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
This commit adds a ManglingOptions struct to IRMaterializationUnit, and replaces
IRCompileLayer::CompileFunction with a new IRCompileLayer::IRCompiler class. The
ManglingOptions struct defines the emulated-TLS state (via a bool member,
EmulatedTLS, which is true if emulated-TLS is enabled and false otherwise). The
IRCompileLayer::IRCompiler class wraps an IRCompiler (the same way that the
CompileFunction typedef used to), but adds a method to return the
IRCompileLayer::ManglingOptions that the compiler will use.
These changes allow us to correctly determine the symbols that will be produced
when a thread local global variable defined at the IR level is compiled with or
without emulated TLS. This is required for ORCv2, where MaterializationUnits
must declare their interface up-front.
Most ORCv2 clients should not require any changes. Clients writing custom IR
compilers will need to wrap their compiler in an IRCompileLayer::IRCompiler,
rather than an IRCompileLayer::CompileFunction, however this should be a
straightforward change (see modifications to CompileUtils.* in this patch for an
example).
The MaterializationResponsibility::defineMaterializing method allows clients to
add new definitions that are in the process of being materialized to the JIT.
This patch adds support to defineMaterializing for symbols with weak linkage
where the new definitions may be rejected if another materializer concurrently
defines the same symbol. If a weak symbol is rejected it will not be added to
the MaterializationResponsibility's responsibility set. Clients can check for
membership in the responsibility set via the
MaterializationResponsibility::getSymbols() method before resolving any
such weak symbols.
This patch also adds code to RTDyldObjectLinkingLayer to tag COFF comdat symbols
introduced during codegen as weak, on the assumption that these are COFF comdat
constants. This fixes http://llvm.org/PR40074.
Based on Don Hinton's patch in https://reviews.llvm.org/D72406. This feature
was accidentally left out of e9e26c01cd, and
would have pessimized concurrent compilation in the default case.
Thanks for spotting this Don!
This patch makes the target triple available via the LLJIT interface, and moves
the IRTransformLayer from LLLazyJIT down into LLJIT. Together these changes make
it easier to use the lazyReexports utility with LLJIT, and to apply IR
transforms to code as it is compiled in LLJIT (rather than requiring transforms
to be applied manually before code is added). An code example is added in
llvm/examples/LLJITExamples/LLJITWithLazyReexports
LLJITBuilder will now use JITLink on supported platforms even if a custom
JITTargetMachineBuilder is supplied, provided that neither the code model,
nor the relocation model, nor the ObjectLinkingLayerCreator is set.
JITLink (which underlies ObjectLinkingLayer) is a replacement for RuntimeDyld.
It supports the native code model, and linker plugins that enable a wider range
of features than RuntimeDyld.
Currently only enabled for MachO/x86-64 and MachO/arm64. New architectures will
be added as JITLink support for them is developed.
This patch removes the magic "main" JITDylib from ExecutionEngine. The main
JITDylib was created automatically at ExecutionSession construction time, and
all subsequently created JITDylibs were added to the main JITDylib's
links-against list by default. This saves a couple of lines of boilerplate for
simple JIT setups, but this isn't worth introducing magical behavior for.
ORCv2 clients should now construct their own main JITDylib using
ExecutionSession::createJITDylib and set up its linkages manually using
JITDylib::setSearchOrder (or related methods in JITDylib).
libraries.
This patch substantially updates ORCv2's lookup API in order to support weak
references, and to better support static archives. Key changes:
-- Each symbol being looked for is now associated with a SymbolLookupFlags
value. If the associated value is SymbolLookupFlags::RequiredSymbol then
the symbol must be defined in one of the JITDylibs being searched (or be
able to be generated in one of these JITDylibs via an attached definition
generator) or the lookup will fail with an error. If the associated value is
SymbolLookupFlags::WeaklyReferencedSymbol then the symbol is permitted to be
undefined, in which case it will simply not appear in the resulting
SymbolMap if the rest of the lookup succeeds.
Since lookup now requires these flags for each symbol, the lookup method now
takes an instance of a new SymbolLookupSet type rather than a SymbolNameSet.
