Instead of performing a transformation, such pass yields a new pass pipeline
to run on the currently visited operation.
This feature can be used for example to implement a sub-pipeline that
would run only on an operation with specific attributes. Another example
would be to compute a cost model and dynamic schedule a pipeline based
on the result of this analysis.
Discussion: https://llvm.discourse.group/t/rfc-dynamic-pass-pipeline/1637
Recommit after fixing an ASAN issue: the callback lambda needs to be
allocated to a temporary to have its lifetime extended to the end of the
current block instead of just the current call expression.
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D86392
This reverts commit 385c3f43fceba227be2e4dce84a59075733541c1.
Test mlir/test/Pass:dynamic-pipeline-fail-on-parent.mlir.test fails
when run with ASAN:
ERROR: AddressSanitizer: stack-use-after-scope on address ...
Reviewed By: bkramer, pifon2a
Differential Revision: https://reviews.llvm.org/D88079
Instead of performing a transformation, such pass yields a new pass pipeline
to run on the currently visited operation.
This feature can be used for example to implement a sub-pipeline that
would run only on an operation with specific attributes. Another example
would be to compute a cost model and dynamic schedule a pipeline based
on the result of this analysis.
Discussion: https://llvm.discourse.group/t/rfc-dynamic-pass-pipeline/1637
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D86392
This crash only happens when a function pass is followed by a module
pass. In this case the splitting of the pass pipeline didn't handle
properly the verifier passes and ended up with an odd number of pass in
the pipeline, breaking an assumption of the local crash reproducer
executor and hitting an assertion.
Differential Revision: https://reviews.llvm.org/D88000
This is allowing to build an OpPassManager from a StringRef instead of an
Identifier, which enables building pipelines without an MLIRContext.
An identifier is still cached on-demand on the OpPassManager for efficiency
during the IR traversal.
This allows to defers the check for traits to the execution instead of forcing it on the pipeline creation.
In particular, this is making our pipeline creation tolerant to dialects not being loaded in the context yet.
Reviewed By: rriddle, GMNGeoffrey
Differential Revision: https://reviews.llvm.org/D86915
This makes OpPassManager more of a "container" of passes and not responsible to drive the execution.
As such we also make it constructible publicly, which will allow to build arbitrary pipeline decoupled from the execution. We'll make use of this facility to expose "dynamic pipeline" in the future.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D86391
This changes the behavior of constructing MLIRContext to no longer load globally
registered dialects on construction. Instead Dialects are only loaded explicitly
on demand:
- the Parser is lazily loading Dialects in the context as it encounters them
during parsing. This is the only purpose for registering dialects and not load
them in the context.
- Passes are expected to declare the dialects they will create entity from
(Operations, Attributes, or Types), and the PassManager is loading Dialects into
the Context when starting a pipeline.
This changes simplifies the configuration of the registration: a compiler only
need to load the dialect for the IR it will emit, and the optimizer is
self-contained and load the required Dialects. For example in the Toy tutorial,
the compiler only needs to load the Toy dialect in the Context, all the others
(linalg, affine, std, LLVM, ...) are automatically loaded depending on the
optimization pipeline enabled.
To adjust to this change, stop using the existing dialect registration: the
global registry will be removed soon.
1) For passes, you need to override the method:
virtual void getDependentDialects(DialectRegistry ®istry) const {}
and registery on the provided registry any dialect that this pass can produce.
Passes defined in TableGen can provide this list in the dependentDialects list
field.
2) For dialects, on construction you can register dependent dialects using the
provided MLIRContext: `context.getOrLoadDialect<DialectName>()`
This is useful if a dialect may canonicalize or have interfaces involving
another dialect.
3) For loading IR, dialect that can be in the input file must be explicitly
registered with the context. `MlirOptMain()` is taking an explicit registry for
this purpose. See how the standalone-opt.cpp example is setup:
mlir::DialectRegistry registry;
registry.insert<mlir::standalone::StandaloneDialect>();
registry.insert<mlir::StandardOpsDialect>();
Only operations from these two dialects can be in the input file. To include all
of the dialects in MLIR Core, you can populate the registry this way:
mlir::registerAllDialects(registry);
4) For `mlir-translate` callback, as well as frontend, Dialects can be loaded in
the context before emitting the IR: context.getOrLoadDialect<ToyDialect>()
Differential Revision: https://reviews.llvm.org/D85622
This changes the behavior of constructing MLIRContext to no longer load globally
registered dialects on construction. Instead Dialects are only loaded explicitly
on demand:
- the Parser is lazily loading Dialects in the context as it encounters them
during parsing. This is the only purpose for registering dialects and not load
them in the context.
