Files
swift-mirror/lib/SILPasses/PerformanceInliner.cpp
2014-02-06 17:34:19 +00:00

404 lines
13 KiB
C++

//===- PerformanceInliner.cpp - Basic cost based inlining for performance -===//
//
// This source file is part of the Swift.org open source project
//
// Copyright (c) 2014 - 2015 Apple Inc. and the Swift project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See http://swift.org/LICENSE.txt for license information
// See http://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "sil-inliner"
#include "swift/SIL/SILInstruction.h"
#include "swift/SIL/CallGraph.h"
#include "swift/SIL/SILModule.h"
#include "swift/SILPasses/Passes.h"
#include "swift/SILPasses/PassManager.h"
#include "swift/SILPasses/Transforms.h"
#include "swift/SILPasses/Utils/Local.h"
#include "swift/SILPasses/Utils/SILInliner.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/ADT/MapVector.h"
using namespace swift;
STATISTIC(NumFunctionsInlined, "Number of functions inlined");
namespace {
class SILPerformanceInliner {
const unsigned InlineCostThreshold;
unsigned getFunctionCost(SILFunction *Callee,
SILFunction *Caller);
public:
explicit SILPerformanceInliner(unsigned threshold)
: InlineCostThreshold(threshold) {}
bool inlineCallsIntoFunction(SILFunction *F);
};
}
//===----------------------------------------------------------------------===//
// Call Graph Creation
//===----------------------------------------------------------------------===//
/// \brief Returns a SILFunction if this ApplyInst calls a recognizable function
/// that is legal to inline.
static SILFunction *getInlinableFunction(ApplyInst *AI) {
// Avoid substituion lists, we don't support them.
if (AI->hasSubstitutions())
return nullptr;
auto *FRI = dyn_cast<FunctionRefInst>(AI->getCallee().getDef());
if (!FRI)
return nullptr;
SILFunction *F = FRI->getReferencedFunction();
if (F->empty() || F->isExternalDeclaration()) {
DEBUG(llvm::dbgs() << " Can't inline " << F->getName() << ".\n");
return nullptr;
}
DEBUG(llvm::dbgs() << " Can inline " << F->getName() << ".\n");
return F;
}
//===----------------------------------------------------------------------===//
// Cost Model
//===----------------------------------------------------------------------===//
namespace {
// For now Free is 0 and Expensive is 1. This can be changed in the future by
// adding more categories.
enum class InlineCost : unsigned {
Free = 0,
Expensive = 1,
CannotBeInlined = UINT_MAX,
};
} // end anonymous namespace
static bool isValidLinkageForTransparentRef(SILLinkage linkage) {
switch (linkage) {
case SILLinkage::Public:
case SILLinkage::PublicExternal:
case SILLinkage::Shared:
return true;
case SILLinkage::Private:
case SILLinkage::Hidden:
case SILLinkage::HiddenExternal:
return false;
}
}
/// For now just assume that every SIL instruction is one to one with an LLVM
/// instruction. This is of course very much so not true.
///
/// TODO: Fill this out.
static InlineCost instructionInlineCost(SILInstruction &I,
SILFunction *Caller) {
switch (I.getKind()) {
case ValueKind::BuiltinFunctionRefInst:
case ValueKind::GlobalAddrInst:
case ValueKind::IntegerLiteralInst:
case ValueKind::FloatLiteralInst:
case ValueKind::DebugValueInst:
case ValueKind::DebugValueAddrInst:
case ValueKind::StringLiteralInst:
return InlineCost::Free;
// Private symbol references cannot be inlined into transparent functions.
case ValueKind::FunctionRefInst:
if (Caller->isTransparent()
&& !isValidLinkageForTransparentRef(
cast<FunctionRefInst>(I).getReferencedFunction()->getLinkage())) {
return InlineCost::CannotBeInlined;
}
return InlineCost::Free;
case ValueKind::SILGlobalAddrInst:
if (Caller->isTransparent()
&& !isValidLinkageForTransparentRef(
cast<SILGlobalAddrInst>(I).getReferencedGlobal()->getLinkage())) {
return InlineCost::CannotBeInlined;
}
return InlineCost::Free;
case ValueKind::TupleElementAddrInst:
case ValueKind::StructElementAddrInst: {
// A gep whose operand is a gep with no other users will get folded by
// LLVM into one gep implying the second should be free.
