8 Written by `Chris Lattner <mailto:sabre@nondot.org>`_ and
9 `Jim Laskey <mailto:jlaskey@mac.com>`_
11 Introduction --- What is a pass?
12 ================================
14 The LLVM Pass Framework is an important part of the LLVM system, because LLVM
15 passes are where most of the interesting parts of the compiler exist. Passes
16 perform the transformations and optimizations that make up the compiler, they
17 build the analysis results that are used by these transformations, and they
18 are, above all, a structuring technique for compiler code.
20 All LLVM passes are subclasses of the `Pass
21 <http://llvm.org/doxygen/classllvm_1_1Pass.html>`_ class, which implement
22 functionality by overriding virtual methods inherited from ``Pass``. Depending
23 on how your pass works, you should inherit from the :ref:`ModulePass
24 <writing-an-llvm-pass-ModulePass>` , :ref:`CallGraphSCCPass
25 <writing-an-llvm-pass-CallGraphSCCPass>`, :ref:`FunctionPass
26 <writing-an-llvm-pass-FunctionPass>` , or :ref:`LoopPass
27 <writing-an-llvm-pass-LoopPass>`, or :ref:`RegionPass
28 <writing-an-llvm-pass-RegionPass>`, or :ref:`BasicBlockPass
29 <writing-an-llvm-pass-BasicBlockPass>` classes, which gives the system more
30 information about what your pass does, and how it can be combined with other
31 passes. One of the main features of the LLVM Pass Framework is that it
32 schedules passes to run in an efficient way based on the constraints that your
33 pass meets (which are indicated by which class they derive from).
35 We start by showing you how to construct a pass, everything from setting up the
36 code, to compiling, loading, and executing it. After the basics are down, more
37 advanced features are discussed.
39 Quick Start --- Writing hello world
40 ===================================
42 Here we describe how to write the "hello world" of passes. The "Hello" pass is
43 designed to simply print out the name of non-external functions that exist in
44 the program being compiled. It does not modify the program at all, it just
45 inspects it. The source code and files for this pass are available in the LLVM
46 source tree in the ``lib/Transforms/Hello`` directory.
48 .. _writing-an-llvm-pass-makefile:
50 Setting up the build environment
51 --------------------------------
53 .. FIXME: Why does this recommend to build in-tree?
55 First, configure and build LLVM. This needs to be done directly inside the
56 LLVM source tree rather than in a separate objects directory. Next, you need
57 to create a new directory somewhere in the LLVM source base. For this example,
58 we'll assume that you made ``lib/Transforms/Hello``. Finally, you must set up
59 a build script (``Makefile``) that will compile the source code for the new
60 pass. To do this, copy the following into ``Makefile``:
64 # Makefile for hello pass
66 # Path to top level of LLVM hierarchy
69 # Name of the library to build
72 # Make the shared library become a loadable module so the tools can
73 # dlopen/dlsym on the resulting library.
76 # Include the makefile implementation stuff
77 include $(LEVEL)/Makefile.common
79 This makefile specifies that all of the ``.cpp`` files in the current directory
80 are to be compiled and linked together into a shared object
81 ``$(LEVEL)/Debug+Asserts/lib/Hello.so`` that can be dynamically loaded by the
82 :program:`opt` or :program:`bugpoint` tools via their :option:`-load` options.
83 If your operating system uses a suffix other than ``.so`` (such as Windows or Mac
84 OS X), the appropriate extension will be used.
86 If you are used CMake to build LLVM, see :ref:`cmake-out-of-source-pass`.
88 Now that we have the build scripts set up, we just need to write the code for
91 .. _writing-an-llvm-pass-basiccode:
96 Now that we have a way to compile our new pass, we just have to write it.
101 #include "llvm/Pass.h"
102 #include "llvm/Function.h"
103 #include "llvm/Support/raw_ostream.h"
105 Which are needed because we are writing a `Pass
106 <http://llvm.org/doxygen/classllvm_1_1Pass.html>`_, we are operating on
107 `Function <http://llvm.org/doxygen/classllvm_1_1Function.html>`_\ s, and we will
108 be doing some printing.
114 using namespace llvm;
116 ... which is required because the functions from the include files live in the
125 ... which starts out an anonymous namespace. Anonymous namespaces are to C++
126 what the "``static``" keyword is to C (at global scope). It makes the things
127 declared inside of the anonymous namespace visible only to the current file.
128 If you're not familiar with them, consult a decent C++ book for more
131 Next, we declare our pass itself:
135 struct Hello : public FunctionPass {
137 This declares a "``Hello``" class that is a subclass of `FunctionPass
138 <writing-an-llvm-pass-FunctionPass>`. The different builtin pass subclasses
139 are described in detail :ref:`later <writing-an-llvm-pass-pass-classes>`, but
140 for now, know that ``FunctionPass`` operates on a function at a time.
145 Hello() : FunctionPass(ID) {}
147 This declares pass identifier used by LLVM to identify pass. This allows LLVM
148 to avoid using expensive C++ runtime information.
152 virtual bool runOnFunction(Function &F) {
154 errs().write_escaped(F.getName()) << "\n";
157 }; // end of struct Hello
158 } // end of anonymous namespace
160 We declare a :ref:`runOnFunction <writing-an-llvm-pass-runOnFunction>` method,
161 which overrides an abstract virtual method inherited from :ref:`FunctionPass
162 <writing-an-llvm-pass-FunctionPass>`. This is where we are supposed to do our
163 thing, so we just print out our message with the name of each function.
169 We initialize pass ID here. LLVM uses ID's address to identify a pass, so
170 initialization value is not important.
174 static RegisterPass<Hello> X("hello", "Hello World Pass",
175 false /* Only looks at CFG */,
176 false /* Analysis Pass */);
178 Lastly, we :ref:`register our class <writing-an-llvm-pass-registration>`
179 ``Hello``, giving it a command line argument "``hello``", and a name "Hello
180 World Pass". The last two arguments describe its behavior: if a pass walks CFG
181 without modifying it then the third argument is set to ``true``; if a pass is
182 an analysis pass, for example dominator tree pass, then ``true`` is supplied as
185 As a whole, the ``.cpp`` file looks like:
189 #include "llvm/Pass.h"
190 #include "llvm/Function.h"
191 #include "llvm/Support/raw_ostream.h"
193 using namespace llvm;
196 struct Hello : public FunctionPass {
198 Hello() : FunctionPass(ID) {}
200 virtual bool runOnFunction(Function &F) {
202 errs().write_escaped(F.getName()) << '\n';
209 static RegisterPass<Hello> X("hello", "Hello World Pass", false, false);
211 Now that it's all together, compile the file with a simple "``gmake``" command
212 in the local directory and you should get a new file
213 "``Debug+Asserts/lib/Hello.so``" under the top level directory of the LLVM
214 source tree (not in the local directory). Note that everything in this file is
215 contained in an anonymous namespace --- this reflects the fact that passes
216 are self contained units that do not need external interfaces (although they
217 can have them) to be useful.
