Updates the object buffer ownership scheme in jitLinkForOrc and related
functions: Ownership of both the object::ObjectFile and underlying
MemoryBuffer is passed into jitLinkForOrc and passed back to the onEmit
callback once linking is complete. This avoids the use-after-free errors
that were seen in 98f2bb44610.
Along the same lines as eb918d8daf1: This code also had to acquire the session
mutex, and this could cause a deadlock under the wrong circumstances. This
patch updates GenericLLVMIRPlatformSupport to just use the session lock for
everything.
In MachOPlatform, obtaining the link-order for a JITDylib requires locking the
session, but also needs to be part of a larger atomic operation that collates
initializer symbols tracked by the platform. Trying to do this under a separate
platform mutex leads to potential locking order issues, e.g.
T1 locks session then tries to lock platform to register a new init symbol
meanwhile
T2 locks platform then tries to lock session to obtain link order.
Removing the platform lock and performing all these operations under the session
lock eliminates this possibility.
At the same time we also need to collate init pointers from the
MachOPlatform::InitScraperPlugin, and we don't need or want to lock the session
for that. The new InitSeqMutex has been added to guard these init pointers, and
the session mutex is never obtained while the InitSeqMutex is held.
The MU may define no symbols, but still contain a non-trivial destructor (e.g.
an LLVM IR module that has been stripped of all externally visible
definitions, but which still needs to lock its context to be destroyed).
Bailing out early ensures that we destroy the unit outside the session lock,
rather than under it which may cause deadlocks.
Also adds some extra sanity-checking assertions.
Follow-up for D74433
What the function returns are almost standard BFD names, except that "ELF" is
in uppercase instead of lowercase.
This patch changes "ELF" to "elf" and changes ARM/AArch64 to use their BFD names.
MIPS and PPC64 have endianness differences as well, but this patch does not intend to address them.
Advantages:
* llvm-objdump: the "file format " line matches GNU objdump on ARM/AArch64 objects
* "file format " line can be extracted and fed into llvm-objcopy -O literally.
(https://github.com/ClangBuiltLinux/linux/issues/779 has such a use case)
Affected tools: llvm-readobj, llvm-objdump, llvm-dwarfdump, MCJIT (internal implementation detail, not exposed)
Reviewed By: jhenderson
Differential Revision: https://reviews.llvm.org/D76046
Enable use of ExecutionEngine JITEventListeners in RTDyldObjectLinkingLayer.
This allows existing MCJIT clients to more easily migrate to LLJIT / ORCv2.
Example usage in llvm/examples/OrcV2Examples/LLJITWithGDBRegistrationListener.
Differential Revision: https://reviews.llvm.org/D75838
Global symbols with linker-private prefixes should be resolvable across object
boundaries, but internal symbols with linker-private prefixes should not.
Renames the llvm/examples/LLJITExamples directory to llvm/examples/OrcV2Examples
since it is becoming a home for all OrcV2 examples, not just LLJIT.
See http://llvm.org/PR31103.
Patch based on https://reviews.llvm.org/D75912 by Alexander Shishkin. Thanks
Alexander!
To minimize disruption to existing clients, who may be relying on the fact that
unused references to unresolved symbols do not generate an error, this patch
makes error checking opt-in: Clients can call ExecutionEngine::hasError or
LLVMExecutionEngineGetError to check whether and error has occurred.
Differential revision: https://reviews.llvm.org/D75912
This patch enables exception handling in code added to LLJIT on Darwin by
adding an orc::EHFrameRegistrationPlugin instance to the ObjectLinkingLayer
(which is currently used on Darwin only).
These may be accessed from multiple threads if concurrent materialization is
enabled in ORC.
Testcase coming in a follow-up patch that enables eh-frame registration for
LLJIT.
Summary:
Enables JIT-linking by RuntimeDyld of COFF objects that contain references to
dllimport symbols. This is done by recognizing symbols that start with the
reserved "__imp_" prefix and building a pointer entry to the target symbol in
the stubs area of the section. References to the "__imp_" symbol are updated to
point to this pointer.
Work in progress: The generic code is in place, but only RuntimeDyldCOFFX86_64
and RuntimeDyldCOFFI386 have been updated to look for and update references to
dllimport symbols.