- Passes are expected to declare the dialects they will create entity from
(Operations, Attributes, or Types), and the PassManager is loading Dialects into
the Context when starting a pipeline.
This changes simplifies the configuration of the registration: a compiler only
need to load the dialect for the IR it will emit, and the optimizer is
self-contained and load the required Dialects. For example in the Toy tutorial,
the compiler only needs to load the Toy dialect in the Context, all the others
(linalg, affine, std, LLVM, ...) are automatically loaded depending on the
optimization pipeline enabled.
To adjust to this change, stop using the existing dialect registration: the
global registry will be removed soon.
1) For passes, you need to override the method:
virtual void getDependentDialects(DialectRegistry ®istry) const {}
and registery on the provided registry any dialect that this pass can produce.
Passes defined in TableGen can provide this list in the dependentDialects list
field.
2) For dialects, on construction you can register dependent dialects using the
provided MLIRContext: `context.getOrLoadDialect<DialectName>()`
This is useful if a dialect may canonicalize or have interfaces involving
another dialect.
3) For loading IR, dialect that can be in the input file must be explicitly
registered with the context. `MlirOptMain()` is taking an explicit registry for
this purpose. See how the standalone-opt.cpp example is setup:
mlir::DialectRegistry registry;
registry.insert<mlir::standalone::StandaloneDialect>();
registry.insert<mlir::StandardOpsDialect>();
Only operations from these two dialects can be in the input file. To include all
of the dialects in MLIR Core, you can populate the registry this way:
mlir::registerAllDialects(registry);
4) For `mlir-translate` callback, as well as frontend, Dialects can be loaded in
the context before emitting the IR: context.getOrLoadDialect<ToyDialect>()
Differential Revision: https://reviews.llvm.org/D85622
This changes the behavior of constructing MLIRContext to no longer load globally
registered dialects on construction. Instead Dialects are only loaded explicitly
on demand:
- the Parser is lazily loading Dialects in the context as it encounters them
during parsing. This is the only purpose for registering dialects and not load
them in the context.
- Passes are expected to declare the dialects they will create entity from
(Operations, Attributes, or Types), and the PassManager is loading Dialects into
the Context when starting a pipeline.
This changes simplifies the configuration of the registration: a compiler only
need to load the dialect for the IR it will emit, and the optimizer is
self-contained and load the required Dialects. For example in the Toy tutorial,
the compiler only needs to load the Toy dialect in the Context, all the others
(linalg, affine, std, LLVM, ...) are automatically loaded depending on the
optimization pipeline enabled.
To adjust to this change, stop using the existing dialect registration: the
global registry will be removed soon.
1) For passes, you need to override the method:
virtual void getDependentDialects(DialectRegistry ®istry) const {}
and registery on the provided registry any dialect that this pass can produce.
Passes defined in TableGen can provide this list in the dependentDialects list
field.
2) For dialects, on construction you can register dependent dialects using the
provided MLIRContext: `context.getOrLoadDialect<DialectName>()`
This is useful if a dialect may canonicalize or have interfaces involving
another dialect.
3) For loading IR, dialect that can be in the input file must be explicitly
registered with the context. `MlirOptMain()` is taking an explicit registry for
this purpose. See how the standalone-opt.cpp example is setup:
mlir::DialectRegistry registry;
mlir::registerDialect<mlir::standalone::StandaloneDialect>();
mlir::registerDialect<mlir::StandardOpsDialect>();
Only operations from these two dialects can be in the input file. To include all
of the dialects in MLIR Core, you can populate the registry this way:
mlir::registerAllDialects(registry);
4) For `mlir-translate` callback, as well as frontend, Dialects can be loaded in
the context before emitting the IR: context.getOrLoadDialect<ToyDialect>()
This changes the behavior of constructing MLIRContext to no longer load globally registered dialects on construction. Instead Dialects are only loaded explicitly on demand:
- the Parser is lazily loading Dialects in the context as it encounters them during parsing. This is the only purpose for registering dialects and not load them in the context.
- Passes are expected to declare the dialects they will create entity from (Operations, Attributes, or Types), and the PassManager is loading Dialects into the Context when starting a pipeline.
This changes simplifies the configuration of the registration: a compiler only need to load the dialect for the IR it will emit, and the optimizer is self-contained and load the required Dialects. For example in the Toy tutorial, the compiler only needs to load the Toy dialect in the Context, all the others (linalg, affine, std, LLVM, ...) are automatically loaded depending on the optimization pipeline enabled.
