llvm-project/bolt/lib/Passes/Inliner.cpp

598 lines
18 KiB
C++
Raw Normal View History

//===--- Passes/Inliner.cpp - Inlining infra for BOLT ---------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// The current inliner has a limited callee support
// (see Inliner::getInliningInfo() for the most up-to-date details):
//
// * No exception handling
// * No jump tables
// * Single entry point
// * CFI update not supported - breaks unwinding
// * Regular Call Sites:
// - only leaf functions (or callees with only tail calls)
// * no invokes (they can't be tail calls)
// - no direct use of %rsp
// * Tail Call Sites:
// - since the stack is unmodified, the regular call limitations are lifted
//
//===----------------------------------------------------------------------===//
#include "bolt/Passes/Inliner.h"
#include "bolt/Core/MCPlus.h"
#include "llvm/Support/CommandLine.h"
#include <map>
#define DEBUG_TYPE "bolt-inliner"
using namespace llvm;
namespace opts {
extern cl::OptionCategory BoltOptCategory;
static cl::opt<bool>
AdjustProfile("inline-ap",
cl::desc("adjust function profile after inlining"),
cl::ZeroOrMore,
cl::cat(BoltOptCategory));
static cl::list<std::string>
ForceInlineFunctions("force-inline",
cl::CommaSeparated,
cl::desc("list of functions to always consider for inlining"),
cl::value_desc("func1,func2,func3,..."),
cl::Hidden,
cl::cat(BoltOptCategory));
static cl::opt<bool>
InlineAll("inline-all",
cl::desc("inline all functions"),
cl::init(false),
cl::ZeroOrMore,
cl::cat(BoltOptCategory));
static cl::opt<bool>
InlineIgnoreLeafCFI("inline-ignore-leaf-cfi",
cl::desc("inline leaf functions with CFI programs (can break unwinding)"),
cl::init(true),
cl::ZeroOrMore,
cl::ReallyHidden,
cl::cat(BoltOptCategory));
static cl::opt<bool>
InlineIgnoreCFI("inline-ignore-cfi",
cl::desc("inline functions with CFI programs (can break exception handling)"),
cl::init(false),
cl::ZeroOrMore,
cl::ReallyHidden,
cl::cat(BoltOptCategory));
static cl::opt<unsigned>
InlineLimit("inline-limit",
cl::desc("maximum number of call sites to inline"),
cl::init(0),
cl::ZeroOrMore,
cl::Hidden,
cl::cat(BoltOptCategory));
static cl::opt<unsigned>
InlineMaxIters("inline-max-iters",
cl::desc("maximum number of inline iterations"),
cl::init(3),
cl::ZeroOrMore,
cl::Hidden,
cl::cat(BoltOptCategory));
static cl::opt<bool>
InlineSmallFunctions("inline-small-functions",
cl::desc("inline functions if increase in size is less than defined by "
"-inline-small-functions-bytes"),
cl::init(false),
cl::ZeroOrMore,
cl::cat(BoltOptCategory));
static cl::opt<unsigned>
InlineSmallFunctionsBytes("inline-small-functions-bytes",
cl::desc("max number of bytes for the function to be considered small for "
"inlining purposes"),
cl::init(4),
cl::ZeroOrMore,
cl::Hidden,
cl::cat(BoltOptCategory));
static cl::opt<bool>
NoInline("no-inline",
cl::desc("disable all inlining (overrides other inlining options)"),
cl::init(false),
cl::ZeroOrMore,
cl::cat(BoltOptCategory));
/// This function returns true if any of inlining options are specified and the
/// inlining pass should be executed. Whenever a new inlining option is added,
/// this function should reflect the change.
