Alexandre Ganea 8404aeb56a [Support] On Windows, ensure hardware_concurrency() extends to all CPU sockets and all NUMA groups
The goal of this patch is to maximize CPU utilization on multi-socket or high core count systems, so that parallel computations such as LLD/ThinLTO can use all hardware threads in the system. Before this patch, on Windows, a maximum of 64 hardware threads could be used at most, in some cases dispatched only on one CPU socket.

== Background ==
Windows doesn't have a flat cpu_set_t like Linux. Instead, it projects hardware CPUs (or NUMA nodes) to applications through a concept of "processor groups". A "processor" is the smallest unit of execution on a CPU, that is, an hyper-thread if SMT is active; a core otherwise. There's a limit of 32-bit processors on older 32-bit versions of Windows, which later was raised to 64-processors with 64-bit versions of Windows. This limit comes from the affinity mask, which historically is represented by the sizeof(void*). Consequently, the concept of "processor groups" was introduced for dealing with systems with more than 64 hyper-threads.

By default, the Windows OS assigns only one "processor group" to each starting application, in a round-robin manner. If the application wants to use more processors, it needs to programmatically enable it, by assigning threads to other "processor groups". This also means that affinity cannot cross "processor group" boundaries; one can only specify a "preferred" group on start-up, but the application is free to allocate more groups if it wants to.

This creates a peculiar situation, where newer CPUs like the AMD EPYC 7702P (64-cores, 128-hyperthreads) are projected by the OS as two (2) "processor groups". This means that by default, an application can only use half of the cores. This situation could only get worse in the years to come, as dies with more cores will appear on the market.

== The problem ==
The heavyweight_hardware_concurrency() API was introduced so that only *one hardware thread per core* was used. Once that API returns, that original intention is lost, only the number of threads is retained. Consider a situation, on Windows, where the system has 2 CPU sockets, 18 cores each, each core having 2 hyper-threads, for a total of 72 hyper-threads. Both heavyweight_hardware_concurrency() and hardware_concurrency() currently return 36, because on Windows they are simply wrappers over std:🧵:hardware_concurrency() -- which can only return processors from the current "processor group".

== The changes in this patch ==
To solve this situation, we capture (and retain) the initial intention until the point of usage, through a new ThreadPoolStrategy class. The number of threads to use is deferred as late as possible, until the moment where the std::threads are created (ThreadPool in the case of ThinLTO).

When using hardware_concurrency(), setting ThreadCount to 0 now means to use all the possible hardware CPU (SMT) threads. Providing a ThreadCount above to the maximum number of threads will have no effect, the maximum will be used instead.
The heavyweight_hardware_concurrency() is similar to hardware_concurrency(), except that only one thread per hardware *core* will be used.

When LLVM_ENABLE_THREADS is OFF, the threading APIs will always return 1, to ensure any caller loops will be exercised at least once.

Differential Revision: https://reviews.llvm.org/D71775
2020-02-14 10:24:22 -05:00

