1//===-- BasicBlockSections.cpp ---=========--------------------------------===// 3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4// See https://llvm.org/LICENSE.txt for license information. 5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 7//===----------------------------------------------------------------------===// 9// BasicBlockSections implementation. 11// The purpose of this pass is to assign sections to basic blocks when 12// -fbasic-block-sections= option is used. Further, with profile information 13// only the subset of basic blocks with profiles are placed in separate sections 14// and the rest are grouped in a cold section. The exception handling blocks are 15// treated specially to ensure they are all in one seciton. 17// Basic Block Sections 18// ==================== 20// With option, -fbasic-block-sections=list, every function may be split into 21// clusters of basic blocks. Every cluster will be emitted into a separate 22// section with its basic blocks sequenced in the given order. To get the 23// optimized performance, the clusters must form an optimal BB layout for the 24// function. We insert a symbol at the beginning of every cluster's section to 25// allow the linker to reorder the sections in any arbitrary sequence. A global 26// order of these sections would encapsulate the function layout. 27// For example, consider the following clusters for a function foo (consisting 28// of 6 basic blocks 0, 1, ..., 5). 33// * Basic blocks 0 and 2 are placed in one section with symbol `foo` 34// referencing the beginning of this section. 35// * Basic blocks 1, 3, 5 are placed in a separate section. A new symbol 36// `foo.__part.1` will reference the beginning of this section. 37// * Basic block 4 (note that it is not referenced in the list) is placed in 38// one section, and a new symbol `foo.cold` will point to it. 40// There are a couple of challenges to be addressed: 42// 1. The last basic block of every cluster should not have any implicit 43// fallthrough to its next basic block, as it can be reordered by the linker. 44// The compiler should make these fallthroughs explicit by adding 45// unconditional jumps.. 47// 2. All inter-cluster branch targets would now need to be resolved by the 48// linker as they cannot be calculated during compile time. This is done 49// using static relocations. Further, the compiler tries to use short branch 50// instructions on some ISAs for small branch offsets. This is not possible 51// for inter-cluster branches as the offset is not determined at compile 52// time, and therefore, long branch instructions have to be used for those. 54// 3. Debug Information (DebugInfo) and Call Frame Information (CFI) emission 55// needs special handling with basic block sections. DebugInfo needs to be 56// emitted with more relocations as basic block sections can break a 57// function into potentially several disjoint pieces, and CFI needs to be 58// emitted per cluster. This also bloats the object file and binary sizes. 60// Basic Block Address Map 63// With -fbasic-block-address-map, we emit the offsets of BB addresses of 64// every function into the .llvm_bb_addr_map section. Along with the function 65// symbols, this allows for mapping of virtual addresses in PMU profiles back to 66// the corresponding basic blocks. This logic is implemented in AsmPrinter. This 67// pass only assigns the BBSectionType of every function to ``labels``. 69//===----------------------------------------------------------------------===// 87// Placing the cold clusters in a separate section mitigates against poor 88// profiles and allows optimizations such as hugepage mapping to be applied at a 89// section granularity. Defaults to ".text.split." which is recognized by lld 90// via the `-z keep-text-section-prefix` flag. 92"bbsections-cold-text-prefix",
93cl::desc(
"The text prefix to use for cold basic block clusters"),
97"bbsections-detect-source-drift",
98cl::desc(
"This checks if there is a fdo instr. profile hash " 99"mismatch for this function"),
115return"Basic Block Sections Analysis";
120 /// Identify basic blocks that need separate sections and prepare to emit them 129}
// end anonymous namespace 131char BasicBlockSections::ID = 0;
133 BasicBlockSections,
"bbsections-prepare",
134"Prepares for basic block sections, by splitting functions " 135"into clusters of basic blocks.",