bool inliningEnabled() {
return !NoInline &&
(InlineAll ||
InlineSmallFunctions ||
!ForceInlineFunctions.empty());
}
bool mustConsider(const llvm::bolt::BinaryFunction &Function) {
for (std::string &Name : opts::ForceInlineFunctions) {
if (Function.hasName(Name))
return true;
}
return false;
}
void syncOptions() {
if (opts::InlineIgnoreCFI)
opts::InlineIgnoreLeafCFI = true;
if (opts::InlineAll)
opts::InlineSmallFunctions = true;
}
} // namespace opts
namespace llvm {
namespace bolt {
uint64_t Inliner::SizeOfCallInst;
uint64_t Inliner::SizeOfTailCallInst;
uint64_t Inliner::getSizeOfCallInst(const BinaryContext &BC) {
if (SizeOfCallInst)
return SizeOfCallInst;
MCInst Inst;
BC.MIB->createCall(Inst, BC.Ctx->createNamedTempSymbol(), BC.Ctx.get());
SizeOfCallInst = BC.computeInstructionSize(Inst);
return SizeOfCallInst;
}
uint64_t Inliner::getSizeOfTailCallInst(const BinaryContext &BC) {
if (SizeOfTailCallInst)
return SizeOfTailCallInst;
MCInst Inst;
BC.MIB->createTailCall(Inst, BC.Ctx->createNamedTempSymbol(), BC.Ctx.get());
SizeOfTailCallInst = BC.computeInstructionSize(Inst);
return SizeOfTailCallInst;
}
Inliner::InliningInfo Inliner::getInliningInfo(const BinaryFunction &BF) const {
if (!shouldOptimize(BF))
return INL_NONE;
const BinaryContext &BC = BF.getBinaryContext();
bool DirectSP = false;
bool HasCFI = false;
bool IsLeaf = true;
// Perform necessary checks unless the option overrides it.
if (!opts::mustConsider(BF)) {
if (BF.hasSDTMarker())
return INL_NONE;
if (BF.hasEHRanges())
return INL_NONE;
if (BF.isMultiEntry())
return INL_NONE;
if (BF.hasJumpTables())
return INL_NONE;
const MCPhysReg SPReg = BC.MIB->getStackPointer();
for (const BinaryBasicBlock *BB : BF.layout()) {
for (const MCInst &Inst : *BB) {
// Tail calls are marked as implicitly using the stack pointer and they
// could be inlined.
if (BC.MIB->isTailCall(Inst))
break;
if (BC.MIB->isCFI(Inst)) {
HasCFI = true;
continue;
}
if (BC.MIB->isCall(Inst))
IsLeaf = false;
// Push/pop instructions are straightforward to handle.
if (BC.MIB->isPush(Inst) || BC.MIB->isPop(Inst))
continue;
DirectSP |= BC.MIB->hasDefOfPhysReg(Inst, SPReg) ||
BC.MIB->hasUseOfPhysReg(Inst, SPReg);
}
}
}
if (HasCFI) {
if (!opts::InlineIgnoreLeafCFI)
return INL_NONE;
if (!IsLeaf && !opts::InlineIgnoreCFI)
return INL_NONE;
}
InliningInfo Info(DirectSP ? INL_TAILCALL : INL_ANY);
size_t Size = BF.estimateSize();
Info.SizeAfterInlining = Size;
Info.SizeAfterTailCallInlining = Size;
// Handle special case of the known size reduction.
if (BF.size() == 1) {
// For a regular call the last return instruction could be removed
// (or converted to a branch).
const MCInst *LastInst = BF.back().getLastNonPseudoInstr();
if (LastInst &&
BC.MIB->isReturn(*LastInst) &&
!BC.MIB->isTailCall(*LastInst)) {
const uint64_t RetInstSize = BC.computeInstructionSize(*LastInst);
assert(Size >= RetInstSize);
Info.SizeAfterInlining -= RetInstSize;
}
}
return Info;
}
void
Inliner::findInliningCandidates(BinaryContext &BC) {
for (const auto &BFI : BC.getBinaryFunctions()) {
const BinaryFunction &Function = BFI.second;
const InliningInfo InlInfo = getInliningInfo(Function);
if (InlInfo.Type != INL_NONE)
InliningCandidates[&Function] = InlInfo;
}
}
std::pair<BinaryBasicBlock *, BinaryBasicBlock::iterator>
Inliner::inlineCall(BinaryBasicBlock &CallerBB,
BinaryBasicBlock::iterator CallInst,
const BinaryFunction &Callee) {
BinaryFunction &CallerFunction = *CallerBB.getFunction();
BinaryContext &BC = CallerFunction.getBinaryContext();
auto &MIB = *BC.MIB;
assert(MIB.isCall(*CallInst) && "can only inline a call or a tail call");
assert(!Callee.isMultiEntry() &&
"cannot inline function with multiple entries");
assert(!Callee.hasJumpTables() &&
"cannot inline function with jump table(s)");
// Get information about the call site.