267 lines
8.5 KiB
C++

//===--------- LLJIT.cpp - An ORC-based JIT for compiling LLVM IR ---------===//
//
// 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
//
//===----------------------------------------------------------------------===//
#include "llvm/ExecutionEngine/Orc/LLJIT.h"
#include "llvm/ExecutionEngine/JITLink/JITLinkMemoryManager.h"
#include "llvm/ExecutionEngine/Orc/ObjectLinkingLayer.h"
#include "llvm/ExecutionEngine/Orc/OrcError.h"
#include "llvm/ExecutionEngine/Orc/RTDyldObjectLinkingLayer.h"
#include "llvm/ExecutionEngine/SectionMemoryManager.h"
#include "llvm/IR/Mangler.h"
namespace llvm {
namespace orc {
Error LLJITBuilderState::prepareForConstruction() {
if (!JTMB) {
if (auto JTMBOrErr = JITTargetMachineBuilder::detectHost())
JTMB = std::move(*JTMBOrErr);
else
return JTMBOrErr.takeError();
}
// If the client didn't configure any linker options then auto-configure the
// JIT linker.
if (!CreateObjectLinkingLayer && JTMB->getCodeModel() == None &&
JTMB->getRelocationModel() == None) {
auto &TT = JTMB->getTargetTriple();
if (TT.isOSBinFormatMachO() &&
(TT.getArch() == Triple::aarch64 || TT.getArch() == Triple::x86_64)) {
JTMB->setRelocationModel(Reloc::PIC_);
JTMB->setCodeModel(CodeModel::Small);
CreateObjectLinkingLayer =
[](ExecutionSession &ES,
const Triple &) -> std::unique_ptr<ObjectLayer> {
return std::make_unique<ObjectLinkingLayer>(
ES, std::make_unique<jitlink::InProcessMemoryManager>());
};
}
}
return Error::success();
}
LLJIT::~LLJIT() {
if (CompileThreads)
CompileThreads->wait();
}
Error LLJIT::defineAbsolute(StringRef Name, JITEvaluatedSymbol Sym) {
auto InternedName = ES->intern(Name);
SymbolMap Symbols({{InternedName, Sym}});
return Main.define(absoluteSymbols(std::move(Symbols)));
}
Error LLJIT::addIRModule(JITDylib &JD, ThreadSafeModule TSM) {
assert(TSM && "Can not add null module");
if (auto Err =
TSM.withModuleDo([&](Module &M) { return applyDataLayout(M); }))
return Err;
return TransformLayer->add(JD, std::move(TSM), ES->allocateVModule());
}
Error LLJIT::addObjectFile(JITDylib &JD, std::unique_ptr<MemoryBuffer> Obj) {
assert(Obj && "Can not add null object");
return ObjTransformLayer.add(JD, std::move(Obj), ES->allocateVModule());
}
Expected<JITEvaluatedSymbol> LLJIT::lookupLinkerMangled(JITDylib &JD,
StringRef Name) {
return ES->lookup(
makeJITDylibSearchOrder(&JD, JITDylibLookupFlags::MatchAllSymbols),
ES->intern(Name));
}
std::unique_ptr<ObjectLayer>
LLJIT::createObjectLinkingLayer(LLJITBuilderState &S, ExecutionSession &ES) {
// If the config state provided an ObjectLinkingLayer factory then use it.
if (S.CreateObjectLinkingLayer)
return S.CreateObjectLinkingLayer(ES, S.JTMB->getTargetTriple());
// Otherwise default to creating an RTDyldObjectLinkingLayer that constructs
// a new SectionMemoryManager for each object.
auto GetMemMgr = []() { return std::make_unique<SectionMemoryManager>(); };
auto ObjLinkingLayer =
std::make_unique<RTDyldObjectLinkingLayer>(ES, std::move(GetMemMgr));
if (S.JTMB->getTargetTriple().isOSBinFormatCOFF()) {
ObjLinkingLayer->setOverrideObjectFlagsWithResponsibilityFlags(true);
ObjLinkingLayer->setAutoClaimResponsibilityForObjectSymbols(true);
}
// FIXME: Explicit conversion to std::unique_ptr<ObjectLayer> added to silence
// errors from some GCC / libstdc++ bots. Remove this conversion (i.e.
// just return ObjLinkingLayer) once those bots are upgraded.
return std::unique_ptr<ObjectLayer>(std::move(ObjLinkingLayer));
}
Expected<std::unique_ptr<IRCompileLayer::IRCompiler>>
LLJIT::createCompileFunction(LLJITBuilderState &S,
JITTargetMachineBuilder JTMB) {
/// If there is a custom compile function creator set then use it.
if (S.CreateCompileFunction)
return S.CreateCompileFunction(std::move(JTMB));
// Otherwise default to creating a SimpleCompiler, or ConcurrentIRCompiler,