Reviewers: compnerd
Subscribers: hiraditya, ributzka, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75884
This patch allows rtdyld-check / jitlink-check expressions to be extended over
multiple lines by terminating each line with a '\'. E.g.
# llvm-rtdyld: *{8}X = \
# llvm-rtdyld: Y
X:
.quad Y
This will be used to break up some long lines in upcoming test cases.
The LLJIT::MachOPlatformSupport class used to unconditionally attempt to
register __objc_selrefs and __objc_classlist sections. If libobjc had not
been loaded this resulted in an assertion, even if no objc sections were
actually present. This patch replaces this unconditional registration with
a check that no objce sections are present if libobjc has not been loaded.
This will allow clients to use MachOPlatform with LLJIT without requiring
libobjc for non-objc code.
Summary: Decompose callThroughToSymbol() into findReexport(), resolveSymbol(), notifyResolved() and reportCallThroughError(). This allows derived classes to reuse the functionality while adding their own code in between.
Reviewers: lhames
Reviewed By: lhames
Subscribers: hiraditya, steven_wu, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75084
ST_File symbols aren't relevant for linking purposes, but can end up shadowing
real symbols if they're not filtered.
No test case yet: The ideal testcase for this would be an ELF llvm-jitlink test,
but llvm-jitlink support for ELF is still under development. We should add a
testcase for this once support lands in tree.
reinterpret_cast'ing the block base address directly to a uint64_t leaves the
high bits in an implementation-defined state, but JITLink expects them to be
zero. Switching to pointerToJITTargetAddress for the cast should fix this.
This should fix the jitlink test failures that we have seen on some of the
32-bit testers.
Lots of headers pass around MemoryBuffer objects, but very few open
them. Let those that do include FileSystem.h.
Saves ~250 includes of Chrono.h & FileSystem.h:
$ diff -u thedeps-before.txt thedeps-after.txt | grep '^[-+] ' | sort | uniq -c | sort -nr
254 - ../llvm/include/llvm/Support/FileSystem.h
253 - ../llvm/include/llvm/Support/Chrono.h
237 - ../llvm/include/llvm/Support/NativeFormatting.h
237 - ../llvm/include/llvm/Support/FormatProviders.h
192 - ../llvm/include/llvm/ADT/StringSwitch.h
190 - ../llvm/include/llvm/Support/FormatVariadicDetails.h
...
This requires duplicating the file_t typedef, which is unfortunate. I
sunk the choice of mapping mode down into the cpp file using variable
template specializations instead of class members in headers.
Summary: A function that creates JITSymbolFlags from a GlobalValueSummary. Similar functions exist: fromGlobalValue(), fromObjectSymbol()
Reviewers: lhames
Reviewed By: lhames
Subscribers: hiraditya, steven_wu, dexonsmith, arphaman, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75082
This optimization bypasses GOT loads and calls/branches through stubs when the
ultimate target of the access/branch is found to be within range of the
reference.
Extra debugging output is also added to the generic JITLink algorithm and
basic GOT and Stubs builder utility to aid debugging.
The GenericLLVMIRPlatformSupport class runs a transform on all LLVM IR added to
the LLJIT instance to replace instances of llvm.global_ctors with a specially
named function that runs the corresponing static initializers (See
(GlobalCtorDtorScraper from lib/ExecutionEngine/Orc/LLJIT.cpp). This patch
updates the GenericIRPlatform class to check for this specially named function
in other materialization units that are added to the JIT and, if found, add
the function to the initializer work queue. Doing this allows object files
that were compiled from IR and cached to be reloaded in subsequent JIT sessions
without their initializers being skipped.
To enable testing this patch also updates the lli tool's -jit-kind=orc-lazy mode
to respect the -enable-cache-manager and -object-cache-dir options, and modifies
the CompileOnDemandLayer to rename extracted submodules to include a hash of the
names of their symbol definitions. This allows a simple object caching scheme
based on module names (which was already implemented in lli) to work with the
lazy JIT.
This patch adds new errors and error checking to the ObjectLinkingLayer to
catch cases where a compiled or loaded object either:
(1) Contains definitions not covered by its responsibility set, or
(2) Is missing definitions that are covered by its responsibility set.