const bool CSIsInvoke = BC.MIB->isInvoke(*CallInst);
const bool CSIsTailCall = BC.MIB->isTailCall(*CallInst);
const int64_t CSGNUArgsSize = BC.MIB->getGnuArgsSize(*CallInst);
const Optional<MCPlus::MCLandingPad> CSEHInfo = BC.MIB->getEHInfo(*CallInst);
// Split basic block at the call site if there will be more incoming edges
// coming from the callee.
BinaryBasicBlock *FirstInlinedBB = &CallerBB;
if (Callee.front().pred_size() && CallInst != CallerBB.begin()) {
FirstInlinedBB = CallerBB.splitAt(CallInst);
CallInst = FirstInlinedBB->begin();
}
// Split basic block after the call instruction unless the callee is trivial
// (i.e. consists of a single basic block). If necessary, obtain a basic block
// for return instructions in the callee to redirect to.
BinaryBasicBlock *NextBB = nullptr;
if (Callee.size() > 1) {
if (std::next(CallInst) != FirstInlinedBB->end()) {
NextBB = FirstInlinedBB->splitAt(std::next(CallInst));
} else {
NextBB = FirstInlinedBB->getSuccessor();
}
}
if (NextBB)
FirstInlinedBB->removeSuccessor(NextBB);
// Remove the call instruction.
auto InsertII = FirstInlinedBB->eraseInstruction(CallInst);
double ProfileRatio = 0;
if (uint64_t CalleeExecCount = Callee.getKnownExecutionCount()) {
ProfileRatio =
(double) FirstInlinedBB->getKnownExecutionCount() / CalleeExecCount;
}
// Save execution count of the first block as we don't want it to change
// later due to profile adjustment rounding errors.
const uint64_t FirstInlinedBBCount = FirstInlinedBB->getKnownExecutionCount();
// Copy basic blocks and maintain a map from their origin.
std::unordered_map<const BinaryBasicBlock *, BinaryBasicBlock *> InlinedBBMap;
InlinedBBMap[&Callee.front()] = FirstInlinedBB;
for (auto BBI = std::next(Callee.begin()); BBI != Callee.end(); ++BBI) {
BinaryBasicBlock *InlinedBB = CallerFunction.addBasicBlock(0);
InlinedBBMap[&*BBI] = InlinedBB;
InlinedBB->setCFIState(FirstInlinedBB->getCFIState());
if (Callee.hasValidProfile()) {
InlinedBB->setExecutionCount(BBI->getKnownExecutionCount());
} else {
InlinedBB->setExecutionCount(FirstInlinedBBCount);
}
}
// Copy over instructions and edges.
for (const BinaryBasicBlock &BB : Callee) {
BinaryBasicBlock *InlinedBB = InlinedBBMap[&BB];
if (InlinedBB != FirstInlinedBB)
InsertII = InlinedBB->begin();
// Copy over instructions making any necessary mods.
for (MCInst Inst : BB) {
if (MIB.isPseudo(Inst))
continue;
MIB.stripAnnotations(Inst, /*KeepTC=*/BC.isX86());
// Fix branch target. Strictly speaking, we don't have to do this as
// targets of direct branches will be fixed later and don't matter
// in the CFG state. However, disassembly may look misleading, and
// hence we do the fixing.