219 Running a pass with ``opt``
220 ---------------------------
222 Now that you have a brand new shiny shared object file, we can use the
223 :program:`opt` command to run an LLVM program through your pass. Because you
224 registered your pass with ``RegisterPass``, you will be able to use the
225 :program:`opt` tool to access it, once loaded.
227 To test it, follow the example at the end of the :doc:`GettingStarted` to
228 compile "Hello World" to LLVM. We can now run the bitcode file (hello.bc) for
229 the program through our transformation like this (or course, any bitcode file
232 .. code-block:: console
234 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello < hello.bc > /dev/null
239 The :option:`-load` option specifies that :program:`opt` should load your pass
240 as a shared object, which makes "``-hello``" a valid command line argument
241 (which is one reason you need to :ref:`register your pass
242 <writing-an-llvm-pass-registration>`). Because the Hello pass does not modify
243 the program in any interesting way, we just throw away the result of
244 :program:`opt` (sending it to ``/dev/null``).
246 To see what happened to the other string you registered, try running
247 :program:`opt` with the :option:`-help` option:
249 .. code-block:: console
251 $ opt -load ../../../Debug+Asserts/lib/Hello.so -help
252 OVERVIEW: llvm .bc -> .bc modular optimizer
254 USAGE: opt [options] <input bitcode>
257 Optimizations available:
259 -globalopt - Global Variable Optimizer
260 -globalsmodref-aa - Simple mod/ref analysis for globals
261 -gvn - Global Value Numbering
262 -hello - Hello World Pass
263 -indvars - Induction Variable Simplification
264 -inline - Function Integration/Inlining
265 -insert-edge-profiling - Insert instrumentation for edge profiling
268 The pass name gets added as the information string for your pass, giving some
269 documentation to users of :program:`opt`. Now that you have a working pass,
270 you would go ahead and make it do the cool transformations you want. Once you
271 get it all working and tested, it may become useful to find out how fast your
272 pass is. The :ref:`PassManager <writing-an-llvm-pass-passmanager>` provides a
273 nice command line option (:option:`--time-passes`) that allows you to get
274 information about the execution time of your pass along with the other passes
275 you queue up. For example:
277 .. code-block:: console
279 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
283 ===============================================================================
284 ... Pass execution timing report ...
285 ===============================================================================
286 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
288 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
289 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bitcode Writer
290 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
291 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
292 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass
293 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
295 As you can see, our implementation above is pretty fast. The additional
296 passes listed are automatically inserted by the :program:`opt` tool to verify
297 that the LLVM emitted by your pass is still valid and well formed LLVM, which
298 hasn't been broken somehow.
300 Now that you have seen the basics of the mechanics behind passes, we can talk
301 about some more details of how they work and how to use them.
303 .. _writing-an-llvm-pass-pass-classes:
305 Pass classes and requirements
306 =============================
308 One of the first things that you should do when designing a new pass is to
309 decide what class you should subclass for your pass. The :ref:`Hello World
310 <writing-an-llvm-pass-basiccode>` example uses the :ref:`FunctionPass
311 <writing-an-llvm-pass-FunctionPass>` class for its implementation, but we did
312 not discuss why or when this should occur. Here we talk about the classes
313 available, from the most general to the most specific.
315 When choosing a superclass for your ``Pass``, you should choose the **most
316 specific** class possible, while still being able to meet the requirements
317 listed. This gives the LLVM Pass Infrastructure information necessary to
318 optimize how passes are run, so that the resultant compiler isn't unnecessarily
321 The ``ImmutablePass`` class
322 ---------------------------
324 The most plain and boring type of pass is the "`ImmutablePass
325 <http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html>`_" class. This pass
326 type is used for passes that do not have to be run, do not change state, and
327 never need to be updated. This is not a normal type of transformation or
328 analysis, but can provide information about the current compiler configuration.
330 Although this pass class is very infrequently used, it is important for
331 providing information about the current target machine being compiled for, and
332 other static information that can affect the various transformations.
334 ``ImmutablePass``\ es never invalidate other transformations, are never
335 invalidated, and are never "run".
337 .. _writing-an-llvm-pass-ModulePass:
339 The ``ModulePass`` class
340 ------------------------
342 The `ModulePass <http://llvm.org/doxygen/classllvm_1_1ModulePass.html>`_ class
343 is the most general of all superclasses that you can use. Deriving from
344 ``ModulePass`` indicates that your pass uses the entire program as a unit,
345 referring to function bodies in no predictable order, or adding and removing
346 functions. Because nothing is known about the behavior of ``ModulePass``
347 subclasses, no optimization can be done for their execution.
349 A module pass can use function level passes (e.g. dominators) using the
350 ``getAnalysis`` interface ``getAnalysis<DominatorTree>(llvm::Function *)`` to
351 provide the function to retrieve analysis result for, if the function pass does
352 not require any module or immutable passes. Note that this can only be done
353 for functions for which the analysis ran, e.g. in the case of dominators you
354 should only ask for the ``DominatorTree`` for function definitions, not
357 To write a correct ``ModulePass`` subclass, derive from ``ModulePass`` and
358 overload the ``runOnModule`` method with the following signature:
360 The ``runOnModule`` method
361 ^^^^^^^^^^^^^^^^^^^^^^^^^^
365 virtual bool runOnModule(Module &M) = 0;
367 The ``runOnModule`` method performs the interesting work of the pass. It
368 should return ``true`` if the module was modified by the transformation and
371 .. _writing-an-llvm-pass-CallGraphSCCPass:
373 The ``CallGraphSCCPass`` class
374 ------------------------------
376 The `CallGraphSCCPass
377 <http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html>`_ is used by
378 passes that need to traverse the program bottom-up on the call graph (callees
379 before callers). Deriving from ``CallGraphSCCPass`` provides some mechanics
380 for building and traversing the ``CallGraph``, but also allows the system to
381 optimize execution of ``CallGraphSCCPass``\ es. If your pass meets the
382 requirements outlined below, and doesn't meet the requirements of a
383 :ref:`FunctionPass <writing-an-llvm-pass-FunctionPass>` or :ref:`BasicBlockPass
384 <writing-an-llvm-pass-BasicBlockPass>`, you should derive from
385 ``CallGraphSCCPass``.
387 ``TODO``: explain briefly what SCC, Tarjan's algo, and B-U mean.
389 To be explicit, CallGraphSCCPass subclasses are:
391 #. ... *not allowed* to inspect or modify any ``Function``\ s other than those
392 in the current SCC and the direct callers and direct callees of the SCC.