143// This function updates and optimizes the branching instructions of every basic 144// block in a given function to account for changes in the layout. 150for (
auto &
MBB : MF) {
153// If this block had a fallthrough before we need an explicit unconditional 154// branch to that block if either 155// 1- the block ends a section, which means its next block may be 156// reorderd by the linker, or 157// 2- the fallthrough block is not adjacent to the block in the new 162// We do not optimize branches for machine basic blocks ending sections, as 163// their adjacent block might be reordered by the linker. 167// It might be possible to optimize branches by flipping the branch 177// This function sorts basic blocks according to the cluster's information. 178// All explicitly specified clusters of basic blocks will be ordered 179// accordingly. All non-specified BBs go into a separate "Cold" section. 180// Additionally, if exception handling landing pads end up in more than one 181// clusters, they are moved into a single "Exception" section. Eventually, 182// clusters are ordered in increasing order of their IDs, with the "Exception" 183// and "Cold" succeeding all other clusters. 184// FuncClusterInfo represents the cluster information for basic blocks. It 185// maps from BBID of basic blocks to their cluster information. If this is 186// empty, it means unique sections for all basic blocks in the function. 191// This variable stores the section ID of the cluster containing eh_pads (if 192// all eh_pads are one cluster). If more than one cluster contain eh_pads, we 193// set it equal to ExceptionSectionID. 194 std::optional<MBBSectionID> EHPadsSectionID;
196for (
auto &
MBB : MF) {
197// With the 'all' option, every basic block is placed in a unique section. 198// With the 'list' option, every basic block is placed in a section 199// associated with its cluster, unless we want individual unique sections 200// for every basic block in this function (if FuncClusterInfo is empty). 202 FuncClusterInfo.
empty()) {
203// If unique sections are desired for all basic blocks of the function, we 204// set every basic block's section ID equal to its original position in 205// the layout (which is equal to its number). This ensures that basic 206// blocks are ordered canonically. 210if (
I != FuncClusterInfo.
end()) {
216if (
TII.isMBBSafeToSplitToCold(
MBB)) {
217// BB goes into the special cold section if it is not specified in the 226// If we already have one cluster containing eh_pads, this must be updated 227// to ExceptionSectionID. Otherwise, we set it equal to the current 234// If EHPads are in more than one section, this places all of them in the 235// special exception section. 252"Entry block should not be displaced by basic block sections");
254// Set IsBeginSection and IsEndSection according to the assigned section IDs. 257// After reordering basic blocks, we must update basic block branches to 258// insert explicit fallthrough branches when required and optimize branches 263// If the exception section begins with a landing pad, that landing pad will 264// assume a zero offset (relative to @LPStart) in the LSDA. However, a value of 265// zero implies "no landing pad." This function inserts a NOP just before the EH 266// pad label to ensure a nonzero offset. 268for (
auto &
MBB : MF) {
271while (!
MI->isEHLabel())
282constchar MetadataName[] =
"instr_prof_hash_mismatch";
285MDTuple *Tuple = cast<MDTuple>(Existing);
287if (
N.equalsStr(MetadataName))
294// Identify, arrange, and modify basic blocks which need separate sections 295// according to the specification provided by the -fbasic-block-sections flag. 298if (BBSectionsType == BasicBlockSection::None)
301// Check for source drift. If the source has changed since the profiles 302// were obtained, optimizing basic blocks might be sub-optimal. 303// This only applies to BasicBlockSection::List as it creates 304// clusters of basic blocks using basic block ids. Source drift can 305// invalidate these groupings leading to sub-optimal code generation with 306// regards to performance. 307if (BBSectionsType == BasicBlockSection::List &&
310// Renumber blocks before sorting them. This is useful for accessing the 311// original layout positions and finding the original fallthroughs. 315if (BBSectionsType == BasicBlockSection::List) {
316auto [HasProfile, ClusterInfo] =
317 getAnalysis<BasicBlockSectionsProfileReaderWrapperPass>()
318 .getClusterInfoForFunction(MF.
