Path: blob/main/docs/source/user-guide/binding/optimization-passes.rst
1154 views
===================
Optimization passes
===================
.. currentmodule:: llvmlite.binding
LLVM gives you the opportunity to fine-tune optimization passes. Optimization
passes are managed by a pass manager. There are two kinds of pass managers:
* :class:`FunctionPassManager`, for optimizations that work on
single functions.
* :class:`ModulePassManager`, for optimizations that work on
whole modules.
llvmlite provides bindings for LLVM's *New Pass Manager*. Prior to llvmlite 0.45
(LLVM 20), llvmlite also supported the *Legacy Pass Manager*, which had a
different API and behavior. The differences between them and the motivations
for the New Pass Manager are outlined in the `LLVM Blog post on the New Pass Manager
<https://blog.llvm.org/posts/2021-03-26-the-new-pass-manager/>`_.
As of llvmlite 0.45 (LLVM 20), support for the Legacy Pass Manager has been
removed. All code should now use the New Pass Manager APIs documented below.
For users migrating from the legacy API, see the :ref:`passes-migration-guide`
later in this document.
New Pass Manager APIs
=====================
To manage the optimization attributes we first need to instantiate a
:class:`PipelineTuningOptions` instance:
.. function:: create_pipeline_tuning_options(speed_level=2, size_level=0)
Create a :class:`PipelineTuningOptions` instance.
.. class:: PipelineTuningOptions(speed_level=2, size_level=0)
Creates a new PipelineTuningOptions object.
The following writable attributes are available, whose default values depend
on the initial setting of the speed and size optimization levels:
* .. attribute:: loop_interleaving
Enable loop interleaving.
* .. attribute:: loop_vectorization
Enable loop vectorization.
* .. attribute:: slp_vectorization
Enable SLP vectorization, which uses a different algorithm to
loop vectorization. Both may be enabled at the same time.
* .. attribute:: loop_unrolling
Enable loop unrolling.
* .. attribute:: speed_level
The level of optimization for speed, as an integer between 0 and 3.
* .. attribute:: size_level
The level of optimization for size, as an integer between 0 and 2.
* .. attribute:: inlining_threshold
The integer threshold for inlining one function into
another. The higher the number, the more likely that
inlining will occur. This attribute is write-only.
We also need a :class:`PassBuilder` object to manage the respective function
and module pass managers:
.. function:: create_pass_builder(tm, pto)
Create a :class:`PassBuilder` instance that uses the given :class:`TargetMachine`
(*tm*) and :class:`PipelineTuningOptions` (*pto*) instances.
.. class:: PassBuilder(target_machine, pipeline_tuning_options)
A pass builder that uses the given :class:`TargetMachine` and
:class:`PipelineTuningOptions` instances.
.. method:: getModulePassManager()
Return a populated :class:`ModulePassManager` object based on PTO settings.
.. method:: getFunctionPassManager()
Return a populated :class:`FunctionPassManager` object based on PTO
settings.
.. method:: start_pass_timing()
Enable the pass timers.
.. method:: finish_pass_timing()
Returns a string containing the LLVM-generated timing report and disables
the pass timers.
The :class:`ModulePassManager` and :class:`FunctionPassManager` classes
implement the module and function pass managers:
.. class:: ModulePassManager()
A pass manager for running optimization passes on an LLVM module.
.. method:: add_verifier()
Add the `Module Verifier
<https://llvm.org/docs/Passes.html#verify-module-verifier>`_ pass.
.. method:: run(module, passbuilder)
Run optimization passes on *module*, a :class:`ModuleRef` instance.
.. class:: FunctionPassManager()
A pass manager for running optimization passes on an LLVM function.
.. method:: run(function, passbuilder)
Run optimization passes on *function*, a :class:`ValueRef` instance.
These can be created with passes populated by using the
:meth:`PassBuilder.getModulePassManager` and
:meth:`PassBuilder.getFunctionPassManager` methods, or they can be instantiated
unpopulated, then passes can be added using the ``add_*`` methods.
To instantiate the unpopulated instances, use:
.. function:: create_new_module_pass_manager()
Create an unpopulated :class:`ModulePassManager` instance.
and
.. function:: create_new_function_pass_manager()
Create an unpopulated :class:`FunctionPassManager` instance.