393 #. ... *required* to preserve the current ``CallGraph`` object, updating it to
394 reflect any changes made to the program.
395 #. ... *not allowed* to add or remove SCC's from the current Module, though
396 they may change the contents of an SCC.
397 #. ... *allowed* to add or remove global variables from the current Module.
398 #. ... *allowed* to maintain state across invocations of :ref:`runOnSCC
399 <writing-an-llvm-pass-runOnSCC>` (including global data).
401 Implementing a ``CallGraphSCCPass`` is slightly tricky in some cases because it
402 has to handle SCCs with more than one node in it. All of the virtual methods
403 described below should return ``true`` if they modified the program, or
404 ``false`` if they didn't.
406 The ``doInitialization(CallGraph &)`` method
407 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
411 virtual bool doInitialization(CallGraph &CG);
413 The ``doInitialization`` method is allowed to do most of the things that
414 ``CallGraphSCCPass``\ es are not allowed to do. They can add and remove
415 functions, get pointers to functions, etc. The ``doInitialization`` method is
416 designed to do simple initialization type of stuff that does not depend on the
417 SCCs being processed. The ``doInitialization`` method call is not scheduled to
418 overlap with any other pass executions (thus it should be very fast).
420 .. _writing-an-llvm-pass-runOnSCC:
422 The ``runOnSCC`` method
423 ^^^^^^^^^^^^^^^^^^^^^^^
427 virtual bool runOnSCC(CallGraphSCC &SCC) = 0;
429 The ``runOnSCC`` method performs the interesting work of the pass, and should
430 return ``true`` if the module was modified by the transformation, ``false``
433 The ``doFinalization(CallGraph &)`` method
434 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
438 virtual bool doFinalization(CallGraph &CG);
440 The ``doFinalization`` method is an infrequently used method that is called
441 when the pass framework has finished calling :ref:`runOnFunction
442 <writing-an-llvm-pass-runOnFunction>` for every function in the program being
445 .. _writing-an-llvm-pass-FunctionPass:
447 The ``FunctionPass`` class
448 --------------------------
450 In contrast to ``ModulePass`` subclasses, `FunctionPass
451 <http://llvm.org/doxygen/classllvm_1_1Pass.html>`_ subclasses do have a
452 predictable, local behavior that can be expected by the system. All
453 ``FunctionPass`` execute on each function in the program independent of all of
454 the other functions in the program. ``FunctionPass``\ es do not require that
455 they are executed in a particular order, and ``FunctionPass``\ es do not modify
458 To be explicit, ``FunctionPass`` subclasses are not allowed to:
460 #. Modify a ``Function`` other than the one currently being processed.
461 #. Add or remove ``Function``\ s from the current ``Module``.
462 #. Add or remove global variables from the current ``Module``.
463 #. Maintain state across invocations of:ref:`runOnFunction
464 <writing-an-llvm-pass-runOnFunction>` (including global data).
466 Implementing a ``FunctionPass`` is usually straightforward (See the :ref:`Hello
467 World <writing-an-llvm-pass-basiccode>` pass for example).
468 ``FunctionPass``\ es may overload three virtual methods to do their work. All
469 of these methods should return ``true`` if they modified the program, or
470 ``false`` if they didn't.
472 .. _writing-an-llvm-pass-doInitialization-mod:
474 The ``doInitialization(Module &)`` method
475 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
479 virtual bool doInitialization(Module &M);
481 The ``doInitialization`` method is allowed to do most of the things that
482 ``FunctionPass``\ es are not allowed to do. They can add and remove functions,
483 get pointers to functions, etc. The ``doInitialization`` method is designed to
484 do simple initialization type of stuff that does not depend on the functions
485 being processed. The ``doInitialization`` method call is not scheduled to
486 overlap with any other pass executions (thus it should be very fast).
488 A good example of how this method should be used is the `LowerAllocations
489 <http://llvm.org/doxygen/LowerAllocations_8cpp-source.html>`_ pass. This pass
490 converts ``malloc`` and ``free`` instructions into platform dependent
491 ``malloc()`` and ``free()`` function calls. It uses the ``doInitialization``
492 method to get a reference to the ``malloc`` and ``free`` functions that it
493 needs, adding prototypes to the module if necessary.
495 .. _writing-an-llvm-pass-runOnFunction:
497 The ``runOnFunction`` method
498 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
502 virtual bool runOnFunction(Function &F) = 0;
504 The ``runOnFunction`` method must be implemented by your subclass to do the
505 transformation or analysis work of your pass. As usual, a ``true`` value
506 should be returned if the function is modified.
508 .. _writing-an-llvm-pass-doFinalization-mod:
510 The ``doFinalization(Module &)`` method
511 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
515 virtual bool doFinalization(Module &M);
517 The ``doFinalization`` method is an infrequently used method that is called
518 when the pass framework has finished calling :ref:`runOnFunction
519 <writing-an-llvm-pass-runOnFunction>` for every function in the program being
522 .. _writing-an-llvm-pass-LoopPass:
524 The ``LoopPass`` class
525 ----------------------
527 All ``LoopPass`` execute on each loop in the function independent of all of the
528 other loops in the function. ``LoopPass`` processes loops in loop nest order
529 such that outer most loop is processed last.
531 ``LoopPass`` subclasses are allowed to update loop nest using ``LPPassManager``
532 interface. Implementing a loop pass is usually straightforward.
533 ``LoopPass``\ es may overload three virtual methods to do their work. All
534 these methods should return ``true`` if they modified the program, or ``false``
537 The ``doInitialization(Loop *, LPPassManager &)`` method
538 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
542 virtual bool doInitialization(Loop *, LPPassManager &LPM);
544 The ``doInitialization`` method is designed to do simple initialization type of
545 stuff that does not depend on the functions being processed. The
546 ``doInitialization`` method call is not scheduled to overlap with any other
547 pass executions (thus it should be very fast). ``LPPassManager`` interface
548 should be used to access ``Function`` or ``Module`` level analysis information.
550 .. _writing-an-llvm-pass-runOnLoop:
552 The ``runOnLoop`` method
553 ^^^^^^^^^^^^^^^^^^^^^^^^
557 virtual bool runOnLoop(Loop *, LPPassManager &LPM) = 0;
559 The ``runOnLoop`` method must be implemented by your subclass to do the
560 transformation or analysis work of your pass. As usual, a ``true`` value
561 should be returned if the function is modified. ``LPPassManager`` interface
562 should be used to update loop nest.
564 The ``doFinalization()`` method
565 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
569 virtual bool doFinalization();
571 The ``doFinalization`` method is an infrequently used method that is called
572 when the pass framework has finished calling :ref:`runOnLoop
573 <writing-an-llvm-pass-runOnLoop>` for every loop in the program being compiled.