getName());
332// Helper function for ordering BB sections as follows: 333// * Entry section (section including the entry block). 334// * Regular sections (in increasing order of their Number). 336// * Exception section 340// We make sure that the section containing the entry block precedes all the 342if (LHS == EntryBBSectionID || RHS == EntryBBSectionID)
343returnLHS == EntryBBSectionID;
347// We sort all basic blocks to make sure the basic blocks of every cluster are 348// contiguous and ordered accordingly. Furthermore, clusters are ordered in 349// increasing order of their section IDs, with the exception and the 350// cold section placed at the end of the function. 351// Also, we force the entry block of the function to be placed at the 352// beginning of the function, regardless of the requested order. 355auto XSectionID =
X.getSectionID();
356auto YSectionID =
Y.getSectionID();
357if (XSectionID != YSectionID)
358return MBBSectionOrder(XSectionID, YSectionID);
359// Make sure that the entry block is placed at the beginning. 360if (&
X == &EntryBB || &
Y == &EntryBB)
361return &
X == &EntryBB;
362// If the two basic block are in the same section, the order is decided by 363// their position within the section. 364if (XSectionID.Type == MBBSectionID::SectionType::Default)
365return FuncClusterInfo.
lookup(*
X.getBBID()).PositionInCluster <
366 FuncClusterInfo.
lookup(*
Y.getBBID()).PositionInCluster;
367returnX.getNumber() <
Y.getNumber();
375// When the BB address map needs to be generated, this renumbers basic blocks to 376// make them appear in increasing order of their IDs in the function. This 377// avoids the need to store basic block IDs in the BB address map section, since 378// they can be determined implicitly. 387// First handle the basic block sections. 388auto R1 = handleBBSections(MF);
389// Handle basic block address map after basic block sections are finalized. 390autoR2 = handleBBAddrMap(MF);
392// We renumber blocks, so update the dominator tree we want to preserve. 393if (
auto *WP = getAnalysisIfAvailable<MachineDominatorTreeWrapperPass>())
394 WP->getDomTree().updateBlockNumbers();
395if (
auto *WP = getAnalysisIfAvailable<MachinePostDominatorTreeWrapperPass>())
396 WP->getPostDomTree().updateBlockNumbers();
401void BasicBlockSections::getAnalysisUsage(
AnalysisUsage &AU)
const{
410returnnew BasicBlockSections();
for(const MachineOperand &MO :llvm::drop_begin(OldMI.operands(), Desc.getNumOperands()))
Lower uses of LDS variables from non kernel functions
static void assignSections(MachineFunction &MF, const DenseMap< UniqueBBID, BBClusterInfo > &FuncClusterInfo)
static cl::opt< bool > BBSectionsDetectSourceDrift("bbsections-detect-source-drift", cl::desc("This checks if there is a fdo instr. profile hash " "mismatch for this function"), cl::init(true), cl::Hidden)
bbsections Prepares for basic block by splitting functions into clusters of basic blocks
bbsections Prepares for basic block by splitting functions into clusters of basic static false void updateBranches(MachineFunction &MF, const SmallVector< MachineBasicBlock * > &PreLayoutFallThroughs)
bbsections Prepares for basic block sections
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
const HexagonInstrInfo * TII
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
#define INITIALIZE_PASS_DEPENDENCY(depName)
#define INITIALIZE_PASS_END(passName, arg, name, cfg, analysis)
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
const SmallVectorImpl< MachineOperand > MachineBasicBlock * TBB
const SmallVectorImpl< MachineOperand > & Cond
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This file defines the SmallVector class.
unify loop Fixup each natural loop to have a single exit block
Represent the analysis usage information of a pass.
AnalysisUsage & addUsedIfAvailable()
Add the specified Pass class to the set of analyses used by this pass.
AnalysisUsage & addRequired()
void setPreservesAll()
Set by analyses that do not transform their input at all.