SILValue Op = I.getOperand(0);
if ((Op->getKind() == ValueKind::TupleElementAddrInst ||
Op->getKind() == ValueKind::StructElementAddrInst) &&
Op->hasOneUse())
return InlineCost::Free;
}
// Aggregates are exploded at the IR level; these are effectively no-ops.
case ValueKind::TupleInst:
case ValueKind::StructInst:
case ValueKind::StructExtractInst:
case ValueKind::TupleExtractInst:
return InlineCost::Free;
// Unchecked casts are free.
case ValueKind::AddressToPointerInst:
case ValueKind::PointerToAddressInst:
case ValueKind::ObjectPointerToRefInst:
case ValueKind::RefToObjectPointerInst:
case ValueKind::RawPointerToRefInst:
case ValueKind::RefToRawPointerInst:
case ValueKind::UpcastExistentialRefInst:
case ValueKind::UpcastInst:
case ValueKind::ThinToThickFunctionInst:
case ValueKind::ConvertFunctionInst:
return InlineCost::Free;
case ValueKind::MetatypeInst:
// Thin metatypes are always free.
if (I.getType(0).castTo<MetatypeType>()->isThin())
return InlineCost::Free;
// TODO: Thick metatypes are free if they don't require generic or lazy
// instantiation.
return InlineCost::Expensive;
// Return and unreachable are free.
case ValueKind::UnreachableInst:
case ValueKind::ReturnInst:
return InlineCost::Free;
// TODO
case ValueKind::ApplyInst: {
// Don't inline functions that contain recursions.
ApplyInst *AI = cast<ApplyInst>(&I);
auto *FRI = dyn_cast<FunctionRefInst>(AI->getCallee().getDef());
if (FRI && FRI->getReferencedFunction() == AI->getFunction())
return InlineCost::CannotBeInlined;
return InlineCost::Expensive;
}
case ValueKind::AllocArrayInst:
case ValueKind::AllocBoxInst:
case ValueKind::AllocRefInst:
case ValueKind::AllocStackInst:
case ValueKind::ArchetypeMetatypeInst:
case ValueKind::ArchetypeMethodInst:
case ValueKind::ArchetypeRefToSuperInst:
case ValueKind::AssignInst:
case ValueKind::AutoreleaseReturnInst:
case ValueKind::BranchInst:
case ValueKind::BridgeToBlockInst:
case ValueKind::CheckedCastBranchInst:
case ValueKind::ClassMetatypeInst:
case ValueKind::ClassMethodInst:
case ValueKind::CondBranchInst:
case ValueKind::CondFailInst:
case ValueKind::CopyAddrInst:
case ValueKind::CopyValueInst:
case ValueKind::DeallocBoxInst:
case ValueKind::DeallocRefInst:
case ValueKind::DeallocStackInst:
case ValueKind::DeinitExistentialInst:
case ValueKind::DestroyAddrInst:
case ValueKind::DestroyValueInst:
case ValueKind::DynamicMethodBranchInst:
case ValueKind::DynamicMethodInst:
case ValueKind::EnumInst:
case ValueKind::IndexAddrInst:
case ValueKind::IndexRawPointerInst:
case ValueKind::InitEnumDataAddrInst:
case ValueKind::InitExistentialInst:
case ValueKind::InitExistentialRefInst:
case ValueKind::InjectEnumAddrInst:
case ValueKind::IsNonnullInst:
case ValueKind::LoadInst:
case ValueKind::LoadWeakInst:
case ValueKind::PartialApplyInst:
case ValueKind::PeerMethodInst:
case ValueKind::ProjectExistentialInst:
case ValueKind::ProjectExistentialRefInst:
case ValueKind::ProtocolMetatypeInst:
case ValueKind::ProtocolMethodInst:
case ValueKind::RefElementAddrInst:
case ValueKind::RefToUnownedInst:
case ValueKind::StoreInst:
case ValueKind::StoreWeakInst:
case ValueKind::StrongReleaseInst:
case ValueKind::StrongRetainAutoreleasedInst:
case ValueKind::StrongRetainInst:
case ValueKind::StrongRetainUnownedInst:
case ValueKind::SuperMethodInst:
case ValueKind::SwitchEnumAddrInst:
case ValueKind::SwitchEnumInst:
case ValueKind::SwitchIntInst:
case ValueKind::TakeEnumDataAddrInst:
case ValueKind::UnconditionalCheckedCastInst:
case ValueKind::UnownedReleaseInst:
case ValueKind::UnownedRetainInst:
case ValueKind::UnownedToRefInst:
case ValueKind::UpcastExistentialInst:
return InlineCost::Expensive;
case ValueKind::SILArgument:
case ValueKind::SILUndef:
llvm_unreachable("Only instructions should be passed into this "
"function.");
case ValueKind::MarkFunctionEscapeInst:
case ValueKind::MarkUninitializedInst:
llvm_unreachable("not valid in canonical sil");
}
}
/// \brief Returns the inlining cost of the function.
unsigned SILPerformanceInliner::getFunctionCost(SILFunction *F,
SILFunction *Caller) {
DEBUG(llvm::dbgs() << " Calculating cost for " << F->getName() << ".\n");
if (F->isTransparent() == IsTransparent_t::IsTransparent)
return 0;
unsigned Cost = 0;
for (auto &BB : *F) {
for (auto &I : BB) {
auto ICost = instructionInlineCost(I, Caller);
if (ICost == InlineCost::CannotBeInlined)
return UINT_MAX;
Cost += unsigned(ICost);
// If we're debugging, continue calculating the total cost even if we
// passed the threshold.