575 .. _writing-an-llvm-pass-RegionPass:
577 The ``RegionPass`` class
578 ------------------------
580 ``RegionPass`` is similar to :ref:`LoopPass <writing-an-llvm-pass-LoopPass>`,
581 but executes on each single entry single exit region in the function.
582 ``RegionPass`` processes regions in nested order such that the outer most
583 region is processed last.
585 ``RegionPass`` subclasses are allowed to update the region tree by using the
586 ``RGPassManager`` interface. You may overload three virtual methods of
587 ``RegionPass`` to implement your own region pass. All these methods should
588 return ``true`` if they modified the program, or ``false`` if they did not.
590 The ``doInitialization(Region *, RGPassManager &)`` method
591 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
595 virtual bool doInitialization(Region *, RGPassManager &RGM);
597 The ``doInitialization`` method is designed to do simple initialization type of
598 stuff that does not depend on the functions being processed. The
599 ``doInitialization`` method call is not scheduled to overlap with any other
600 pass executions (thus it should be very fast). ``RPPassManager`` interface
601 should be used to access ``Function`` or ``Module`` level analysis information.
603 .. _writing-an-llvm-pass-runOnRegion:
605 The ``runOnRegion`` method
606 ^^^^^^^^^^^^^^^^^^^^^^^^^^
610 virtual bool runOnRegion(Region *, RGPassManager &RGM) = 0;
612 The ``runOnRegion`` method must be implemented by your subclass to do the
613 transformation or analysis work of your pass. As usual, a true value should be
614 returned if the region is modified. ``RGPassManager`` interface should be used to
617 The ``doFinalization()`` method
618 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
622 virtual bool doFinalization();
624 The ``doFinalization`` method is an infrequently used method that is called
625 when the pass framework has finished calling :ref:`runOnRegion
626 <writing-an-llvm-pass-runOnRegion>` for every region in the program being
629 .. _writing-an-llvm-pass-BasicBlockPass:
631 The ``BasicBlockPass`` class
632 ----------------------------
634 ``BasicBlockPass``\ es are just like :ref:`FunctionPass's
635 <writing-an-llvm-pass-FunctionPass>` , except that they must limit their scope
636 of inspection and modification to a single basic block at a time. As such,
637 they are **not** allowed to do any of the following:
639 #. Modify or inspect any basic blocks outside of the current one.
640 #. Maintain state across invocations of :ref:`runOnBasicBlock
641 <writing-an-llvm-pass-runOnBasicBlock>`.
642 #. Modify the control flow graph (by altering terminator instructions)
643 #. Any of the things forbidden for :ref:`FunctionPasses
644 <writing-an-llvm-pass-FunctionPass>`.
646 ``BasicBlockPass``\ es are useful for traditional local and "peephole"
647 optimizations. They may override the same :ref:`doInitialization(Module &)
648 <writing-an-llvm-pass-doInitialization-mod>` and :ref:`doFinalization(Module &)
649 <writing-an-llvm-pass-doFinalization-mod>` methods that :ref:`FunctionPass's
650 <writing-an-llvm-pass-FunctionPass>` have, but also have the following virtual
651 methods that may also be implemented:
653 The ``doInitialization(Function &)`` method
654 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
658 virtual bool doInitialization(Function &F);
660 The ``doInitialization`` method is allowed to do most of the things that
661 ``BasicBlockPass``\ es are not allowed to do, but that ``FunctionPass``\ es
662 can. The ``doInitialization`` method is designed to do simple initialization
663 that does not depend on the ``BasicBlock``\ s being processed. The
664 ``doInitialization`` method call is not scheduled to overlap with any other
665 pass executions (thus it should be very fast).
667 .. _writing-an-llvm-pass-runOnBasicBlock:
669 The ``runOnBasicBlock`` method
670 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
674 virtual bool runOnBasicBlock(BasicBlock &BB) = 0;
676 Override this function to do the work of the ``BasicBlockPass``. This function
677 is not allowed to inspect or modify basic blocks other than the parameter, and
678 are not allowed to modify the CFG. A ``true`` value must be returned if the
679 basic block is modified.
681 The ``doFinalization(Function &)`` method
682 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
686 virtual bool doFinalization(Function &F);
688 The ``doFinalization`` method is an infrequently used method that is called
689 when the pass framework has finished calling :ref:`runOnBasicBlock
690 <writing-an-llvm-pass-runOnBasicBlock>` for every ``BasicBlock`` in the program
691 being compiled. This can be used to perform per-function finalization.
693 The ``MachineFunctionPass`` class
694 ---------------------------------
696 A ``MachineFunctionPass`` is a part of the LLVM code generator that executes on
697 the machine-dependent representation of each LLVM function in the program.
699 Code generator passes are registered and initialized specially by
700 ``TargetMachine::addPassesToEmitFile`` and similar routines, so they cannot
701 generally be run from the :program:`opt` or :program:`bugpoint` commands.
703 A ``MachineFunctionPass`` is also a ``FunctionPass``, so all the restrictions
704 that apply to a ``FunctionPass`` also apply to it. ``MachineFunctionPass``\ es
705 also have additional restrictions. In particular, ``MachineFunctionPass``\ es
706 are not allowed to do any of the following:
708 #. Modify or create any LLVM IR ``Instruction``\ s, ``BasicBlock``\ s,
709 ``Argument``\ s, ``Function``\ s, ``GlobalVariable``\ s,
710 ``GlobalAlias``\ es, or ``Module``\ s.
711 #. Modify a ``MachineFunction`` other than the one currently being processed.
712 #. Maintain state across invocations of :ref:`runOnMachineFunction
713 <writing-an-llvm-pass-runOnMachineFunction>` (including global data).
715 .. _writing-an-llvm-pass-runOnMachineFunction:
717 The ``runOnMachineFunction(MachineFunction &MF)`` method
718 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
722 virtual bool runOnMachineFunction(MachineFunction &MF) = 0;
724 ``runOnMachineFunction`` can be considered the main entry point of a
725 ``MachineFunctionPass``; that is, you should override this method to do the
726 work of your ``MachineFunctionPass``.
728 The ``runOnMachineFunction`` method is called on every ``MachineFunction`` in a
729 ``Module``, so that the ``MachineFunctionPass`` may perform optimizations on
730 the machine-dependent representation of the function. If you want to get at
731 the LLVM ``Function`` for the ``MachineFunction`` you're working on, use
732 ``MachineFunction``'s ``getFunction()`` accessor method --- but remember, you
733 may not modify the LLVM ``Function`` or its contents from a
734 ``MachineFunctionPass``.