Differential Revision: https://reviews.llvm.org/D85622
This changes the behavior of constructing MLIRContext to no longer load globally registered dialects on construction. Instead Dialects are only loaded explicitly on demand:
- the Parser is lazily loading Dialects in the context as it encounters them during parsing. This is the only purpose for registering dialects and not load them in the context.
- Passes are expected to declare the dialects they will create entity from (Operations, Attributes, or Types), and the PassManager is loading Dialects into the Context when starting a pipeline.
This changes simplifies the configuration of the registration: a compiler only need to load the dialect for the IR it will emit, and the optimizer is self-contained and load the required Dialects. For example in the Toy tutorial, the compiler only needs to load the Toy dialect in the Context, all the others (linalg, affine, std, LLVM, ...) are automatically loaded depending on the optimization pipeline enabled.
Essentially takes the lld/Common/Threads.h wrappers and moves them to
the llvm/Support/Paralle.h algorithm header.
The changes are:
- Remove policy parameter, since all clients use `par`.
- Rename the methods to `parallelSort` etc to match LLVM style, since
they are no longer C++17 pstl compatible.
- Move algorithms from llvm::parallel:: to llvm::, since they have
"parallel" in the name and are no longer overloads of the regular
algorithms.
- Add range overloads
- Use the sequential algorithm directly when 1 thread is requested
(skips task grouping)
- Fix the index type of parallelForEachN to size_t. Nobody in LLVM was
using any other parameter, and it made overload resolution hard for
for_each_n(par, 0, foo.size(), ...) because 0 is int, not size_t.
Remove Threads.h and update LLD for that.
This is a prerequisite for parallel public symbol processing in the PDB
library, which is in LLVM.
Reviewed By: MaskRay, aganea
Differential Revision: https://reviews.llvm.org/D79390
This is useful for several reasons:
* In some situations the user can guarantee that thread-safety isn't necessary and don't want to pay the cost of synchronization, e.g., when parsing a very large module.
* For things like logging threading is not desirable as the output is not guaranteed to be in stable order.
This flag also subsumes the pass manager flag for multi-threading.
Differential Revision: https://reviews.llvm.org/D79266
These libraries are distinct from other things in Analysis in that they
operate only on core IR concepts. This also simplifies dependencies
so that Dialect -> Analysis -> Parser -> IR. Previously, the parser depended
on portions of the the Analysis directory as well, which sometimes
caused issues with the way the cmake makefile generator discovers
dependencies on generated files during compilation.
Differential Revision: https://reviews.llvm.org/D79240
The current implementation uses CrashRecoveryContext, but this only supports recovering in a certain number of cases. This revision adds a signal handler to support even more situations.
This revision was able to properly generate a reproducer for a segfault in the Inliner, that the current recovery couldn't.
Differential Revision: https://reviews.llvm.org/D78315
This revision adds a mode to the crash reproducer generator to attempt to generate a more local reproducer. This will attempt to generate a reproducer right before the offending pass that fails. This is useful for the majority of failures that are specific to a single pass, and situations where some passes in the pipeline are not registered with a specific tool.
Differential Revision: https://reviews.llvm.org/D78314
This moves the threading check to runOnOperation. This produces a much cleaner interface for the adaptor pass, and will allow for the ability to enable/disable threading in a much cleaner way in the future.
Differential Revision: https://reviews.llvm.org/D78313
These have proved incredibly useful for interleaving values between a range w.r.t to streams. After this revision, the mlir/Support/STLExtras.h is empty. A followup revision will remove it from the tree.
Differential Revision: https://reviews.llvm.org/D78067
Summary: ClassID is a bit janky right now as it involves passing a magic pointer around. This revision hides the internal implementation mechanism within a new class TypeID. This class is a value-typed wrapper around the original ClassID implementation.
Differential Revision: https://reviews.llvm.org/D77768
Summary: This hook allows for passes to specify the command line argument without the need for registration. More concretely this will allow for generating pass crash reproducers without needing to have the passes registered. This should remove the need for production tools to register passes, leaving that solely to development tools like mlir-opt.
Differential Revision: https://reviews.llvm.org/D77907
This is avoid the user to shoot themselves in the foot and encounter
strange crashes that are confusing until one run with TSAN.
Differential Revision: https://reviews.llvm.org/D75399
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 is how it should've been and brings it more in line with
std::string_view. There should be no functional change here.
This is mostly mechanical from a custom clang-tidy check, with a lot of
manual fixups. It uncovers a lot of minor inefficiencies.