SymbolLookupSet is a vector-backed set of (name, flags) pairs. Clients are
responsible for ensuring that the set property (i.e. unique elements) holds,
though this is usually simple and SymbolLookupSet provides convenience
methods to support this.
-- Lookups now have an associated LookupKind value, which is either
LookupKind::Static or LookupKind::DLSym. Definition generators can inspect
the lookup kind when determining whether or not to generate new definitions.
The StaticLibraryDefinitionGenerator is updated to only pull in new objects
from the archive if the lookup kind is Static. This allows lookup to be
re-used to emulate dlsym for JIT'd symbols without pulling in new objects
from archives (which would not happen in a normal dlsym call).
-- JITLink is updated to allow externals to be assigned weak linkage, and
weak externals now use the SymbolLookupFlags::WeaklyReferencedSymbol value
for lookups. Unresolved weak references will be assigned the default value of
zero.
Since this patch was modifying the lookup API anyway, it alo replaces all of the
"MatchNonExported" boolean arguments with a "JITDylibLookupFlags" enum for
readability. If a JITDylib's associated value is
JITDylibLookupFlags::MatchExportedSymbolsOnly then the lookup will only
match against exported (non-hidden) symbols in that JITDylib. If a JITDylib's
associated value is JITDylibLookupFlags::MatchAllSymbols then the lookup will
match against any symbol defined in the JITDylib.
Adds a DumpObjects utility that can be used to dump JIT'd objects to disk.
Instances of DebugObjects may be used by ObjectTransformLayer as no-op
transforms.
This patch also adds an ObjectTransformLayer to LLJIT and an example of how
to use this utility to dump JIT'd objects in LLJIT.
Now that we've moved to C++14, we no longer need the llvm::make_unique
implementation from STLExtras.h. This patch is a mechanical replacement
of (hopefully) all the llvm::make_unique instances across the monorepo.
llvm-svn: 369013
libObject does not apply the Exported flag to symbols in COFF object files,
which can lead to assertions when the symbol flags initially derived from
IR added to the JIT clash with the flags seen by the JIT linker. Both
RTDyldObjectLinkingLayer and ObjectLinkingLayer have a workaround for this:
they can be told to override the flags seen by the linker with the flags
attached to the materialization responsibility object that was passed down
to the linker. This patch modifies LLJIT's setup code to enable this override
by default on platforms where COFF is the default object format.
llvm-svn: 367712
ThreadSafeModule/ThreadSafeContext are used to manage lifetimes and locking
for LLVMContexts in ORCv2. Prior to this patch contexts were locked as soon
as an associated Module was emitted (to be compiled and linked), and were not
unlocked until the emit call returned. This could lead to deadlocks if
interdependent modules that shared contexts were compiled on different threads:
when, during emission of the first module, the dependence was discovered the
second module (which would provide the required symbol) could not be emitted as
the thread emitting the first module still held the lock.
This patch eliminates this possibility by moving to a finer-grained locking
scheme. Each client holds the module lock only while they are actively operating
on it. To make this finer grained locking simpler/safer to implement this patch
removes the explicit lock method, 'getContextLock', from ThreadSafeModule and
replaces it with a new method, 'withModuleDo', that implicitly locks the context,
calls a user-supplied function object to operate on the Module, then implicitly
unlocks the context before returning the result.
ThreadSafeModule TSM = getModule(...);
size_t NumFunctions = TSM.withModuleDo(
[](Module &M) { // <- context locked before entry to lambda.
return M.size();
});
Existing ORCv2 layers that operate on ThreadSafeModules are updated to use the
new method.
