8 Introduction --- What is a pass?
9 ================================
11 The LLVM Pass Framework is an important part of the LLVM system, because LLVM
12 passes are where most of the interesting parts of the compiler exist. Passes
13 perform the transformations and optimizations that make up the compiler, they
14 build the analysis results that are used by these transformations, and they
15 are, above all, a structuring technique for compiler code.
17 All LLVM passes are subclasses of the `Pass
18 <http://llvm.org/doxygen/classllvm_1_1Pass.html>`_ class, which implement
19 functionality by overriding virtual methods inherited from ``Pass``. Depending
20 on how your pass works, you should inherit from the :ref:`ModulePass
21 <writing-an-llvm-pass-ModulePass>` , :ref:`CallGraphSCCPass
22 <writing-an-llvm-pass-CallGraphSCCPass>`, :ref:`FunctionPass
23 <writing-an-llvm-pass-FunctionPass>` , or :ref:`LoopPass
24 <writing-an-llvm-pass-LoopPass>`, or :ref:`RegionPass
25 <writing-an-llvm-pass-RegionPass>`, or :ref:`BasicBlockPass
26 <writing-an-llvm-pass-BasicBlockPass>` classes, which gives the system more
27 information about what your pass does, and how it can be combined with other
28 passes. One of the main features of the LLVM Pass Framework is that it
29 schedules passes to run in an efficient way based on the constraints that your
30 pass meets (which are indicated by which class they derive from).
32 We start by showing you how to construct a pass, everything from setting up the
33 code, to compiling, loading, and executing it. After the basics are down, more
34 advanced features are discussed.
36 Quick Start --- Writing hello world
37 ===================================
39 Here we describe how to write the "hello world" of passes. The "Hello" pass is
40 designed to simply print out the name of non-external functions that exist in
41 the program being compiled. It does not modify the program at all, it just
42 inspects it. The source code and files for this pass are available in the LLVM
43 source tree in the ``lib/Transforms/Hello`` directory.
45 .. _writing-an-llvm-pass-makefile:
47 Setting up the build environment
48 --------------------------------
50 First, configure and build LLVM. Next, you need to create a new directory
51 somewhere in the LLVM source base. For this example, we'll assume that you
52 made ``lib/Transforms/Hello``. Finally, you must set up a build script
53 (``Makefile``) that will compile the source code for the new pass. To do this,
54 copy the following into ``Makefile``:
58 # Makefile for hello pass
60 # Path to top level of LLVM hierarchy
63 # Name of the library to build
66 # Make the shared library become a loadable module so the tools can
67 # dlopen/dlsym on the resulting library.
70 # Include the makefile implementation stuff
71 include $(LEVEL)/Makefile.common
73 This makefile specifies that all of the ``.cpp`` files in the current directory
74 are to be compiled and linked together into a shared object
75 ``$(LEVEL)/Debug+Asserts/lib/Hello.so`` that can be dynamically loaded by the
76 :program:`opt` or :program:`bugpoint` tools via their :option:`-load` options.
77 If your operating system uses a suffix other than ``.so`` (such as Windows or Mac
78 OS X), the appropriate extension will be used.
80 If you are used CMake to build LLVM, see :ref:`cmake-out-of-source-pass`.
82 Now that we have the build scripts set up, we just need to write the code for
85 .. _writing-an-llvm-pass-basiccode:
90 Now that we have a way to compile our new pass, we just have to write it.
95 #include "llvm/Pass.h"
96 #include "llvm/IR/Function.h"
97 #include "llvm/Support/raw_ostream.h"
99 Which are needed because we are writing a `Pass
100 <http://llvm.org/doxygen/classllvm_1_1Pass.html>`_, we are operating on
101 `Function <http://llvm.org/doxygen/classllvm_1_1Function.html>`_\ s, and we will
102 be doing some printing.
108 using namespace llvm;
110 ... which is required because the functions from the include files live in the
119 ... which starts out an anonymous namespace. Anonymous namespaces are to C++
120 what the "``static``" keyword is to C (at global scope). It makes the things
121 declared inside of the anonymous namespace visible only to the current file.
122 If you're not familiar with them, consult a decent C++ book for more
125 Next, we declare our pass itself:
129 struct Hello : public FunctionPass {
131 This declares a "``Hello``" class that is a subclass of :ref:`FunctionPass
132 <writing-an-llvm-pass-FunctionPass>`. The different builtin pass subclasses
133 are described in detail :ref:`later <writing-an-llvm-pass-pass-classes>`, but
134 for now, know that ``FunctionPass`` operates on a function at a time.
