Previously we stripped Weak flags from JITDylib symbol table entries once they
were resolved (there was no particularly good reason for this). Now we want to
retain them and query them when setting the Linkage on external symbols in
LinkGraphs during symbol resolution (this was the motivation for 75404e9ef8).
Making weak linkage of external definitions discoverable in the LinkGraph will
in turn allow future plugins to implement correct handling for them (by
recording locations that depend on exported weak definitions and pointing all
of these at one chosen definition at runtime).
This code previously used cantFail, but both steps (resolution and emission)
can fail if the resource tracker associated with the
AbsoluteSymbolsMaterializationUnit is removed. Checking these errors is
necessary for correct error propagation.
Idiomatic llvm::Error usage can result in a FailedToMaterialize error tearing
down an ExecutionSession instance. Since the FailedToMaterialize error holds
SymbolStringPtrs and JITDylib references this leads to crashes when accessing
or logging the error.
This patch modifies FailedToMaterialize to retain the SymbolStringPool and
JITDylibs involved in the failure so that we can safely report an error message
to the client, even if the error tears down the session.
The contract for JITDylibs allows the getName method to be used even after the
session has been torn down, but no other JITDylib fields should be accessed via
the FailedToMaterialize error if the ssesion has been torn down. Logging the
error is guaranteed to be safe in all cases.
Clients are required to call ExecutionSession::endSession before destroying the
ExecutionSession. Failure to do so can lead to memory leaks and other difficult
to debug issues. Enforcing this requirement by assertion makes it easy to spot
or debug situations where the contract was not followed.
This patch updates the MachO platform (both the ORC MachOPlatform class and the
ORC-Runtime macho_platform.* files) to use allocation actions, rather than EPC
calls, to transfer the initializer information scraped from each linked object.
Interactions between the ORC and ORC-Runtime sides of the platform are
substantially redesigned to accomodate the change.
The high-level changes in this patch are:
1. The MachOPlatform::setupJITDylib method now calls into the runtime to set up
a dylib name <-> header mapping, and a dylib state object (JITDylibState).
2. The MachOPlatformPlugin builds an allocation action that calls the
__orc_rt_macho_register_object_platform_sections and
__orc_rt_macho_deregister_object_platform_sections functions in the runtime
to register the address ranges for all "interesting" sections in the object
being allocated (TLS data sections, initializers, language runtime metadata
sections, etc.).
3. The MachOPlatform::rt_getInitializers method (the entry point in the
controller for requests from the runtime for initializer information) is
replaced by MachOPlatform::rt_pushInitializers. The former returned a data
structure containing the "interesting" section address ranges, but these are
now handled by __orc_rt_macho_register_object_platform_sections. The new
rt_pushInitializers method first issues a lookup to trigger materialization
of the "interesting" sections, then returns the dylib dependence tree rooted
at the requested dylib for dlopen to consume. (The dylib dependence tree is
returned by rt_pushInitializers, rather than being handled by some dedicated
call, because rt_pushInitializers can alter the dependence tree).
The advantage of these changes (beyond the performance advantages of using
allocation actions) is that it moves more information about the materialized
portions of the JITDylib into the executor. This tends to make the runtime
easier to reason about, e.g. the implementation of dlopen in the runtime is now
recursive, rather than relying on recursive calls in the controller to build a
linear data structure for consumption by the runtime. This change can also make
some operations more efficient, e.g. JITDylibs can be dlclosed and then
re-dlopened without having to pull all initializers over from the controller
again.
In addition to the high-level changes, there are some low-level changes to ORC
and the runtime:
* In ORC, at ExecutionSession teardown time JITDylibs are now destroyed in
reverse creation order. This is on the assumption that the ORC runtime will be
loaded into an earlier dylib that will be used by later JITDylibs. This is a
short-term solution to crashes that arose during testing when the runtime was
torn down before its users. Longer term we will likely destroy dylibs in
dependence order.
* toSPSSerializable(Expected<T> E) is updated to explicitly initialize the T
value, allowing it to be used by Ts that have explicit constructors.