The ``add_*`` methods supported by both pass manager classes are:
.. currentmodule:: None
.. method:: add_aa_eval_pass()
Add the `Exhaustive Alias Analysis Precision Evaluator
<https://llvm.org/docs/Passes.html#aa-eval-exhaustive-alias-analysis-precision-evaluator>`_
pass.
.. method:: add_loop_unroll_pass()
Add the `Loop Unroll
<https://llvm.org/docs/Passes.html#loop-unroll-unroll-loops>`_ pass.
.. method:: add_loop_rotate_pass()
Add the `Loop Rotate
<https://llvm.org/docs/Passes.html#loop-rotate-rotate-loops>`_ pass.
.. method:: add_instruction_combine_pass()
Add the `Combine Redundant Instructions
<https://llvm.org/docs/Passes.html#instcombine-combine-redundant-instructions>`_
pass.
.. method:: add_jump_threading_pass()
Add the `Jump Threading
<https://llvm.org/docs/Passes.html#jump-threading-jump-threading>`_ pass.
.. method:: add_simplify_cfg_pass()
Add the `Simplify CFG
<https://llvm.org/docs/Passes.html#simplifycfg-simplify-the-cfg>`_ pass.
.. method:: add_refprune_pass()
Add the `Reference pruning
<https://github.com/numba/llvmlite/blob/main/ffi/custom_passes.cpp>`_ pass.
.. currentmodule:: llvmlite.binding
Example Usage of New API
------------------------
Here's a complete example showing how to use the New Pass Manager APIs:
.. literalinclude:: ../../../../examples/newpassmanagers.py
:caption: examples/newpassmanagers.py
.. _passes-migration-guide:
Legacy API Migration Guide
==========================
As of llvmlite 0.45 (LLVM 20), the legacy pass manager API has been removed.
This section provides a migration guide comparing the New Pass Manager API
with the removed legacy API to help users understand the differences and
migrate existing code:
.. list-table:: Pass Manager API Comparison
:header-rows: 1
:widths: 50 50
* - **Legacy Pass Manager**
- **New Pass Manager**
* - .. code-block:: python
# Setup
pmb = PassManagerBuilder()
pmb.opt_level = 2
pmb.size_level = 0
pmb.loop_vectorize = True
- .. code-block:: python
# Setup
pto = create_pipeline_tuning_options(
speed_level=2, size_level=0)
pto.loop_vectorization = True
pass_builder = create_pass_builder(
target_machine, pto)
* - .. code-block:: python
# Module optimization
mpm = ModulePassManager()
pmb.populate(mpm)
mpm.run(module)
- .. code-block:: python
# Module optimization
mpm = pass_builder.getModulePassManager()
mpm.run(module, pass_builder)
* - .. code-block:: python
# Function optimization
fpm = FunctionPassManager(module)
pmb.populate(fpm)
fpm.initialize()
fpm.run(function)
fpm.finalize()
- .. code-block:: python
# Function optimization
fpm = pass_builder.getFunctionPassManager()
fpm.run(function, pass_builder)
Legacy Pass Manager APIs (Removed)
==================================
.. warning::
The Legacy Pass Manager API has been removed as of llvmlite 0.45 (LLVM 20).
This documentation is kept for reference purposes only. New code should use
the New Pass Manager API documented above.
The legacy API required creating and configuring a :class:`PassManagerBuilder`
to instantiate pass managers. This approach has been superseded by the factory
functions and direct class instantiation provided by the New Pass Manager.
.. class:: PassManagerBuilder()
Create a new pass manager builder. This object centralizes
optimization settings.
The ``populate`` method is available:
.. method:: populate(pm)
Populate the pass manager *pm* with the optimization passes
configured in this pass manager builder.
The following writable attributes are available:
* .. attribute:: disable_unroll_loops
If ``True``, disable loop unrolling.
* .. attribute:: inlining_threshold
The integer threshold for inlining one function into
another. The higher the number, the more likely that
inlining will occur. This attribute is write-only.
* .. attribute:: loop_vectorize
If ``True``, allow vectorizing loops.
* .. attribute:: opt_level
The general optimization level, as an integer between 0
and 3.
* .. attribute:: size_level
Whether and how much to optimize for size, as an integer
between 0 and 2.
* .. attribute:: slp_vectorize
If ``True``, enable the SLP vectorizer, which uses a
different algorithm than the loop vectorizer. Both may
be enabled at the same time.