// depending on the number of threads requested.
if (S.NumCompileThreads > 0)
return std::make_unique<ConcurrentIRCompiler>(std::move(JTMB));
auto TM = JTMB.createTargetMachine();
if (!TM)
return TM.takeError();
return std::make_unique<TMOwningSimpleCompiler>(std::move(*TM));
}
LLJIT::LLJIT(LLJITBuilderState &S, Error &Err)
: ES(S.ES ? std::move(S.ES) : std::make_unique<ExecutionSession>()),
Main(this->ES->createJITDylib("<main>")), DL(""),
TT(S.JTMB->getTargetTriple()),
ObjLinkingLayer(createObjectLinkingLayer(S, *ES)),
ObjTransformLayer(*this->ES, *ObjLinkingLayer), CtorRunner(Main),
DtorRunner(Main) {
ErrorAsOutParameter _(&Err);
if (auto DLOrErr = S.JTMB->getDefaultDataLayoutForTarget())
DL = std::move(*DLOrErr);
else {
Err = DLOrErr.takeError();
return;
}
{
auto CompileFunction = createCompileFunction(S, std::move(*S.JTMB));
if (!CompileFunction) {
Err = CompileFunction.takeError();
return;
}
CompileLayer = std::make_unique<IRCompileLayer>(
*ES, ObjTransformLayer, std::move(*CompileFunction));
TransformLayer = std::make_unique<IRTransformLayer>(*ES, *CompileLayer);
}
if (S.NumCompileThreads > 0) {
TransformLayer->setCloneToNewContextOnEmit(true);
CompileThreads =
std::make_unique<ThreadPool>(hardware_concurrency(S.NumCompileThreads));
ES->setDispatchMaterialization(
[this](JITDylib &JD, std::unique_ptr<MaterializationUnit> MU) {
// FIXME: Switch to move capture once we have c++14.
auto SharedMU = std::shared_ptr<MaterializationUnit>(std::move(MU));
auto Work = [SharedMU, &JD]() { SharedMU->doMaterialize(JD); };
CompileThreads->async(std::move(Work));
});
}
}
std::string LLJIT::mangle(StringRef UnmangledName) {
std::string MangledName;
{
raw_string_ostream MangledNameStream(MangledName);
Mangler::getNameWithPrefix(MangledNameStream, UnmangledName, DL);
}
return MangledName;
}
Error LLJIT::applyDataLayout(Module &M) {
if (M.getDataLayout().isDefault())
M.setDataLayout(DL);
if (M.getDataLayout() != DL)
return make_error<StringError>(
"Added modules have incompatible data layouts",
inconvertibleErrorCode());
return Error::success();
}
void LLJIT::recordCtorDtors(Module &M) {
CtorRunner.add(getConstructors(M));
DtorRunner.add(getDestructors(M));
}
Error LLLazyJITBuilderState::prepareForConstruction() {
if (auto Err = LLJITBuilderState::prepareForConstruction())
return Err;
TT = JTMB->getTargetTriple();
return Error::success();
}
Error LLLazyJIT::addLazyIRModule(JITDylib &JD, ThreadSafeModule TSM) {
assert(TSM && "Can not add null module");
if (auto Err = TSM.withModuleDo([&](Module &M) -> Error {
if (auto Err = applyDataLayout(M))
return Err;
recordCtorDtors(M);
return Error::success();
}))
return Err;
return CODLayer->add(JD, std::move(TSM), ES->allocateVModule());
}
LLLazyJIT::LLLazyJIT(LLLazyJITBuilderState &S, Error &Err) : LLJIT(S, Err) {
// If LLJIT construction failed then bail out.
if (Err)
return;
ErrorAsOutParameter _(&Err);
/// Take/Create the lazy-compile callthrough manager.
if (S.LCTMgr)
LCTMgr = std::move(S.LCTMgr);
else {
if (auto LCTMgrOrErr = createLocalLazyCallThroughManager(
S.TT, *ES, S.LazyCompileFailureAddr))
LCTMgr = std::move(*LCTMgrOrErr);
else {
Err = LCTMgrOrErr.takeError();
return;
}
}
// Take/Create the indirect stubs manager builder.
auto ISMBuilder = std::move(S.ISMBuilder);
// If none was provided, try to build one.
if (!ISMBuilder)
ISMBuilder = createLocalIndirectStubsManagerBuilder(S.TT);
// No luck. Bail out.
if (!ISMBuilder) {
Err = make_error<StringError>("Could not construct "
"IndirectStubsManagerBuilder for target " +
S.TT.str(),
inconvertibleErrorCode());
return;
}
// Create the COD layer.
CODLayer = std::make_unique<CompileOnDemandLayer>(
*ES, *TransformLayer, *LCTMgr, std::move(ISMBuilder));
if (S.NumCompileThreads > 0)
CODLayer->setCloneToNewContextOnEmit(true);
}
} // End namespace orc.
} // End namespace llvm.