736 .. _writing-an-llvm-pass-registration:
741 In the :ref:`Hello World <writing-an-llvm-pass-basiccode>` example pass we
742 illustrated how pass registration works, and discussed some of the reasons that
743 it is used and what it does. Here we discuss how and why passes are
746 As we saw above, passes are registered with the ``RegisterPass`` template. The
747 template parameter is the name of the pass that is to be used on the command
748 line to specify that the pass should be added to a program (for example, with
749 :program:`opt` or :program:`bugpoint`). The first argument is the name of the
750 pass, which is to be used for the :option:`-help` output of programs, as well
751 as for debug output generated by the :option:`--debug-pass` option.
753 If you want your pass to be easily dumpable, you should implement the virtual
761 virtual void print(llvm::raw_ostream &O, const Module *M) const;
763 The ``print`` method must be implemented by "analyses" in order to print a
764 human readable version of the analysis results. This is useful for debugging
765 an analysis itself, as well as for other people to figure out how an analysis
766 works. Use the opt ``-analyze`` argument to invoke this method.
768 The ``llvm::raw_ostream`` parameter specifies the stream to write the results
769 on, and the ``Module`` parameter gives a pointer to the top level module of the
770 program that has been analyzed. Note however that this pointer may be ``NULL``
771 in certain circumstances (such as calling the ``Pass::dump()`` from a
772 debugger), so it should only be used to enhance debug output, it should not be
775 .. _writing-an-llvm-pass-interaction:
777 Specifying interactions between passes
778 --------------------------------------
780 One of the main responsibilities of the ``PassManager`` is to make sure that
781 passes interact with each other correctly. Because ``PassManager`` tries to
782 :ref:`optimize the execution of passes <writing-an-llvm-pass-passmanager>` it
783 must know how the passes interact with each other and what dependencies exist
784 between the various passes. To track this, each pass can declare the set of
785 passes that are required to be executed before the current pass, and the passes
786 which are invalidated by the current pass.
788 Typically this functionality is used to require that analysis results are
789 computed before your pass is run. Running arbitrary transformation passes can
790 invalidate the computed analysis results, which is what the invalidation set
791 specifies. If a pass does not implement the :ref:`getAnalysisUsage
792 <writing-an-llvm-pass-getAnalysisUsage>` method, it defaults to not having any
793 prerequisite passes, and invalidating **all** other passes.
795 .. _writing-an-llvm-pass-getAnalysisUsage:
797 The ``getAnalysisUsage`` method
798 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
802 virtual void getAnalysisUsage(AnalysisUsage &Info) const;
804 By implementing the ``getAnalysisUsage`` method, the required and invalidated
805 sets may be specified for your transformation. The implementation should fill
806 in the `AnalysisUsage
807 <http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html>`_ object with
808 information about which passes are required and not invalidated. To do this, a
809 pass may call any of the following methods on the ``AnalysisUsage`` object:
811 The ``AnalysisUsage::addRequired<>`` and ``AnalysisUsage::addRequiredTransitive<>`` methods
812 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
814 If your pass requires a previous pass to be executed (an analysis for example),
815 it can use one of these methods to arrange for it to be run before your pass.
816 LLVM has many different types of analyses and passes that can be required,
817 spanning the range from ``DominatorSet`` to ``BreakCriticalEdges``. Requiring
818 ``BreakCriticalEdges``, for example, guarantees that there will be no critical
819 edges in the CFG when your pass has been run.
821 Some analyses chain to other analyses to do their job. For example, an
822 `AliasAnalysis <AliasAnalysis>` implementation is required to :ref:`chain
823 <aliasanalysis-chaining>` to other alias analysis passes. In cases where
824 analyses chain, the ``addRequiredTransitive`` method should be used instead of
825 the ``addRequired`` method. This informs the ``PassManager`` that the
826 transitively required pass should be alive as long as the requiring pass is.
828 The ``AnalysisUsage::addPreserved<>`` method
829 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
831 One of the jobs of the ``PassManager`` is to optimize how and when analyses are
832 run. In particular, it attempts to avoid recomputing data unless it needs to.
833 For this reason, passes are allowed to declare that they preserve (i.e., they
834 don't invalidate) an existing analysis if it's available. For example, a
835 simple constant folding pass would not modify the CFG, so it can't possibly
836 affect the results of dominator analysis. By default, all passes are assumed
837 to invalidate all others.
839 The ``AnalysisUsage`` class provides several methods which are useful in
840 certain circumstances that are related to ``addPreserved``. In particular, the
841 ``setPreservesAll`` method can be called to indicate that the pass does not
842 modify the LLVM program at all (which is true for analyses), and the
843 ``setPreservesCFG`` method can be used by transformations that change
844 instructions in the program but do not modify the CFG or terminator
845 instructions (note that this property is implicitly set for
846 :ref:`BasicBlockPass <writing-an-llvm-pass-BasicBlockPass>`\ es).
848 ``addPreserved`` is particularly useful for transformations like
849 ``BreakCriticalEdges``. This pass knows how to update a small set of loop and
850 dominator related analyses if they exist, so it can preserve them, despite the
851 fact that it hacks on the CFG.
853 Example implementations of ``getAnalysisUsage``
854 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
858 // This example modifies the program, but does not modify the CFG
859 void LICM::getAnalysisUsage(AnalysisUsage &AU) const {
860 AU.setPreservesCFG();
861 AU.addRequired<LoopInfo>();
864 .. _writing-an-llvm-pass-getAnalysis:
866 The ``getAnalysis<>`` and ``getAnalysisIfAvailable<>`` methods
867 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
869 The ``Pass::getAnalysis<>`` method is automatically inherited by your class,
870 providing you with access to the passes that you declared that you required
871 with the :ref:`getAnalysisUsage <writing-an-llvm-pass-getAnalysisUsage>`
872 method. It takes a single template argument that specifies which pass class
873 you want, and returns a reference to that pass. For example:
877 bool LICM::runOnFunction(Function &F) {
878 LoopInfo &LI = getAnalysis<LoopInfo>();
882 This method call returns a reference to the pass desired. You may get a
883 runtime assertion failure if you attempt to get an analysis that you did not
884 declare as required in your :ref:`getAnalysisUsage
885 <writing-an-llvm-pass-getAnalysisUsage>` implementation. This method can be
886 called by your ``run*`` method implementation, or by any other local method
887 invoked by your ``run*`` method.
889 A module level pass can use function level analysis info using this interface.
894 bool ModuleLevelPass::runOnModule(Module &M) {
896 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
900 In above example, ``runOnFunction`` for ``DominatorTree`` is called by pass
901 manager before returning a reference to the desired pass.
903 If your pass is capable of updating analyses if they exist (e.g.,
904 ``BreakCriticalEdges``, as described above), you can use the
905 ``getAnalysisIfAvailable`` method, which returns a pointer to the analysis if
906 it is active. For example:
910 if (DominatorSet *DS = getAnalysisIfAvailable<DominatorSet>()) {
911 // A DominatorSet is active. This code will update it.