This method is used to introduce Module locking into each of the existing
layers.
llvm-svn: 367686
LLJITBuilder now has a setCompileFunctionCreator method which can be used to
construct a CompileFunction for the LLJIT instance being created. The motivating
use-case for this is supporting ObjectCaches, which can now be set up at
compile-function construction time. To demonstrate this an example project,
LLJITWithObjectCache, is included.
llvm-svn: 365671
LLJITBuilder and LLLazyJITBuilder construct LLJIT and LLLazyJIT instances
respectively. Over time these will allow more configurable options to be
added while remaining easy to use in the default case, which for default
in-process JITing is now:
auto J = ExitOnErr(LLJITBuilder.create());
llvm-svn: 359511
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Doesn't build on Windows. The call to 'lookup' is ambiguous. Clang and
MSVC agree, anyway.
http://lab.llvm.org:8011/builders/clang-x64-windows-msvc/builds/787
C:\b\slave\clang-x64-windows-msvc\build\llvm.src\unittests\ExecutionEngine\Orc\CoreAPIsTest.cpp(315): error C2668: 'llvm::orc::ExecutionSession::lookup': ambiguous call to overloaded function
C:\b\slave\clang-x64-windows-msvc\build\llvm.src\include\llvm/ExecutionEngine/Orc/Core.h(823): note: could be 'llvm::Expected<llvm::JITEvaluatedSymbol> llvm::orc::ExecutionSession::lookup(llvm::ArrayRef<llvm::orc::JITDylib *>,llvm::orc::SymbolStringPtr)'
C:\b\slave\clang-x64-windows-msvc\build\llvm.src\include\llvm/ExecutionEngine/Orc/Core.h(817): note: or 'llvm::Expected<llvm::JITEvaluatedSymbol> llvm::orc::ExecutionSession::lookup(const llvm::orc::JITDylibSearchList &,llvm::orc::SymbolStringPtr)'
C:\b\slave\clang-x64-windows-msvc\build\llvm.src\unittests\ExecutionEngine\Orc\CoreAPIsTest.cpp(315): note: while trying to match the argument list '(initializer list, llvm::orc::SymbolStringPtr)'
llvm-svn: 345078
In the new scheme the client passes a list of (JITDylib&, bool) pairs, rather
than a list of JITDylibs. For each JITDylib the boolean indicates whether or not
to match against non-exported symbols (true means that they should be found,
false means that they should not). The MatchNonExportedInJD and MatchNonExported
parameters on lookup are removed.
The new scheme is more flexible, and easier to understand.
This patch also updates JITDylib search orders to be lists of (JITDylib&, bool)
pairs to match the new lookup scheme. Error handling is also plumbed through
the LLJIT class to allow regression tests to fail predictably when a lookup from
a lazy call-through fails.
llvm-svn: 345077
MaterializationResponsibility.
VModuleKeys are intended to enable selective removal of modules from a JIT
session, however for a wide variety of use cases selective removal is not
needed and introduces unnecessary overhead. As of this commit, the default
constructed VModuleKey value is reserved as a "do not track" value, and
becomes the default when adding a new module to the JIT.
This commit also changes the propagation of VModuleKeys. They were passed
alongside the MaterializationResponsibity instance in XXLayer::emit methods,
but are now propagated as part of the MaterializationResponsibility instance
itself (and as part of MaterializationUnit when stored in a JITDylib).
Associating VModuleKeys with MaterializationUnits in this way should allow
for a thread-safe module removal mechanism in the future, even when a module
is in the process of being compiled, by having the
MaterializationResponsibility object check in on its VModuleKey's state
before commiting its results to the JITDylib.
llvm-svn: 344643
This commit adds a 'Legacy' prefix to old ORC layers and utilities, and removes
the '2' suffix from the new ORC layers. If you wish to continue using the old
ORC layers you will need to add a 'Legacy' prefix to your classes. If you were
already using the new ORC layers you will need to drop the '2' suffix.
The legacy layers will remain in-tree until the new layers reach feature
parity with them. This will involve adding support for removing code from the
new layers, and ensuring that performance is comperable.
llvm-svn: 344572
The new name is a better fit: This class does not actually spawn any new
threads for compilation, it is just safe to call from multiple threads
concurrently.
The "Simple" part of the name did not convey much either, so it was
dropped.
llvm-svn: 344567
This adds two arguments to the main ExecutionSession::lookup method:
MatchNonExportedInJD, and MatchNonExported. These control whether and where
hidden symbols should be matched when searching a list of JITDylibs.