Proir to this patch providing the correct set of definitions was treated as
an API contract requirement, however this requires that the client be confident
in the correctness of the whole compiler / object-cache pipeline and results
in difficult-to-debug assertions upon failure. Treating this as a recoverable
error results in clearer diagnostics.
The performance overhead of this check is one comparison of densemap keys
(symbol string pointers) per linking object, which is minimal. If this overhead
ever becomes a problem we can add the check under a flag that can be turned off
if the client fully trusts the rest of the pipeline.
Initializers and deinitializers are used to implement C++ static constructors
and destructors, runtime registration for some languages (e.g. with the
Objective-C runtime for Objective-C/C++ code) and other tasks that would
typically be performed when a shared-object/dylib is loaded or unloaded by a
statically compiled program.
MCJIT and ORC have historically provided limited support for discovering and
running initializers/deinitializers by scanning the llvm.global_ctors and
llvm.global_dtors variables and recording the functions to be run. This approach
suffers from several drawbacks: (1) It only works for IR inputs, not for object
files (including cached JIT'd objects). (2) It only works for initializers
described by llvm.global_ctors and llvm.global_dtors, however not all
initializers are described in this way (Objective-C, for example, describes
initializers via specially named metadata sections). (3) To make the
initializer/deinitializer functions described by llvm.global_ctors and
llvm.global_dtors searchable they must be promoted to extern linkage, polluting
the JIT symbol table (extra care must be taken to ensure this promotion does
not result in symbol name clashes).
This patch introduces several interdependent changes to ORCv2 to support the
construction of new initialization schemes, and includes an implementation of a
backwards-compatible llvm.global_ctor/llvm.global_dtor scanning scheme, and a
MachO specific scheme that handles Objective-C runtime registration (if the
Objective-C runtime is available) enabling execution of LLVM IR compiled from
Objective-C and Swift.
The major changes included in this patch are:
(1) The MaterializationUnit and MaterializationResponsibility classes are
extended to describe an optional "initializer" symbol for the module (see the
getInitializerSymbol method on each class). The presence or absence of this
symbol indicates whether the module contains any initializers or
deinitializers. The initializer symbol otherwise behaves like any other:
searching for it triggers materialization.
(2) A new Platform interface is introduced in llvm/ExecutionEngine/Orc/Core.h
which provides the following callback interface:
- Error setupJITDylib(JITDylib &JD): Can be used to install standard symbols
in JITDylibs upon creation. E.g. __dso_handle.
- Error notifyAdding(JITDylib &JD, const MaterializationUnit &MU): Generally
used to record initializer symbols.
- Error notifyRemoving(JITDylib &JD, VModuleKey K): Used to notify a platform
that a module is being removed.
Platform implementations can use these callbacks to track outstanding
initializers and implement a platform-specific approach for executing them. For
example, the MachOPlatform installs a plugin in the JIT linker to scan for both
__mod_inits sections (for C++ static constructors) and ObjC metadata sections.
If discovered, these are processed in the usual platform order: Objective-C
registration is carried out first, then static initializers are executed,
ensuring that calls to Objective-C from static initializers will be safe.
This patch updates LLJIT to use the new scheme for initialization. Two
LLJIT::PlatformSupport classes are implemented: A GenericIR platform and a MachO
platform. The GenericIR platform implements a modified version of the previous
llvm.global-ctor scraping scheme to provide support for Windows and
Linux. LLJIT's MachO platform uses the MachOPlatform class to provide MachO
specific initialization as described above.
Reviewers: sgraenitz, dblaikie
Subscribers: mgorny, hiraditya, mgrang, ributzka, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74300
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
ObjectLinkingLayer was not correctly propagating dependencies through local
symbols within an object. This could cause symbol lookup to return before a
searched-for symbol is ready if the following conditions are met:
(1) The definition of the symbol being searched for transitively depends on a
local symbol within the same object, and that local symbol in turn
transitively depends on an external symbol provided by some other module
in the JIT.
(2) Concurrent compilation is enabled.
(3) Thread scheduling causes the lookup of the searched-for symbol to return
before all transitive dependencies of the looked-up symbol are emitted.
This bug was found by inspection and has not been observed in practice.
A jitlink test case has been added to verify that symbol dependencies are
correctly propagated through local symbol definitions.