139 Hello() : FunctionPass(ID) {}
141 This declares pass identifier used by LLVM to identify pass. This allows LLVM
142 to avoid using expensive C++ runtime information.
146 bool runOnFunction(Function &F) override {
148 errs().write_escaped(F.getName()) << "\n";
151 }; // end of struct Hello
152 } // end of anonymous namespace
154 We declare a :ref:`runOnFunction <writing-an-llvm-pass-runOnFunction>` method,
155 which overrides an abstract virtual method inherited from :ref:`FunctionPass
156 <writing-an-llvm-pass-FunctionPass>`. This is where we are supposed to do our
157 thing, so we just print out our message with the name of each function.
163 We initialize pass ID here. LLVM uses ID's address to identify a pass, so
164 initialization value is not important.
168 static RegisterPass<Hello> X("hello", "Hello World Pass",
169 false /* Only looks at CFG */,
170 false /* Analysis Pass */);
172 Lastly, we :ref:`register our class <writing-an-llvm-pass-registration>`
173 ``Hello``, giving it a command line argument "``hello``", and a name "Hello
174 World Pass". The last two arguments describe its behavior: if a pass walks CFG
175 without modifying it then the third argument is set to ``true``; if a pass is
176 an analysis pass, for example dominator tree pass, then ``true`` is supplied as
179 As a whole, the ``.cpp`` file looks like:
183 #include "llvm/Pass.h"
184 #include "llvm/IR/Function.h"
185 #include "llvm/Support/raw_ostream.h"
187 using namespace llvm;
190 struct Hello : public FunctionPass {
192 Hello() : FunctionPass(ID) {}
194 bool runOnFunction(Function &F) override {
196 errs().write_escaped(F.getName()) << '\n';
203 static RegisterPass<Hello> X("hello", "Hello World Pass", false, false);
205 Now that it's all together, compile the file with a simple "``gmake``" command
206 in the local directory and you should get a new file
207 "``Debug+Asserts/lib/Hello.so``" under the top level directory of the LLVM
208 source tree (not in the local directory). Note that everything in this file is
209 contained in an anonymous namespace --- this reflects the fact that passes
210 are self contained units that do not need external interfaces (although they
211 can have them) to be useful.
213 Running a pass with ``opt``
214 ---------------------------
216 Now that you have a brand new shiny shared object file, we can use the
217 :program:`opt` command to run an LLVM program through your pass. Because you
218 registered your pass with ``RegisterPass``, you will be able to use the
219 :program:`opt` tool to access it, once loaded.
221 To test it, follow the example at the end of the :doc:`GettingStarted` to
222 compile "Hello World" to LLVM. We can now run the bitcode file (hello.bc) for
223 the program through our transformation like this (or course, any bitcode file
226 .. code-block:: console
228 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello < hello.bc > /dev/null
233 The :option:`-load` option specifies that :program:`opt` should load your pass
234 as a shared object, which makes "``-hello``" a valid command line argument
235 (which is one reason you need to :ref:`register your pass
236 <writing-an-llvm-pass-registration>`). Because the Hello pass does not modify
237 the program in any interesting way, we just throw away the result of
238 :program:`opt` (sending it to ``/dev/null``).
240 To see what happened to the other string you registered, try running
241 :program:`opt` with the :option:`-help` option:
243 .. code-block:: console
245 $ opt -load ../../../Debug+Asserts/lib/Hello.so -help
246 OVERVIEW: llvm .bc -> .bc modular optimizer
248 USAGE: opt [options] <input bitcode>
251 Optimizations available:
253 -globalopt - Global Variable Optimizer
254 -globalsmodref-aa - Simple mod/ref analysis for globals
255 -gvn - Global Value Numbering
256 -hello - Hello World Pass
257 -indvars - Induction Variable Simplification
258 -inline - Function Integration/Inlining
261 The pass name gets added as the information string for your pass, giving some
262 documentation to users of :program:`opt`. Now that you have a working pass,
263 you would go ahead and make it do the cool transformations you want. Once you
264 get it all working and tested, it may become useful to find out how fast your
265 pass is. The :ref:`PassManager <writing-an-llvm-pass-passmanager>` provides a
266 nice command line option (:option:`--time-passes`) that allows you to get
267 information about the execution time of your pass along with the other passes
268 you queue up. For example:
270 .. code-block:: console
272 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
276 ===============================================================================
277 ... Pass execution timing report ...