* The ORC runtime now (1) attempts to track ref-counts, and (2) distinguishes
not-yet-processed "interesting" sections from previously processed ones. (1)
is necessary for standard dlopen/dlclose emulation. (2) is intended as a step
towards better REPL support -- it should enable future runtime calls that
run only newly registered initializers ("dlopen_more", "dlopen_additions",
...?).
Calls to JITDylib's getDFSLinkOrder and getReverseDFSLinkOrder methods (both
static an non-static versions) are now valid to make on defunct JITDylibs, but
will return an error if any JITDylib in the link order is defunct.
This means that platforms can safely lookup link orders by name in response to
jit-dlopen calls from the ORC runtime, even if the call names a defunct
JITDylib -- the call will just fail with an error.
This is a counterpart to Platform::setupJITDylib, and is called when JITDylib
instances are removed (via ExecutionSession::removeJITDylib).
Upcoming MachOPlatform patches will use this to clear per-JITDylib data when
JITDylibs are removed.
If all symbols in a lookup match before we reach the end of the search order
then bail out of the search-order loop early.
This should reduce unnecessary contention on the session lock and improve
readability of the debug logs.
MaterializationUnit::Interface holds the values that make up the interface
(for ORC's purposes) of a materialization unit: the symbol flags map and
initializer symbol.
Having a type for this will make functions that build materializer interfaces
more readable and maintainable.
This allows JITDylibs to be removed from the ExecutionSession. Calling
ExecutionSession::removeJITDylib will disconnect the JITDylib from the
ExecutionSession and clear it (removing all trackers associated with it). The
JITDylib object will then be destroyed as soon as the last JITDylibSP pointing
at it is destroyed.
GeneratorsMutex should prevent lookups from proceeding through the
generators of a single JITDylib concurrently (since this could
result in redundant attempts to generate definitions). Mutation of
the generators list itself should be done under the session lock.
This keeps the tracker alive for the lifetime of the MR. This is needed so that
we can check whether the tracker has become defunct before posting results (or
failure) for the MR.
The callWrapperAsync and callSPSWrapperAsync methods take a handler object
that is run on the return value of the call when it is ready. The new RunPolicy
parameters allow clients to control how these handlers are run. If no policy is
specified then the handler will be packaged as a GenericNamedTask and dispatched
using the ExecutorProcessControl's TaskDispatch member. Callers can use the
ExecutorProcessControl::RunInPlace policy to cause the handler to be run
directly instead, which may be preferrable for simple handlers, or they can
write their own policy object (e.g. to dispatch as some other kind of Task,
rather than GenericNamedTask).
ExecutorProcessControl objects will now have a TaskDispatcher member which
should be used to dispatch work (in particular, handling incoming packets in
the implementation of remote EPC implementations like SimpleRemoteEPC).
The GenericNamedTask template can be used to wrap function objects that are
callable as 'void()' (along with an optional name to describe the task).
The makeGenericNamedTask functions can be used to create GenericNamedTask
instances without having to name the function object type.
In a future patch ExecutionSession will be updated to use the
ExecutorProcessControl's dispatcher, instead of its DispatchTaskFunction.
Wrapper function call and dispatch handler helpers are moved to
ExecutionSession, and existing EPC-based tools are re-written to take an
ExecutionSession argument instead.
Requiring an ExecutorProcessControl instance simplifies existing EPC based
utilities (which only need to take an ES now), and should encourage more
utilities to use the EPC interface. It also simplifies process termination,
since the session can automatically call ExecutorProcessControl::disconnect
(previously this had to be done manually, and carefully ordered with the
rest of JIT tear-down to work correctly).
Adds support for MachO static initializers/deinitializers and eh-frame
registration via the ORC runtime.
This commit introduces cooperative support code into the ORC runtime and ORC
LLVM libraries (especially the MachOPlatform class) to support macho runtime
features for JIT'd code. This commit introduces support for static
initializers, static destructors (via cxa_atexit interposition), and eh-frame
registration. Near-future commits will add support for MachO native
thread-local variables, and language runtime registration (e.g. for Objective-C
and Swift).
The llvm-jitlink tool is updated to use the ORC runtime where available, and
regression tests for the new MachOPlatform support are added to compiler-rt.
