->arg_end();
+ I != E; ++I) {
+ unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
+ if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
+ if (!DTCalculated) {
+ DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
+ ->getParent()));
+ DTCalculated = true;
+ }
+
+ // If we can already prove the asserted alignment in the context of the
+ // caller, then don't bother inserting the assumption.
+ Value *Arg = CS.getArgument(I->getArgNo());
+ if (getKnownAlignment(Arg, IFI.DL,
+ &IFI.ACT->getAssumptionCache(*CalledFunc),
+ CS.getInstruction(), &DT) >= Align)
+ continue;
+
+ IRBuilder<>(CS.getInstruction()).CreateAlignmentAssumption(*IFI.DL, Arg,
+ Align);
}
}
}
static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
BasicBlock *InsertBlock,
InlineFunctionInfo &IFI) {
- LLVMContext &Context = Src->getContext();
- Type *VoidPtrTy = Type::getInt8PtrTy(Context);
Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
- Type *Tys[3] = { VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context) };
- Function *MemCpyFn = Intrinsic::getDeclaration(M, Intrinsic::memcpy, Tys);
- IRBuilder<> builder(InsertBlock->begin());
- Value *DstCast = builder.CreateBitCast(Dst, VoidPtrTy, "tmp");
- Value *SrcCast = builder.CreateBitCast(Src, VoidPtrTy, "tmp");
+ IRBuilder<> Builder(InsertBlock->begin());
Value *Size;
if (IFI.DL == nullptr)
Size = ConstantExpr::getSizeOf(AggTy);
else
- Size = ConstantInt::get(Type::getInt64Ty(Context),
- IFI.DL->getTypeStoreSize(AggTy));
+ Size = Builder.getInt64(IFI.DL->getTypeStoreSize(AggTy));
// Always generate a memcpy of alignment 1 here because we don't know
// the alignment of the src pointer. Other optimizations can infer
// better alignment.
- Value *CallArgs[] = {
- DstCast, SrcCast, Size,
- ConstantInt::get(Type::getInt32Ty(Context), 1),
- ConstantInt::getFalse(Context) // isVolatile
- };
- builder.CreateCall(MemCpyFn, CallArgs);
+ Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
}
/// HandleByValArgument - When inlining a call site that has a byval argument,
PointerType *ArgTy = cast<PointerType>(Arg->getType());
Type *AggTy = ArgTy->getElementType();
+ Function *Caller = TheCall->getParent()->getParent();
+
// If the called function is readonly, then it could not mutate the caller's
// copy of the byval'd memory. In this case, it is safe to elide the copy and
// temporary.
// If the pointer is already known to be sufficiently aligned, or if we can
// round it up to a larger alignment, then we don't need a temporary.
- if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
- IFI.DL) >= ByValAlignment)
+ if (getOrEnforceKnownAlignment(Arg, ByValAlignment, IFI.DL,
+ &IFI.ACT->getAssumptionCache(*Caller),
+ TheCall) >= ByValAlignment)
return Arg;
// Otherwise, we have to make a memcpy to get a safe alignment. This is bad
// pointer inside the callee).
Align = std::max(Align, ByValAlignment);
- Function *Caller = TheCall->getParent()->getParent();
-
Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
&*Caller->begin()->begin());
IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
return false;
}
-/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to
-/// recursively update InlinedAtEntry of a DebugLoc.
-static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
- const DebugLoc &InlinedAtDL,
- LLVMContext &Ctx) {
- if (MDNode *IA = DL.getInlinedAt(Ctx)) {
- DebugLoc NewInlinedAtDL
- = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
- return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
- NewInlinedAtDL.getAsMDNode(Ctx));
+/// Rebuild the entire inlined-at chain for this instruction so that the top of
+/// the chain now is inlined-at the new call site.