278 ===============================================================================
279 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
281 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
282 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bitcode Writer
283 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
284 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
285 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass
286 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
288 As you can see, our implementation above is pretty fast. The additional
289 passes listed are automatically inserted by the :program:`opt` tool to verify
290 that the LLVM emitted by your pass is still valid and well formed LLVM, which
291 hasn't been broken somehow.
293 Now that you have seen the basics of the mechanics behind passes, we can talk
294 about some more details of how they work and how to use them.
296 .. _writing-an-llvm-pass-pass-classes:
298 Pass classes and requirements
299 =============================
301 One of the first things that you should do when designing a new pass is to
302 decide what class you should subclass for your pass. The :ref:`Hello World
303 <writing-an-llvm-pass-basiccode>` example uses the :ref:`FunctionPass
304 <writing-an-llvm-pass-FunctionPass>` class for its implementation, but we did
305 not discuss why or when this should occur. Here we talk about the classes
306 available, from the most general to the most specific.
308 When choosing a superclass for your ``Pass``, you should choose the **most
309 specific** class possible, while still being able to meet the requirements
310 listed. This gives the LLVM Pass Infrastructure information necessary to
311 optimize how passes are run, so that the resultant compiler isn't unnecessarily
314 The ``ImmutablePass`` class
315 ---------------------------
317 The most plain and boring type of pass is the "`ImmutablePass
318 <http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html>`_" class. This pass
319 type is used for passes that do not have to be run, do not change state, and
320 never need to be updated. This is not a normal type of transformation or
321 analysis, but can provide information about the current compiler configuration.
323 Although this pass class is very infrequently used, it is important for
324 providing information about the current target machine being compiled for, and
325 other static information that can affect the various transformations.
327 ``ImmutablePass``\ es never invalidate other transformations, are never
328 invalidated, and are never "run".
330 .. _writing-an-llvm-pass-ModulePass:
332 The ``ModulePass`` class
333 ------------------------
335 The `ModulePass <http://llvm.org/doxygen/classllvm_1_1ModulePass.html>`_ class
336 is the most general of all superclasses that you can use. Deriving from
337 ``ModulePass`` indicates that your pass uses the entire program as a unit,
338 referring to function bodies in no predictable order, or adding and removing
339 functions. Because nothing is known about the behavior of ``ModulePass``
340 subclasses, no optimization can be done for their execution.
342 A module pass can use function level passes (e.g. dominators) using the
343 ``getAnalysis`` interface ``getAnalysis<DominatorTree>(llvm::Function *)`` to
344 provide the function to retrieve analysis result for, if the function pass does
345 not require any module or immutable passes. Note that this can only be done
346 for functions for which the analysis ran, e.g. in the case of dominators you
347 should only ask for the ``DominatorTree`` for function definitions, not
350 To write a correct ``ModulePass`` subclass, derive from ``ModulePass`` and
351 overload the ``runOnModule`` method with the following signature:
353 The ``runOnModule`` method
354 ^^^^^^^^^^^^^^^^^^^^^^^^^^
358 virtual bool runOnModule(Module &M) = 0;
360 The ``runOnModule`` method performs the interesting work of the pass. It
361 should return ``true`` if the module was modified by the transformation and
364 .. _writing-an-llvm-pass-CallGraphSCCPass:
366 The ``CallGraphSCCPass`` class
367 ------------------------------
369 The `CallGraphSCCPass
370 <http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html>`_ is used by
371 passes that need to traverse the program bottom-up on the call graph (callees
372 before callers). Deriving from ``CallGraphSCCPass`` provides some mechanics
373 for building and traversing the ``CallGraph``, but also allows the system to
374 optimize execution of ``CallGraphSCCPass``\ es. If your pass meets the
375 requirements outlined below, and doesn't meet the requirements of a
376 :ref:`FunctionPass <writing-an-llvm-pass-FunctionPass>` or :ref:`BasicBlockPass
377 <writing-an-llvm-pass-BasicBlockPass>`, you should derive from
378 ``CallGraphSCCPass``.