DEBUG(continue);
// If i is greater than the InlineCostThreshold, we already know we are
// not going to inline this given function, so there is no point in
// continuing to visit instructions.
if (Cost > InlineCostThreshold)
return Cost;
}
}
DEBUG(llvm::dbgs() << " Found cost: " << Cost << "\n");
return Cost;
}
//===----------------------------------------------------------------------===//
// Inliner
//===----------------------------------------------------------------------===//
/// \brief Attempt to inline all calls smaller than our threshold into F until.
/// returns True if a function was inlined.
bool SILPerformanceInliner::inlineCallsIntoFunction(SILFunction *Caller) {
bool Changed = false;
SILInliner Inliner(*Caller, SILInliner::InlineKind::PerformanceInline);
DEBUG(llvm::dbgs() << "Visiting Function: " << Caller->getName() << "\n");
llvm::SmallVector<ApplyInst*, 8> CallSites;
// Collect all of the ApplyInsts in this function. We will be changing the
// control flow and collecting the AIs simplifies the scan.
for (auto &BB : *Caller) {
auto I = BB.begin(), E = BB.end();
while (I != E) {
// Check if this is a call site.
ApplyInst *AI = dyn_cast<ApplyInst>(I++);
if (AI)
CallSites.push_back(AI);
}
}
for (auto AI : CallSites) {
DEBUG(llvm::dbgs() << " Found call site:" << *AI);
// Get the callee.
SILFunction *Callee = getInlinableFunction(AI);
if (!Callee)
continue;
DEBUG(llvm::dbgs() << " Found callee:" << Callee->getName() << ".\n");
// Prevent circular inlining.
if (Callee == Caller) {
DEBUG(llvm::dbgs() << " Skipping recursive calls.\n");
continue;
}
// Calculate the inlining cost of the callee.
unsigned CalleeCost = getFunctionCost(Callee, Caller);
if (CalleeCost > InlineCostThreshold) {
DEBUG(llvm::dbgs() << " Function too big to inline. Skipping.\n");
continue;
}
// Add the arguments from AI into a SILValue list.
SmallVector<SILValue, 8> Args;
for (const auto &Arg : AI->getArguments())
Args.push_back(Arg);
// Ok, we are within budget. Attempt to inline.
DEBUG(llvm::dbgs() << " Inlining " << Callee->getName() << " Into " <<
Caller->getName() << "\n");
// We already moved the iterator to the next instruction because the AI
// will be erased by the inliner. Notice that we will skip all of the
// newly inlined ApplyInsts. That's okay because we will visit them in
// our next invocation of the inliner.
Inliner.inlineFunction(AI, Callee, ArrayRef<Substitution>(), Args);
NumFunctionsInlined++;
Changed = true;
}
return Changed;
}
class SILPerformanceInlinerPass : public SILModuleTransform {
unsigned Threshold;
public:
SILPerformanceInlinerPass(unsigned th) : Threshold(th) {}
virtual void runOnModule(SILModule &M, SILPassManager *PM) {
CallGraphAnalysis* CGA = PM->getAnalysis<CallGraphAnalysis>();
if (Threshold == 0) {
DEBUG(llvm::dbgs() << "*** The Performance Inliner is disabled ***\n");
return;
}
// Initialize the worklist with a bottom-up call-graph order list of
// functions.
const std::vector<SILFunction *> &Order = CGA->bottomUpCallGraphOrder();
std::vector<SILFunction *> Worklist(Order);
std::reverse(Worklist.begin(), Worklist.end());
SILPerformanceInliner inliner(Threshold);
bool Changed = false;
while (!Worklist.empty()) {
SILFunction *F = Worklist.back();
Worklist.pop_back();
Changed |= inliner.inlineCallsIntoFunction(F);
}
// Invalidate the call graph.
if (Changed)
CGA->invalidate(SILAnalysis::InvalidationKind::CallGraph);
}
};
SILTransform *swift::createPerfInliner(unsigned threshold) {
return new SILPerformanceInlinerPass(threshold);
}