914 Implementing Analysis Groups
915 ----------------------------
917 Now that we understand the basics of how passes are defined, how they are used,
918 and how they are required from other passes, it's time to get a little bit
919 fancier. All of the pass relationships that we have seen so far are very
920 simple: one pass depends on one other specific pass to be run before it can
921 run. For many applications, this is great, for others, more flexibility is
924 In particular, some analyses are defined such that there is a single simple
925 interface to the analysis results, but multiple ways of calculating them.
926 Consider alias analysis for example. The most trivial alias analysis returns
927 "may alias" for any alias query. The most sophisticated analysis a
928 flow-sensitive, context-sensitive interprocedural analysis that can take a
929 significant amount of time to execute (and obviously, there is a lot of room
930 between these two extremes for other implementations). To cleanly support
931 situations like this, the LLVM Pass Infrastructure supports the notion of
934 Analysis Group Concepts
935 ^^^^^^^^^^^^^^^^^^^^^^^
937 An Analysis Group is a single simple interface that may be implemented by
938 multiple different passes. Analysis Groups can be given human readable names
939 just like passes, but unlike passes, they need not derive from the ``Pass``
940 class. An analysis group may have one or more implementations, one of which is
941 the "default" implementation.
943 Analysis groups are used by client passes just like other passes are: the
944 ``AnalysisUsage::addRequired()`` and ``Pass::getAnalysis()`` methods. In order
945 to resolve this requirement, the :ref:`PassManager
946 <writing-an-llvm-pass-passmanager>` scans the available passes to see if any
947 implementations of the analysis group are available. If none is available, the
948 default implementation is created for the pass to use. All standard rules for
949 :ref:`interaction between passes <writing-an-llvm-pass-interaction>` still
952 Although :ref:`Pass Registration <writing-an-llvm-pass-registration>` is
953 optional for normal passes, all analysis group implementations must be
954 registered, and must use the :ref:`INITIALIZE_AG_PASS
955 <writing-an-llvm-pass-RegisterAnalysisGroup>` template to join the
956 implementation pool. Also, a default implementation of the interface **must**
957 be registered with :ref:`RegisterAnalysisGroup
958 <writing-an-llvm-pass-RegisterAnalysisGroup>`.
960 As a concrete example of an Analysis Group in action, consider the
961 `AliasAnalysis <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_
962 analysis group. The default implementation of the alias analysis interface
963 (the `basicaa <http://llvm.org/doxygen/structBasicAliasAnalysis.html>`_ pass)
964 just does a few simple checks that don't require significant analysis to
965 compute (such as: two different globals can never alias each other, etc).
966 Passes that use the `AliasAnalysis
967 <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_ interface (for
968 example the `gcse <http://llvm.org/doxygen/structGCSE.html>`_ pass), do not
969 care which implementation of alias analysis is actually provided, they just use
970 the designated interface.
972 From the user's perspective, commands work just like normal. Issuing the
973 command ``opt -gcse ...`` will cause the ``basicaa`` class to be instantiated
974 and added to the pass sequence. Issuing the command ``opt -somefancyaa -gcse
975 ...`` will cause the ``gcse`` pass to use the ``somefancyaa`` alias analysis
976 (which doesn't actually exist, it's just a hypothetical example) instead.
978 .. _writing-an-llvm-pass-RegisterAnalysisGroup:
980 Using ``RegisterAnalysisGroup``
981 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
983 The ``RegisterAnalysisGroup`` template is used to register the analysis group
984 itself, while the ``INITIALIZE_AG_PASS`` is used to add pass implementations to
985 the analysis group. First, an analysis group should be registered, with a
986 human readable name provided for it. Unlike registration of passes, there is
987 no command line argument to be specified for the Analysis Group Interface
988 itself, because it is "abstract":
992 static RegisterAnalysisGroup<AliasAnalysis> A("Alias Analysis");
994 Once the analysis is registered, passes can declare that they are valid
995 implementations of the interface by using the following code:
1000 // Declare that we implement the AliasAnalysis interface
1001 INITIALIZE_AG_PASS(FancyAA, AliasAnalysis , "somefancyaa",
1002 "A more complex alias analysis implementation",
1003 false, // Is CFG Only?
1004 true, // Is Analysis?
1005 false); // Is default Analysis Group implementation?
1008 This just shows a class ``FancyAA`` that uses the ``INITIALIZE_AG_PASS`` macro
1009 both to register and to "join" the `AliasAnalysis
1010 <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_ analysis group.
1011 Every implementation of an analysis group should join using this macro.
1016 // Declare that we implement the AliasAnalysis interface
1017 INITIALIZE_AG_PASS(BasicAA, AliasAnalysis, "basicaa",
1018 "Basic Alias Analysis (default AA impl)",
1019 false, // Is CFG Only?
1020 true, // Is Analysis?
1021 true); // Is default Analysis Group implementation?
1024 Here we show how the default implementation is specified (using the final
1025 argument to the ``INITIALIZE_AG_PASS`` template). There must be exactly one
1026 default implementation available at all times for an Analysis Group to be used.
1027 Only default implementation can derive from ``ImmutablePass``. Here we declare
1028 that the `BasicAliasAnalysis
1029 <http://llvm.org/doxygen/structBasicAliasAnalysis.html>`_ pass is the default
1030 implementation for the interface.
1035 The `Statistic <http://llvm.org/doxygen/Statistic_8h-source.html>`_ class is
1036 designed to be an easy way to expose various success metrics from passes.
1037 These statistics are printed at the end of a run, when the :option:`-stats`
1038 command line option is enabled on the command line. See the :ref:`Statistics
1039 section <Statistic>` in the Programmer's Manual for details.
1041 .. _writing-an-llvm-pass-passmanager:
1043 What PassManager does
1044 ---------------------
1046 The `PassManager <http://llvm.org/doxygen/PassManager_8h-source.html>`_ `class
1047 <http://llvm.org/doxygen/classllvm_1_1PassManager.html>`_ takes a list of
1048 passes, ensures their :ref:`prerequisites <writing-an-llvm-pass-interaction>`
1049 are set up correctly, and then schedules passes to run efficiently. All of the
1050 LLVM tools that run passes use the PassManager for execution of these passes.
1052 The PassManager does two main things to try to reduce the execution time of a
1055 #. **Share analysis results.** The ``PassManager`` attempts to avoid
1056 recomputing analysis results as much as possible. This means keeping track
1057 of which analyses are available already, which analyses get invalidated, and
1058 which analyses are needed to be run for a pass. An important part of work
1059 is that the ``PassManager`` tracks the exact lifetime of all analysis
1060 results, allowing it to :ref:`free memory
1061 <writing-an-llvm-pass-releaseMemory>` allocated to holding analysis results
1062 as soon as they are no longer needed.