This doesn't actually modify StringRef yet, I'll do that in a follow-up.
This change refactors pass options to be more similar to how statistics are modeled. More specifically, the options are specified directly on the pass instead of in a separate options class. (Note that the behavior and specification for pass pipelines remains the same.) This brings about several benefits:
* The specification of options is much simpler
* The round-trip format of a pass can be generated automatically
* This gives a somewhat deeper integration with "configuring" a pass, which we could potentially expose to users in the future.
PiperOrigin-RevId: 286953824
Statistics are a way to keep track of what the compiler is doing and how effective various optimizations are. It is useful to see what optimizations are contributing to making a particular program run faster. Pass-instance specific statistics take this even further as you can see the effect of placing a particular pass at specific places within the pass pipeline, e.g. they could help answer questions like "what happens if I run CSE again here".
Statistics can be added to a pass by simply adding members of type 'Pass::Statistics'. This class takes as a constructor arguments: the parent pass pointer, a name, and a description. Statistics can be dumped by the pass manager in a similar manner to how pass timing information is dumped, i.e. via PassManager::enableStatistics programmatically; or -pass-statistics and -pass-statistics-display via the command line pass manager options.
Below is an example:
struct MyPass : public OperationPass<MyPass> {
Statistic testStat{this, "testStat", "A test statistic"};
void runOnOperation() {
...
++testStat;
...
}
};
$ mlir-opt -pass-pipeline='func(my-pass,my-pass)' foo.mlir -pass-statistics
Pipeline Display:
===-------------------------------------------------------------------------===
... Pass statistics report ...
===-------------------------------------------------------------------------===
'func' Pipeline
MyPass
(S) 15 testStat - A test statistic
MyPass
(S) 6 testStat - A test statistic
List Display:
===-------------------------------------------------------------------------===
... Pass statistics report ...
===-------------------------------------------------------------------------===
MyPass
(S) 21 testStat - A test statistic
PiperOrigin-RevId: 284022014
This cl adds support for generating a .mlir file containing a reproducer for crashes and failures that happen during pass execution. The reproducer contains a comment detailing the configuration of the pass manager(e.g. the textual description of the pass pipeline that the pass manager was executing), along with the original input module.
Example Output:
// configuration: -pass-pipeline='func(cse, canonicalize), inline'
// note: verifyPasses=false
module {
...
}
PiperOrigin-RevId: 274088134
Allow printing out pipelines in a format that is as close as possible to the
textual pass pipeline format. Individual passes can override the print function
in order to format any options that may have been used to construct that pass.
PiperOrigin-RevId: 273813627
For the cases where there are multiple levels of nested pass managers, the parent thread ID is not enough to distinguish the parent of a given pass pipeline. Passing in the parent pass gives an exact anchor point.
PiperOrigin-RevId: 272105461
OperationPass' are defined exactly the same way as they are now:
class DerivedPass : public OperationPass<DerivedPass>;
OpPass' are now defined as OperationPass, but with an additional template parameter for the operation type:
class DerivedPass : public OperationPass<DerivedPass, FuncOp>;
PiperOrigin-RevId: 269122410
This allows for explicitly specifying the pipeline to add to the pass manager. This includes the nesting structure, as well as the passes/pipelines to run. A textual pipeline string is defined as a series of names, each of which may in itself recursively contain a nested pipeline description. A name is either the name of a registered pass, or pass pipeline, (e.g. "cse") or the name of an operation type (e.g. "func").
For example, the following pipeline:
$ mlir-opt foo.mlir -cse -canonicalize -lower-to-llvm
Could now be specified as:
$ mlir-opt foo.mlir -pass-pipeline='func(cse, canonicalize), lower-to-llvm'
This will allow for running pipelines on nested operations, like say spirv modules. This does not remove any of the current functionality, and in fact can be used in unison. The new option is available via 'pass-pipeline'.
PiperOrigin-RevId: 268954279
Some compilers will try to auto-generate the destructor, instead of using the user provided destructor, when creating a default move constructor.
PiperOrigin-RevId: 268067367
This allows for parallelizing across pipelines of multiple operation types. AdaptorPasses can now hold pass managers for multiple operation types and will dispatch based upon the operation being operated on.
PiperOrigin-RevId: 268017344
This is done via a new set of instrumentation hooks runBeforePipeline/runAfterPipeline, that signal the lifetime of a pass pipeline on a specific operation type. These hooks also provide the parent thread of the pipeline, allowing for accurate merging of timers running on different threads.
PiperOrigin-RevId: 267909193