A similar effect could have been achieved by filtering search results, but
this would have involved materializing symbol definitions (since materialization
is triggered on lookup) only to throw the results away, among other issues.
llvm-svn: 344467
rather than require them to have been promoted before being passed in.
Dropping this precondition is better for layer composition (CompileOnDemandLayer
was the only one that placed pre-conditions on the modules that could be added).
It also means that the promoted private symbols do not show up in the target
JITDylib's symbol table. Instead, they are confined to the hidden implementation
dylib that contains the actual definitions.
For the 403.gcc testcase this cut down the public symbol table size from ~15,000
symbols to ~4000, substantially reducing symbol dependence tracking costs.
llvm-svn: 344078
flag to true in LLJIT when running in multithreaded mode.
The IRLayer::setCloneToNewContextOnEmit method sets a flag within the IRLayer
that causes modules added to that layer to be moved to a new context (by
serializing to/from a memory buffer) when they are emitted. This allows modules
that were all loaded on the same context to be compiled in parallel.
llvm-svn: 343266
one SymbolLinkagePromoter utility.
SymbolLinkagePromoter renames anonymous and private symbols, and bumps all
linkages to at least global/hidden-visibility. Modules whose symbols have been
promoted by this utility can be decomposed into sub-modules without introducing
link errors. This is used by the CompileOnDemandLayer to extract single-function
modules for lazy compilation.
llvm-svn: 343257
Modifies lit to add a 'thread_support' feature that can be used in lit test
REQUIRES clauses. The thread_support flag is set if -DLLVM_ENABLE_THREADS=ON
and unset if -DLLVM_ENABLE_THREADS=OFF. The lit flag is used to disable the
multiple-compile-threads-basic.ll testcase when threading is disabled.
llvm-svn: 343122
This doesn't work well in builds configured with LLVM_ENABLE_THREADS=OFF,
causing the following assert when running
ExecutionEngine/OrcLazy/multiple-compile-threads-basic.ll:
lib/ExecutionEngine/Orc/Core.cpp:1748: Expected<llvm::JITEvaluatedSymbol>
llvm::orc::lookup(const llvm::orc::JITDylibList &, llvm::orc::SymbolStringPtr):
Assertion `ResultMap->size() == 1 && "Unexpected number of results"' failed.
> LLJIT and LLLazyJIT can now be constructed with an optional NumCompileThreads
> arguments. If this is non-zero then a thread-pool will be created with the
> given number of threads, and compile tasks will be dispatched to the thread
> pool.
>
> To enable testing of this feature, two new flags are added to lli:
>
> (1) -compile-threads=N (N = 0 by default) controls the number of compile threads
> to use.
>
> (2) -thread-entry can be used to execute code on additional threads. For each
> -thread-entry argument supplied (multiple are allowed) a new thread will be
> created and the given symbol called. These additional thread entry points are
> called after static constructors are run, but before main.
llvm-svn: 343099
for lazy compilation, rather than a callback manager.
The new mechanism does not block compile threads, and does not require
function bodies to be renamed.
Future modifications should allow laziness on a per-module basis to work
without any modification of the input module.
llvm-svn: 343065
This will allow trampoline pools to be re-used for a new lazy-reexport utility
that generates looks up function bodies using the standard symbol lookup process
(rather than using a user provided compile function). This new utility provides
the same capabilities (since MaterializationUnits already allow user supplied
compile functions to be run) as JITCompileCallbackManager, but can use the new
asynchronous lookup functions to avoid blocking a compile thread.
This patch also updates createLocalCompileCallbackManager to return an error if
a callback manager can not be created, and updates clients of that API to
account for the change. Finally, the OrcCBindingsStack is updates so that if
a callback manager is not available for the target platform a valid stack
(without support for lazy compilation) can still be constructed.
llvm-svn: 343059
LLJIT and LLLazyJIT can now be constructed with an optional NumCompileThreads
arguments. If this is non-zero then a thread-pool will be created with the
given number of threads, and compile tasks will be dispatched to the thread
pool.
To enable testing of this feature, two new flags are added to lli:
(1) -compile-threads=N (N = 0 by default) controls the number of compile threads
to use.