.. class:: PassManager
The base class for pass managers. Use individual ``add_*``
methods or :meth:`PassManagerBuilder.populate` to add
optimization passes.
* .. function:: add_constant_merge_pass()
See `constmerge pass documentation <http://llvm.org/docs/Passes.html#constmerge-merge-duplicate-global-constants>`_.
* .. function:: add_dead_arg_elimination_pass()
See `deadargelim pass documentation <http://llvm.org/docs/Passes.html#deadargelim-dead-argument-elimination>`_.
* .. function:: add_function_attrs_pass()
See `functionattrs pass documentation <http://llvm.org/docs/Passes.html#functionattrs-deduce-function-attributes>`_.
* .. function:: add_function_inlining_pass(self, )
See `inline pass documentation <http://llvm.org/docs/Passes.html#inline-function-integration-inlining>`_.
* .. function:: add_global_dce_pass()
See `globaldce pass documentation <http://llvm.org/docs/Passes.html#globaldce-dead-global-elimination>`_.
* .. function:: add_global_optimizer_pass()
See `globalopt pass documentation <http://llvm.org/docs/Passes.html#globalopt-global-variable-optimizer>`_.
* .. function:: add_ipsccp_pass()
See `ipsccp pass documentation <http://llvm.org/docs/Passes.html#ipsccp-interprocedural-sparse-conditional-constant-propagation>`_.
* .. function:: add_dead_code_elimination_pass()
See `dce pass documentation <http://llvm.org/docs/Passes.html#dce-dead-code-elimination>`_.
* .. function:: add_cfg_simplification_pass()
See `simplifycfg pass documentation <http://llvm.org/docs/Passes.html#simplifycfg-simplify-the-cfg>`_.
* .. function:: add_gvn_pass()
See `gvn pass documentation <http://llvm.org/docs/Passes.html#gvn-global-value-numbering>`_.
* .. function:: add_instruction_combining_pass()
See `instcombine pass documentation <http://llvm.org/docs/Passes.html#passes-instcombine>`_.
* .. function:: add_licm_pass()
See `licm pass documentation <http://llvm.org/docs/Passes.html#licm-loop-invariant-code-motion>`_.
* .. function:: add_sccp_pass()
See `sccp pass documentation <http://llvm.org/docs/Passes.html#sccp-sparse-conditional-constant-propagation>`_.
* .. function:: add_sroa_pass()
See `scalarrepl pass documentation <http://llvm.org/docs/Passes.html#scalarrepl-scalar-replacement-of-aggregates>`_.
While the scalarrepl pass documentation describes the
transformation performed by the pass added by this
function, the pass corresponds to the ``opt -sroa``
command-line option and not to ``opt -scalarrepl``.
* .. function:: add_type_based_alias_analysis_pass()
See `tbaa metadata documentation <http://llvm.org/docs/LangRef.html#tbaa-metadata>`_.
* .. function:: add_basic_alias_analysis_pass()
See `basicaa pass documentation <http://llvm.org/docs/AliasAnalysis.html#the-basicaa-pass>`_.
* .. function:: add_instruction_namer_pass()
See `instnamer pass documentation <http://llvm.org/docs/Passes.html#instnamer-assign-names-to-anonymous-instructions>`_.
* .. function:: add_refprune_pass()
Add the `Reference pruning
<https://github.com/numba/llvmlite/blob/main/ffi/custom_passes.cpp>`_ pass.
.. class:: ModulePassManager()
:noindex:
Create a new pass manager to run optimization passes on a
module.
The ``run`` method is available:
.. method:: run(module)
:noindex:
Run optimization passes on the
*module*, a :class:`ModuleRef` instance.
Returns ``True`` if the optimizations made any modification
to the module. Otherwise returns ``False``.
.. class:: FunctionPassManager(module)
:noindex:
Create a new pass manager to run optimization passes on a
function of the given *module*, a :class:`ModuleRef` instance.
The following methods are available:
* .. method:: finalize()
Run all the finalizers of the optimization passes.
* .. method:: initialize()
Run all the initializers of the optimization passes.
* .. method:: run(function)
:noindex:
Run optimization passes on *function*, a
:class:`ValueRef` instance.
Returns ``True`` if the optimizations made any
modification to the module. Otherwise returns ``False``.