+static DebugLoc
+updateInlinedAtInfo(DebugLoc DL, MDLocation *InlinedAtNode,
+ LLVMContext &Ctx,
+ DenseMap<const MDLocation *, MDLocation *> &IANodes) {
+ SmallVector<MDLocation*, 3> InlinedAtLocations;
+ MDLocation *Last = InlinedAtNode;
+ DebugLoc CurInlinedAt = DL;
+
+ // Gather all the inlined-at nodes
+ while (MDLocation *IA =
+ cast_or_null<MDLocation>(CurInlinedAt.getInlinedAt(Ctx))) {
+ // Skip any we've already built nodes for
+ if (MDLocation *Found = IANodes[IA]) {
+ Last = Found;
+ break;
+ }
+
+ InlinedAtLocations.push_back(IA);
+ CurInlinedAt = DebugLoc::getFromDILocation(IA);
+ }
+
+ // Starting from the top, rebuild the nodes to point to the new inlined-at
+ // location (then rebuilding the rest of the chain behind it) and update the
+ // map of already-constructed inlined-at nodes.
+ for (auto I = InlinedAtLocations.rbegin(), E = InlinedAtLocations.rend();
+ I != E; ++I) {
+ const MDLocation *MD = *I;
+ Last = IANodes[MD] = MDLocation::getDistinct(
+ Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
}
- return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
- InlinedAtDL.getAsMDNode(Ctx));
+ // And finally create the normal location for this instruction, referring to
+ // the new inlined-at chain.
+ return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), Last);
}
/// fixupLineNumbers - Update inlined instructions' line numbers to
if (TheCallDL.isUnknown())
return;
+ auto &Ctx = Fn->getContext();
+ auto *InlinedAtNode = cast<MDLocation>(TheCallDL.getAsMDNode(Ctx));
+
+ // Create a unique call site, not to be confused with any other call from the
+ // same location.
+ InlinedAtNode = MDLocation::getDistinct(
+ Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
+ InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
+
+ // Cache the inlined-at nodes as they're built so they are reused, without
+ // this every instruction's inlined-at chain would become distinct from each
+ // other.
+ DenseMap<const MDLocation *, MDLocation *> IANodes;
+
for (; FI != Fn->end(); ++FI) {
for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
BI != BE; ++BI) {
// originates from the call location. This is important for
// ((__always_inline__, __nodebug__)) functions which must use caller
// location for all instructions in their function body.
+
+ // Don't update static allocas, as they may get moved later.
+ if (auto *AI = dyn_cast<AllocaInst>(BI))
+ if (isa<Constant>(AI->getArraySize()))
+ continue;
+
BI->setDebugLoc(TheCallDL);
} else {
- BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
+ BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
LLVMContext &Ctx = BI->getContext();
MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
- DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
- InlinedAt, Ctx));
+ DVI->setOperand(2, MetadataAsValue::get(
+ Ctx, createInlinedVariable(DVI->getVariable(),
+ InlinedAt, Ctx)));
+ } else if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI)) {
+ LLVMContext &Ctx = BI->getContext();
+ MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
+ DDI->setOperand(1, MetadataAsValue::get(
+ Ctx, createInlinedVariable(DDI->getVariable(),
+ InlinedAt, Ctx)));
}
}
}
VMap[I] = ActualArg;
}
+ // Add alignment assumptions if necessary. We do this before the inlined
+ // instructions are actually cloned into the caller so that we can easily
+ // check what will be known at the start of the inlined code.
+ AddAlignmentAssumptions(CS, IFI);
+
// We want the inliner to prune the code as it copies. We would LOVE to
// have no dead or constant instructions leftover after inlining occurs
// (which can happen, e.g., because an argument was constant), but we'll be
CloneAliasScopeMetadata(CS, VMap);
// Add noalias metadata if necessary.
- AddAliasScopeMetadata(CS, VMap, IFI.DL);
+ AddAliasScopeMetadata(CS, VMap, IFI.DL, IFI.AA);
+
+ // FIXME: We could register any cloned assumptions instead of clearing the
+ // whole function's cache.
+ if (IFI.ACT)
+ IFI.ACT->getAssumptionCache(*Caller).clear();
}
// If there are any alloca instructions in the block that used to be the entry
// the entries are the same or undef). If so, remove the PHI so it doesn't
// block other optimizations.
if (PHI) {
- if (Value *V = SimplifyInstruction(PHI, IFI.DL)) {
+ if (Value *V = SimplifyInstruction(PHI, IFI.DL, nullptr, nullptr,
+ &IFI.ACT->getAssumptionCache(*Caller))) {
PHI->replaceAllUsesWith(V);
PHI->eraseFromParent();
}