if (MIB.isBranch(Inst)) {
assert(!MIB.isIndirectBranch(Inst) &&
"unexpected indirect branch in callee");
const BinaryBasicBlock *TargetBB =
Callee.getBasicBlockForLabel(MIB.getTargetSymbol(Inst));
assert(TargetBB && "cannot find target block in callee");
MIB.replaceBranchTarget(Inst, InlinedBBMap[TargetBB]->getLabel(),
BC.Ctx.get());
}
if (CSIsTailCall || (!MIB.isCall(Inst) && !MIB.isReturn(Inst))) {
InsertII = std::next(InlinedBB->insertInstruction(InsertII,
std::move(Inst)));
continue;
}
// Handle special instructions for a non-tail call site.
if (!MIB.isCall(Inst)) {
// Returns are removed.
break;
}
MIB.convertTailCallToCall(Inst);
// Propagate EH-related info to call instructions.
if (CSIsInvoke) {
MIB.addEHInfo(Inst, *CSEHInfo);
if (CSGNUArgsSize >= 0)
MIB.addGnuArgsSize(Inst, CSGNUArgsSize);
}
InsertII = std::next(InlinedBB->insertInstruction(InsertII,
std::move(Inst)));
}
// Add CFG edges to the basic blocks of the inlined instance.
std::vector<BinaryBasicBlock *> Successors(BB.succ_size());
std::transform(
BB.succ_begin(),
BB.succ_end(),
Successors.begin(),
[&InlinedBBMap](const BinaryBasicBlock *BB) {
return InlinedBBMap.at(BB);
});
if (CallerFunction.hasValidProfile() && Callee.hasValidProfile()) {
InlinedBB->addSuccessors(
Successors.begin(),
Successors.end(),
BB.branch_info_begin(),
BB.branch_info_end());
} else {
InlinedBB->addSuccessors(
Successors.begin(),
Successors.end());
}
if (!CSIsTailCall && BB.succ_size() == 0 && NextBB) {
// Either it's a return block or the last instruction never returns.
InlinedBB->addSuccessor(NextBB, InlinedBB->getExecutionCount());
}
// Scale profiling info for blocks and edges after inlining.
if (CallerFunction.hasValidProfile() && Callee.size() > 1) {
if (opts::AdjustProfile) {
InlinedBB->adjustExecutionCount(ProfileRatio);
} else {
InlinedBB->setExecutionCount(
InlinedBB->getKnownExecutionCount() * ProfileRatio);
}
}
}
// Restore the original execution count of the first inlined basic block.
FirstInlinedBB->setExecutionCount(FirstInlinedBBCount);
CallerFunction.recomputeLandingPads();
if (NextBB)
return std::make_pair(NextBB, NextBB->begin());
if (Callee.size() == 1)
return std::make_pair(FirstInlinedBB, InsertII);
return std::make_pair(FirstInlinedBB, FirstInlinedBB->end());
}
bool Inliner::inlineCallsInFunction(BinaryFunction &Function) {
BinaryContext &BC = Function.getBinaryContext();
std::vector<BinaryBasicBlock *> Blocks(Function.layout().begin(),
Function.layout().end());
std::sort(Blocks.begin(), Blocks.end(),
[](const BinaryBasicBlock *BB1, const BinaryBasicBlock *BB2) {
return BB1->getKnownExecutionCount() > BB2->getKnownExecutionCount();
});
bool DidInlining = false;
for (BinaryBasicBlock *BB : Blocks) {
for (auto InstIt = BB->begin(); InstIt != BB->end(); ) {
MCInst &Inst = *InstIt;
if (!BC.MIB->isCall(Inst) || MCPlus::getNumPrimeOperands(Inst) != 1 ||
!Inst.getOperand(0).isExpr()) {
++InstIt;
continue;
}
const MCSymbol *TargetSymbol = BC.MIB->getTargetSymbol(Inst);
assert(TargetSymbol && "target symbol expected for direct call");
// Don't inline calls to a secondary entry point in a target function.
uint64_t EntryID = 0;
BinaryFunction *TargetFunction =
BC.getFunctionForSymbol(TargetSymbol, &EntryID);
if (!TargetFunction || EntryID != 0) {
++InstIt;
continue;
}
// Don't do recursive inlining.