1064 #. **Pipeline the execution of passes on the program.** The ``PassManager``
1065 attempts to get better cache and memory usage behavior out of a series of
1066 passes by pipelining the passes together. This means that, given a series
1067 of consecutive :ref:`FunctionPass <writing-an-llvm-pass-FunctionPass>`, it
1068 will execute all of the :ref:`FunctionPass
1069 <writing-an-llvm-pass-FunctionPass>` on the first function, then all of the
1070 :ref:`FunctionPasses <writing-an-llvm-pass-FunctionPass>` on the second
1071 function, etc... until the entire program has been run through the passes.
1073 This improves the cache behavior of the compiler, because it is only
1074 touching the LLVM program representation for a single function at a time,
1075 instead of traversing the entire program. It reduces the memory consumption
1076 of compiler, because, for example, only one `DominatorSet
1077 <http://llvm.org/doxygen/classllvm_1_1DominatorSet.html>`_ needs to be
1078 calculated at a time. This also makes it possible to implement some
1079 :ref:`interesting enhancements <writing-an-llvm-pass-SMP>` in the future.
1081 The effectiveness of the ``PassManager`` is influenced directly by how much
1082 information it has about the behaviors of the passes it is scheduling. For
1083 example, the "preserved" set is intentionally conservative in the face of an
1084 unimplemented :ref:`getAnalysisUsage <writing-an-llvm-pass-getAnalysisUsage>`
1085 method. Not implementing when it should be implemented will have the effect of
1086 not allowing any analysis results to live across the execution of your pass.
1088 The ``PassManager`` class exposes a ``--debug-pass`` command line options that
1089 is useful for debugging pass execution, seeing how things work, and diagnosing
1090 when you should be preserving more analyses than you currently are. (To get
1091 information about all of the variants of the ``--debug-pass`` option, just type
1092 "``opt -help-hidden``").
1094 By using the --debug-pass=Structure option, for example, we can see how our
1095 :ref:`Hello World <writing-an-llvm-pass-basiccode>` pass interacts with other
1096 passes. Lets try it out with the gcse and licm passes:
1098 .. code-block:: console
1100 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1102 Function Pass Manager
1103 Dominator Set Construction
1104 Immediate Dominators Construction
1105 Global Common Subexpression Elimination
1106 -- Immediate Dominators Construction
1107 -- Global Common Subexpression Elimination
1108 Natural Loop Construction
1109 Loop Invariant Code Motion
1110 -- Natural Loop Construction
1111 -- Loop Invariant Code Motion
1113 -- Dominator Set Construction
1118 This output shows us when passes are constructed and when the analysis results
1119 are known to be dead (prefixed with "``--``"). Here we see that GCSE uses
1120 dominator and immediate dominator information to do its job. The LICM pass
1121 uses natural loop information, which uses dominator sets, but not immediate
1122 dominators. Because immediate dominators are no longer useful after the GCSE
1123 pass, it is immediately destroyed. The dominator sets are then reused to
1124 compute natural loop information, which is then used by the LICM pass.
1126 After the LICM pass, the module verifier runs (which is automatically added by
1127 the :program:`opt` tool), which uses the dominator set to check that the
1128 resultant LLVM code is well formed. After it finishes, the dominator set
1129 information is destroyed, after being computed once, and shared by three
1132 Lets see how this changes when we run the :ref:`Hello World
1133 <writing-an-llvm-pass-basiccode>` pass in between the two passes:
1135 .. code-block:: console
1137 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1139 Function Pass Manager
1140 Dominator Set Construction
1141 Immediate Dominators Construction
1142 Global Common Subexpression Elimination
1143 -- Dominator Set Construction
1144 -- Immediate Dominators Construction
1145 -- Global Common Subexpression Elimination
1148 Dominator Set Construction
1149 Natural Loop Construction
1150 Loop Invariant Code Motion
1151 -- Natural Loop Construction
1152 -- Loop Invariant Code Motion
1154 -- Dominator Set Construction
1162 Here we see that the :ref:`Hello World <writing-an-llvm-pass-basiccode>` pass
1163 has killed the Dominator Set pass, even though it doesn't modify the code at
1164 all! To fix this, we need to add the following :ref:`getAnalysisUsage
1165 <writing-an-llvm-pass-getAnalysisUsage>` method to our pass:
1169 // We don't modify the program, so we preserve all analyses
1170 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1171 AU.setPreservesAll();
1174 Now when we run our pass, we get this output:
1176 .. code-block:: console
1178 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1179 Pass Arguments: -gcse -hello -licm
1181 Function Pass Manager
1182 Dominator Set Construction
1183 Immediate Dominators Construction
1184 Global Common Subexpression Elimination
1185 -- Immediate Dominators Construction
1186 -- Global Common Subexpression Elimination
1189 Natural Loop Construction
1190 Loop Invariant Code Motion
1191 -- Loop Invariant Code Motion
1192 -- Natural Loop Construction
1194 -- Dominator Set Construction
1202 Which shows that we don't accidentally invalidate dominator information
1203 anymore, and therefore do not have to compute it twice.
1205 .. _writing-an-llvm-pass-releaseMemory:
1207 The ``releaseMemory`` method
1208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1212 virtual void releaseMemory();
1214 The ``PassManager`` automatically determines when to compute analysis results,
1215 and how long to keep them around for. Because the lifetime of the pass object
1216 itself is effectively the entire duration of the compilation process, we need
1217 some way to free analysis results when they are no longer useful. The
1218 ``releaseMemory`` virtual method is the way to do this.
1220 If you are writing an analysis or any other pass that retains a significant
1221 amount of state (for use by another pass which "requires" your pass and uses
1222 the :ref:`getAnalysis <writing-an-llvm-pass-getAnalysis>` method) you should
1223 implement ``releaseMemory`` to, well, release the memory allocated to maintain
1224 this internal state. This method is called after the ``run*`` method for the
1225 class, before the next call of ``run*`` in your pass.
1227 Registering dynamically loaded passes
1228 =====================================
1230 *Size matters* when constructing production quality tools using LLVM, both for
1231 the purposes of distribution, and for regulating the resident code size when
1232 running on the target system. Therefore, it becomes desirable to selectively
1233 use some passes, while omitting others and maintain the flexibility to change
1234 configurations later on. You want to be able to do all this, and, provide
1235 feedback to the user. This is where pass registration comes into play.
1237 The fundamental mechanisms for pass registration are the
1238 ``MachinePassRegistry`` class and subclasses of ``MachinePassRegistryNode``.
1240 An instance of ``MachinePassRegistry`` is used to maintain a list of
1241 ``MachinePassRegistryNode`` objects. This instance maintains the list and
1242 communicates additions and deletions to the command line interface.