(2) -thread-entry can be used to execute code on additional threads. For each
-thread-entry argument supplied (multiple are allowed) a new thread will be
created and the given symbol called. These additional thread entry points are
called after static constructors are run, but before main.
llvm-svn: 343058
compilation of IR in the JIT.
ThreadSafeContext is a pair of an LLVMContext and a mutex that can be used to
lock that context when it needs to be accessed from multiple threads.
ThreadSafeModule is a pair of a unique_ptr<Module> and a
shared_ptr<ThreadSafeContext>. This allows the lifetime of a ThreadSafeContext
to be managed automatically in terms of the ThreadSafeModules that refer to it:
Once all modules using a ThreadSafeContext are destructed, and providing the
client has not held on to a copy of shared context pointer, the context will be
automatically destructed.
This scheme is necessary due to the following constraits: (1) We need multiple
contexts for multithreaded compilation (at least one per compile thread plus
one to store any IR not currently being compiled, though one context per module
is simpler). (2) We need to free contexts that are no longer being used so that
the JIT does not leak memory over time. (3) Module lifetimes are not
predictable (modules are compiled as needed depending on the flow of JIT'd
code) so there is no single point where contexts could be reclaimed.
JIT clients not using concurrency can safely use one ThreadSafeContext for all
ThreadSafeModules.
JIT clients who want to be able to compile concurrently should use a different
ThreadSafeContext for each module, or call setCloneToNewContextOnEmit on their
top-level IRLayer. The former reduces compile latency (since no clone step is
needed) at the cost of additional memory overhead for uncompiled modules (as
every uncompiled module will duplicate the LLVM types, constants and metadata
that have been shared).
llvm-svn: 343055
The existing memory manager API can not be shared between objects when linking
concurrently (since there is no way to know which concurrent allocations were
performed on behalf of which object, and hence which allocations would be safe
to finalize when finalizeMemory is called). For now, we can work around this by
requiring a new memory manager for each object.
This change only affects the concurrent version of the ORC APIs.
llvm-svn: 341579
The addObjectFile method adds the given object file to the JIT session, making
its code available for execution.
Support for the -extra-object flag is added to lli when operating in
-jit-kind=orc-lazy mode to support testing of this feature.
llvm-svn: 340870
VSO was a little close to VDSO (an acronym on Linux for Virtual Dynamic Shared
Object) for comfort. It also risks giving the impression that instances of this
class could be shared between ExecutionSessions, which they can not.
JITDylib seems moderately less confusing, while still hinting at how this
class is intended to be used, i.e. as a JIT-compiled stand-in for a dynamic
library (code that would have been a dynamic library if you had wanted to
compile it ahead of time).
llvm-svn: 340084
A search order is a list of VSOs to be searched linearly to find symbols. Each
VSO now has a search order that will be used when fixing up definitions in that
VSO. Each VSO's search order defaults to just that VSO itself.
This is a first step towards removing symbol resolvers from ORC altogether. In
practice symbol resolvers tended to be used to implement a search order anyway,
sometimes with additional programatic generation of symbols. Now that VSOs
support programmatic generation of definitions via fallback generators, search
orders provide a cleaner way to achieve the desired effect (while removing a lot
of boilerplate).
llvm-svn: 337593
LLJIT is a prefabricated ORC based JIT class that is meant to be the go-to
replacement for MCJIT. Unlike OrcMCJITReplacement (which will continue to be
supported) it is not API or bug-for-bug compatible, but targets the same
use cases: Simple, non-lazy compilation and execution of LLVM IR.
LLLazyJIT extends LLJIT with support for function-at-a-time lazy compilation,
similar to what was provided by LLVM's original (now long deprecated) JIT APIs.
This commit also contains some simple utility classes (CtorDtorRunner2,
LocalCXXRuntimeOverrides2, JITTargetMachineBuilder) to support LLJIT and
LLLazyJIT.
Both of these classes are works in progress. Feedback from JIT clients is very
welcome!
llvm-svn: 335670
Alexandre Ganea