if (TargetFunction == &Function) {
++InstIt;
continue;
}
auto IInfo = InliningCandidates.find(TargetFunction);
if (IInfo == InliningCandidates.end()) {
++InstIt;
continue;
}
const bool IsTailCall = BC.MIB->isTailCall(Inst);
if (!IsTailCall && IInfo->second.Type == INL_TAILCALL) {
++InstIt;
continue;
}
int64_t SizeAfterInlining;
if (IsTailCall) {
SizeAfterInlining = IInfo->second.SizeAfterTailCallInlining -
getSizeOfTailCallInst(BC);
} else {
SizeAfterInlining = IInfo->second.SizeAfterInlining -
getSizeOfCallInst(BC);
}
if (!opts::InlineAll && !opts::mustConsider(*TargetFunction)) {
if (!opts::InlineSmallFunctions ||
SizeAfterInlining > opts::InlineSmallFunctionsBytes) {
++InstIt;
continue;
}
}
LLVM_DEBUG(dbgs() << "BOLT-DEBUG: inlining call to " << *TargetFunction
<< " in " << Function << " : " << BB->getName()
<< ". Count: " << BB->getKnownExecutionCount()
<< ". Size change: " << SizeAfterInlining
<< " bytes.\n");
std::tie(BB, InstIt) = inlineCall(*BB, InstIt, *TargetFunction);
DidInlining = true;
TotalInlinedBytes += SizeAfterInlining;
++NumInlinedCallSites;
NumInlinedDynamicCalls += BB->getExecutionCount();
// Subtract basic block execution count from the callee execution count.
if (opts::AdjustProfile) {
TargetFunction->adjustExecutionCount(BB->getKnownExecutionCount());
}
// Check if the caller inlining status has to be adjusted.
if (IInfo->second.Type == INL_TAILCALL) {
auto CallerIInfo = InliningCandidates.find(&Function);
if (CallerIInfo != InliningCandidates.end() &&
CallerIInfo->second.Type == INL_ANY) {
LLVM_DEBUG(dbgs() << "adjusting inlining status for function "
<< Function << '\n');
CallerIInfo->second.Type = INL_TAILCALL;
}
}
if (NumInlinedCallSites == opts::InlineLimit) {
return true;
}
}
}
return DidInlining;
}
void Inliner::runOnFunctions(BinaryContext &BC) {
opts::syncOptions();
if (!opts::inliningEnabled())
return;
uint64_t TotalSize = 0;
for (auto &BFI : BC.getBinaryFunctions())
TotalSize += BFI.second.getSize();
bool InlinedOnce;
unsigned NumIters = 0;
do {
if (opts::InlineLimit && NumInlinedCallSites >= opts::InlineLimit)
break;
InlinedOnce = false;
InliningCandidates.clear();
findInliningCandidates(BC);
std::vector<BinaryFunction *> ConsideredFunctions;
for (auto &BFI : BC.getBinaryFunctions()) {
BinaryFunction &Function = BFI.second;
if (!shouldOptimize(Function))
continue;
ConsideredFunctions.push_back(&Function);
}
std::sort(ConsideredFunctions.begin(), ConsideredFunctions.end(),
[](const BinaryFunction *A, const BinaryFunction *B) {
return B->getKnownExecutionCount() < A->getKnownExecutionCount();
});
for (BinaryFunction *Function : ConsideredFunctions) {
if (opts::InlineLimit && NumInlinedCallSites >= opts::InlineLimit)
break;
const bool DidInline = inlineCallsInFunction(*Function);
if (DidInline)
Modified.insert(Function);
InlinedOnce |= DidInline;
}
++NumIters;
} while (InlinedOnce && NumIters < opts::InlineMaxIters);
if (NumInlinedCallSites) {
outs() << "BOLT-INFO: inlined " << NumInlinedDynamicCalls << " calls at "
<< NumInlinedCallSites << " call sites in " << NumIters
<< " iteration(s). Change in binary size: " << TotalInlinedBytes
<< " bytes.\n";
}
}
} // namespace bolt
} // namespace llvm