1244 An instance of ``MachinePassRegistryNode`` subclass is used to maintain
1245 information provided about a particular pass. This information includes the
1246 command line name, the command help string and the address of the function used
1247 to create an instance of the pass. A global static constructor of one of these
1248 instances *registers* with a corresponding ``MachinePassRegistry``, the static
1249 destructor *unregisters*. Thus a pass that is statically linked in the tool
1250 will be registered at start up. A dynamically loaded pass will register on
1251 load and unregister at unload.
1253 Using existing registries
1254 -------------------------
1256 There are predefined registries to track instruction scheduling
1257 (``RegisterScheduler``) and register allocation (``RegisterRegAlloc``) machine
1258 passes. Here we will describe how to *register* a register allocator machine
1261 Implement your register allocator machine pass. In your register allocator
1262 ``.cpp`` file add the following include:
1266 #include "llvm/CodeGen/RegAllocRegistry.h"
1268 Also in your register allocator ``.cpp`` file, define a creator function in the
1273 FunctionPass *createMyRegisterAllocator() {
1274 return new MyRegisterAllocator();
1277 Note that the signature of this function should match the type of
1278 ``RegisterRegAlloc::FunctionPassCtor``. In the same file add the "installing"
1279 declaration, in the form:
1283 static RegisterRegAlloc myRegAlloc("myregalloc",
1284 "my register allocator help string",
1285 createMyRegisterAllocator);
1287 Note the two spaces prior to the help string produces a tidy result on the
1288 :option:`-help` query.
1290 .. code-block:: console
1294 -regalloc - Register allocator to use (default=linearscan)
1295 =linearscan - linear scan register allocator
1296 =local - local register allocator
1297 =simple - simple register allocator
1298 =myregalloc - my register allocator help string
1301 And that's it. The user is now free to use ``-regalloc=myregalloc`` as an
1302 option. Registering instruction schedulers is similar except use the
1303 ``RegisterScheduler`` class. Note that the
1304 ``RegisterScheduler::FunctionPassCtor`` is significantly different from
1305 ``RegisterRegAlloc::FunctionPassCtor``.
1307 To force the load/linking of your register allocator into the
1308 :program:`llc`/:program:`lli` tools, add your creator function's global
1309 declaration to ``Passes.h`` and add a "pseudo" call line to
1310 ``llvm/Codegen/LinkAllCodegenComponents.h``.
1312 Creating new registries
1313 -----------------------
1315 The easiest way to get started is to clone one of the existing registries; we
1316 recommend ``llvm/CodeGen/RegAllocRegistry.h``. The key things to modify are
1317 the class name and the ``FunctionPassCtor`` type.
1319 Then you need to declare the registry. Example: if your pass registry is
1320 ``RegisterMyPasses`` then define:
1324 MachinePassRegistry RegisterMyPasses::Registry;
1326 And finally, declare the command line option for your passes. Example:
1330 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1331 RegisterPassParser<RegisterMyPasses> >
1333 cl::init(&createDefaultMyPass),
1334 cl::desc("my pass option help"));
1336 Here the command option is "``mypass``", with ``createDefaultMyPass`` as the
1339 Using GDB with dynamically loaded passes
1340 ----------------------------------------
1342 Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1343 should be. First of all, you can't set a breakpoint in a shared object that
1344 has not been loaded yet, and second of all there are problems with inlined
1345 functions in shared objects. Here are some suggestions to debugging your pass
1348 For sake of discussion, I'm going to assume that you are debugging a
1349 transformation invoked by :program:`opt`, although nothing described here
1352 Setting a breakpoint in your pass
1353 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1355 First thing you do is start gdb on the opt process:
1357 .. code-block:: console
1361 Copyright 2000 Free Software Foundation, Inc.
1362 GDB is free software, covered by the GNU General Public License, and you are
1363 welcome to change it and/or distribute copies of it under certain conditions.
1364 Type "show copying" to see the conditions.
1365 There is absolutely no warranty for GDB. Type "show warranty" for details.
1366 This GDB was configured as "sparc-sun-solaris2.6"...
1369 Note that :program:`opt` has a lot of debugging information in it, so it takes
1370 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1371 (the shared object isn't loaded until runtime), we must execute the process,
1372 and have it stop before it invokes our pass, but after it has loaded the shared
1373 object. The most foolproof way of doing this is to set a breakpoint in
1374 ``PassManager::run`` and then run the process with the arguments you want:
1376 .. code-block:: console
1378 $ (gdb) break llvm::PassManager::run
1379 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1380 (gdb) run test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1381 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1382 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1383 70 bool PassManager::run(Module &M) { return PM->run(M); }
1386 Once the :program:`opt` stops in the ``PassManager::run`` method you are now
1387 free to set breakpoints in your pass so that you can trace through execution or
1388 do other standard debugging stuff.
1390 Miscellaneous Problems
1391 ^^^^^^^^^^^^^^^^^^^^^^
1393 Once you have the basics down, there are a couple of problems that GDB has,
1394 some with solutions, some without.
1396 * Inline functions have bogus stack information. In general, GDB does a pretty
1397 good job getting stack traces and stepping through inline functions. When a
1398 pass is dynamically loaded however, it somehow completely loses this
1399 capability. The only solution I know of is to de-inline a function (move it
1400 from the body of a class to a ``.cpp`` file).
1402 * Restarting the program breaks breakpoints. After following the information
1403 above, you have succeeded in getting some breakpoints planted in your pass.
1404 Nex thing you know, you restart the program (i.e., you type "``run``" again),
1405 and you start getting errors about breakpoints being unsettable. The only
1406 way I have found to "fix" this problem is to delete the breakpoints that are
1407 already set in your pass, run the program, and re-set the breakpoints once
1408 execution stops in ``PassManager::run``.
1410 Hopefully these tips will help with common case debugging situations. If you'd
1411 like to contribute some tips of your own, just contact `Chris
1412 <mailto:sabre@nondot.org>`_.
1414 Future extensions planned
1415 -------------------------
1417 Although the LLVM Pass Infrastructure is very capable as it stands, and does
1418 some nifty stuff, there are things we'd like to add in the future. Here is
1421 .. _writing-an-llvm-pass-SMP:
1426 Multiple CPU machines are becoming more common and compilation can never be
1427 fast enough: obviously we should allow for a multithreaded compiler. Because
1428 of the semantics defined for passes above (specifically they cannot maintain
1429 state across invocations of their ``run*`` methods), a nice clean way to
1430 implement a multithreaded compiler would be for the ``PassManager`` class to
1431 create multiple instances of each pass object, and allow the separate instances
1432 to be hacking on different parts of the program at the same time.
1434 This implementation would prevent each of the passes from having to implement
1435 multithreaded constructs, requiring only the LLVM core to have locking in a few
1436 places (for global resources). Although this is a simple extension, we simply
1437 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1438 Despite that, we have kept the LLVM passes SMP ready, and you should too.