Notable changes on the ORC runtime side:
1. The new macho_platform.h / macho_platform.cpp files contain the bulk of the
runtime-side support. This includes eh-frame registration; jit versions of
dlopen, dlsym, and dlclose; a cxa_atexit interpose to record static destructors,
and an '__orc_rt_macho_run_program' function that defines running a JIT'd MachO
program in terms of the jit- dlopen/dlsym/dlclose functions.
2. Replaces JITTargetAddress (and casting operations) with ExecutorAddress
(copied from LLVM) to improve type-safety of address management.
3. Adds serialization support for ExecutorAddress and unordered_map types to
the runtime-side Simple Packed Serialization code.
4. Adds orc-runtime regression tests to ensure that static initializers and
cxa-atexit interposes work as expected.
Notable changes on the LLVM side:
1. The MachOPlatform class is updated to:
1.1. Load the ORC runtime into the ExecutionSession.
1.2. Set up standard aliases for macho-specific runtime functions. E.g.
___cxa_atexit -> ___orc_rt_macho_cxa_atexit.
1.3. Install the MachOPlatformPlugin to scrape LinkGraphs for information
needed to support MachO features (e.g. eh-frames, mod-inits), and
communicate this information to the runtime.
1.4. Provide entry-points that the runtime can call to request initializers,
perform symbol lookup, and request deinitialiers (the latter is
implemented as an empty placeholder as macho object deinits are rarely
used).
1.5. Create a MachO header object for each JITDylib (defining the __mh_header
and __dso_handle symbols).
2. The llvm-jitlink tool (and llvm-jitlink-executor) are updated to use the
runtime when available.
3. A `lookupInitSymbolsAsync` method is added to the Platform base class. This
can be used to issue an async lookup for initializer symbols. The existing
`lookupInitSymbols` method is retained (the GenericIRPlatform code is still
using it), but is deprecated and will be removed soon.
4. JIT-dispatch support code is added to ExecutorProcessControl.
The JIT-dispatch system allows handlers in the JIT process to be associated with
'tag' symbols in the executor, and allows the executor to make remote procedure
calls back to the JIT process (via __orc_rt_jit_dispatch) using those tags.
The primary use case is ORC runtime code that needs to call bakc to handlers in
orc::Platform subclasses. E.g. __orc_rt_macho_jit_dlopen calling back to
MachOPlatform::rt_getInitializers using __orc_rt_macho_get_initializers_tag.
(The system is generic however, and could be used by non-runtime code).
The new ExecutorProcessControl::JITDispatchInfo struct provides the address
(in the executor) of the jit-dispatch function and a jit-dispatch context
object, and implementations of the dispatch function are added to
SelfExecutorProcessControl and OrcRPCExecutorProcessControl.
5. OrcRPCTPCServer is updated to support JIT-dispatch calls over ORC-RPC.
6. Serialization support for StringMap is added to the LLVM-side Simple Packed
Serialization code.
7. A JITLink::allocateBuffer operation is introduced to allocate writable memory
attached to the graph. This is used by the MachO header synthesis code, and will
be generically useful for other clients who want to create new graph content
from scratch.
Generalizing this API allows work to be distributed more evenly. In particular,
query callbacks can now be dispatched (rather than running immediately on the
thread that satisfied the query). This avoids the pathalogical case where an
operation on one thread satisfies many queries simultaneously, causing large
amounts of work to be run on that thread while other threads potentially sit
idle.
C API clients can now define a custom definition generator by providing a
callback function (to implement DefinitionGenerator::tryToGenerate) and context
object. All arguments for the DefinitionGenerator::tryToGenerate method have
been given C API counterparts, and the API allows for optionally asynchronous
generation.
This patch moves definition generation out from the session lock, instead
running it under a per-dylib generator lock. It also makes the
DefinitionGenerator::tryToGenerate method optionally asynchronous: Generators
are handed an opaque LookupState object which can be captured to stop/restart
the lookup process.
The new scheme provides the following benefits and guarantees:
(1) Queries that do not need to attempt definition generation (because all
requested symbols matched against existing definitions in the JITDylib)
can proceed without being blocked by any running definition generators.