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
iterator find(const_arg_type_t< KeyT > Val)
std::pair< iterator, bool > try_emplace(KeyT &&Key, Ts &&...Args)
MDNode * getMetadata(unsigned KindID) const
Get the current metadata attachments for the given kind, if any.
bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB, SmallVectorImpl< MachineOperand > &Cond, bool AllowModify) const override
Analyze the branching code at the end of MBB, returning true if it cannot be understood (e....
ArrayRef< MDOperand > operands() const
bool isEHPad() const
Returns true if the block is a landing pad.
MachineBasicBlock * getFallThrough(bool JumpToFallThrough=true)
Return the fallthrough block if the block can implicitly transfer control to the block after it by fa...
int getNumber() const
MachineBasicBlocks are uniquely numbered at the function level, unless they're not in a MachineFuncti...
std::optional< UniqueBBID > getBBID() const
void updateTerminator(MachineBasicBlock *PreviousLayoutSuccessor)
Update the terminator instructions in block to account for changes to block layout which may have bee...
MBBSectionID getSectionID() const
Returns the section ID of this basic block.
void setSectionID(MBBSectionID V)
Sets the section ID for this basic block.
const MachineFunction * getParent() const
Return the MachineFunction containing this basic block.
bool isBeginSection() const
Returns true if this block begins any section.
DebugLoc findBranchDebugLoc()
Find and return the merged DebugLoc of the branch instructions of the block.
bool isEndSection() const
Returns true if this block ends any section.
Analysis pass which computes a MachineDominatorTree.
MachineFunctionPass - This class adapts the FunctionPass interface to allow convenient creation of pa...
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - Subclasses that override getAnalysisUsage must call this.
virtual bool runOnMachineFunction(MachineFunction &MF)=0
runOnMachineFunction - This method must be overloaded to perform the desired machine code transformat...
void setBBSectionsType(BasicBlockSection V)
const TargetSubtargetInfo & getSubtarget() const
getSubtarget - Return the subtarget for which this machine code is being compiled.
StringRef getName() const
getName - Return the name of the corresponding LLVM function.
bool hasBBSections() const
Returns true if this function has basic block sections enabled.
Function & getFunction()
Return the LLVM function that this machine code represents.
unsigned getNumBlockIDs() const
getNumBlockIDs - Return the number of MBB ID's allocated.
void RenumberBlocks(MachineBasicBlock *MBBFrom=nullptr)
RenumberBlocks - This discards all of the MachineBasicBlock numbers and recomputes them.
const MachineBasicBlock & front() const
void assignBeginEndSections()
Assign IsBeginSection IsEndSection fields for basic blocks in this function.
const TargetMachine & getTarget() const
getTarget - Return the target machine this machine code is compiled with
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
virtual StringRef getPassName() const
getPassName - Return a nice clean name for a pass.
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
StringRef - Represent a constant reference to a string, i.e.
TargetInstrInfo - Interface to description of machine instruction set.
virtual void insertNoop(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI) const
Insert a noop into the instruction stream at the specified point.
llvm::BasicBlockSection getBBSectionsType() const
If basic blocks should be emitted into their own section, corresponding to -fbasic-block-sections.
virtual const TargetInstrInfo * getInstrInfo() const
An efficient, type-erasing, non-owning reference to a callable.
self_iterator getIterator()
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
initializer< Ty > init(const Ty &Val)
This is an optimization pass for GlobalISel generic memory operations.
MachineFunctionPass * createBasicBlockSectionsPass()
createBasicBlockSections Pass - This pass assigns sections to machine basic blocks and is enabled wit...
bool hasInstrProfHashMismatch(MachineFunction &MF)
This checks if the source of this function has drifted since this binary was profiled previously.
void initializeBasicBlockSectionsPass(PassRegistry &)
void avoidZeroOffsetLandingPad(MachineFunction &MF)
cl::opt< std::string > BBSectionsColdTextPrefix
void sortBasicBlocksAndUpdateBranches(MachineFunction &MF, MachineBasicBlockComparator MBBCmp)
static const MBBSectionID ExceptionSectionID
static const MBBSectionID ColdSectionID