(2) Definition generators can capture the LookupState to continue their work
asynchronously. This allows generators to run for an arbitrary amount of
time without blocking a thread. Definition generators that do not need to
run asynchronously can return without capturing the LookupState to eliminate
unnecessary recursion and improve lookup performance.
(3) Definition generators still do not need to worry about concurrency or
re-entrance: Since they are still run under a (per-dylib) lock, generators
will never be re-entered concurrently, or given overlapping symbol sets to
generate.
Finally, the new system distinguishes between symbols that are candidates for
generation (generation candidates) and symbols that failed to match for a query
(due to symbol visibility). This fixes a bug where an unresolved symbol could
trigger generation of a duplicate definition for an existing hidden symbol.
MaterializationResponsibility, JITDylib, and ExecutionSession collectively
manage the OrcV2 core JIT state. Responsibility for maintaining and
updating this state has previously been spread among these classes, resulting
in implementations that are each non-trivial, but all tightly coupled. This has
in turn made reading the code and reasoning about state update and locking
rules difficult.
The core state model can be simplified by thinking of
MaterializationResponsibility and JITDylib as facets of ExecutionSession. This
commit is the first in a series intended to refactor Core.cpp to reflect this
model. Operations on MaterializationResponsibility and JITDylib will forward to
implementation methods inside ExecutionSession. Raw state will remain with the
original classes, but in most cases will only be modified by the
ExecutionSession.
This patch introduces new APIs to support resource tracking and removal in Orc.
It is intended as a thread-safe generalization of the removeModule concept from
OrcV1.
Clients can now create ResourceTracker objects (using
JITDylib::createResourceTracker) to track resources for each MaterializationUnit
(code, data, aliases, absolute symbols, etc.) added to the JIT. Every
MaterializationUnit will be associated with a ResourceTracker, and
ResourceTrackers can be re-used for multiple MaterializationUnits. Each JITDylib
has a default ResourceTracker that will be used for MaterializationUnits added
to that JITDylib if no ResourceTracker is explicitly specified.
Two operations can be performed on ResourceTrackers: transferTo and remove. The
transferTo operation transfers tracking of the resources to a different
ResourceTracker object, allowing ResourceTrackers to be merged to reduce
administrative overhead (the source tracker is invalidated in the process). The
remove operation removes all resources associated with a ResourceTracker,
including any symbols defined by MaterializationUnits associated with the
tracker, and also invalidates the tracker. These operations are thread safe, and
should work regardless of the the state of the MaterializationUnits. In the case
of resource transfer any existing resources associated with the source tracker
will be transferred to the destination tracker, and all future resources for
those units will be automatically associated with the destination tracker. In
the case of resource removal all already-allocated resources will be
deallocated, any if any program representations associated with the tracker have
not been compiled yet they will be destroyed. If any program representations are
currently being compiled then they will be prevented from completing: their
MaterializationResponsibility will return errors on any attempt to update the
JIT state.
Clients (usually Layer writers) wishing to track resources can implement the
ResourceManager API to receive notifications when ResourceTrackers are
transferred or removed. The MaterializationResponsibility::withResourceKeyDo
method can be used to create associations between the key for a ResourceTracker
and an allocated resource in a thread-safe way.
RTDyldObjectLinkingLayer and ObjectLinkingLayer are updated to use the
ResourceManager API to enable tracking and removal of memory allocated by the
JIT linker.
The new JITDylib::clear method can be used to trigger removal of every
ResourceTracker associated with the JITDylib (note that this will only
remove resources for the JITDylib, it does not run static destructors).
This patch includes unit tests showing basic usage. A follow-up patch will
update the Kaleidoscope and BuildingAJIT tutorial series to OrcV2 and will
use this API to release code associated with anonymous expressions.
This removes all legacy layers, legacy utilities, the old Orc C bindings,
OrcMCJITReplacement, and OrcMCJITReplacement regression tests.
ExecutionEngine and MCJIT are not affected by this change.
Making MaterializationResponsibility instances immovable allows their
associated VModuleKeys to be updated by the ExecutionSession while the
responsibility is still in-flight. This will be used in the upcoming
removable code feature to enable safe merging of resource keys even if
there are active compiles using the keys being merged.
Kazu Hirata