380 ``TODO``: explain briefly what SCC, Tarjan's algo, and B-U mean.
382 To be explicit, CallGraphSCCPass subclasses are:
384 #. ... *not allowed* to inspect or modify any ``Function``\ s other than those
385 in the current SCC and the direct callers and direct callees of the SCC.
386 #. ... *required* to preserve the current ``CallGraph`` object, updating it to
387 reflect any changes made to the program.
388 #. ... *not allowed* to add or remove SCC's from the current Module, though
389 they may change the contents of an SCC.
390 #. ... *allowed* to add or remove global variables from the current Module.
391 #. ... *allowed* to maintain state across invocations of :ref:`runOnSCC
392 <writing-an-llvm-pass-runOnSCC>` (including global data).
394 Implementing a ``CallGraphSCCPass`` is slightly tricky in some cases because it
395 has to handle SCCs with more than one node in it. All of the virtual methods
396 described below should return ``true`` if they modified the program, or
397 ``false`` if they didn't.
399 The ``doInitialization(CallGraph &)`` method
400 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
404 virtual bool doInitialization(CallGraph &CG);
406 The ``doInitialization`` method is allowed to do most of the things that
407 ``CallGraphSCCPass``\ es are not allowed to do. They can add and remove
408 functions, get pointers to functions, etc. The ``doInitialization`` method is
409 designed to do simple initialization type of stuff that does not depend on the
410 SCCs being processed. The ``doInitialization`` method call is not scheduled to
411 overlap with any other pass executions (thus it should be very fast).
413 .. _writing-an-llvm-pass-runOnSCC:
415 The ``runOnSCC`` method
416 ^^^^^^^^^^^^^^^^^^^^^^^
420 virtual bool runOnSCC(CallGraphSCC &SCC) = 0;
422 The ``runOnSCC`` method performs the interesting work of the pass, and should
423 return ``true`` if the module was modified by the transformation, ``false``
426 The ``doFinalization(CallGraph &)`` method
427 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
431 virtual bool doFinalization(CallGraph &CG);
433 The ``doFinalization`` method is an infrequently used method that is called
434 when the pass framework has finished calling :ref:`runOnSCC
435 <writing-an-llvm-pass-runOnSCC>` for every SCC in the program being compiled.
437 .. _writing-an-llvm-pass-FunctionPass:
439 The ``FunctionPass`` class
440 --------------------------
442 In contrast to ``ModulePass`` subclasses, `FunctionPass
443 <http://llvm.org/doxygen/classllvm_1_1Pass.html>`_ subclasses do have a
444 predictable, local behavior that can be expected by the system. All
445 ``FunctionPass`` execute on each function in the program independent of all of
446 the other functions in the program. ``FunctionPass``\ es do not require that
447 they are executed in a particular order, and ``FunctionPass``\ es do not modify
450 To be explicit, ``FunctionPass`` subclasses are not allowed to:
452 #. Inspect or modify a ``Function`` other than the one currently being processed.
453 #. Add or remove ``Function``\ s from the current ``Module``.
454 #. Add or remove global variables from the current ``Module``.
455 #. Maintain state across invocations of :ref:`runOnFunction
456 <writing-an-llvm-pass-runOnFunction>` (including global data).
458 Implementing a ``FunctionPass`` is usually straightforward (See the :ref:`Hello
459 World <writing-an-llvm-pass-basiccode>` pass for example).
460 ``FunctionPass``\ es may overload three virtual methods to do their work. All
461 of these methods should return ``true`` if they modified the program, or
462 ``false`` if they didn't.
464 .. _writing-an-llvm-pass-doInitialization-mod:
466 The ``doInitialization(Module &)`` method
467 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
471 virtual bool doInitialization(Module &M);
473 The ``doInitialization`` method is allowed to do most of the things that
474 ``FunctionPass``\ es are not allowed to do. They can add and remove functions,
475 get pointers to functions, etc. The ``doInitialization`` method is designed to
476 do simple initialization type of stuff that does not depend on the functions
477 being processed. The ``doInitialization`` method call is not scheduled to
478 overlap with any other pass executions (thus it should be very fast).
480 A good example of how this method should be used is the `LowerAllocations
481 <http://llvm.org/doxygen/LowerAllocations_8cpp-source.html>`_ pass. This pass
482 converts ``malloc`` and ``free`` instructions into platform dependent
483 ``malloc()`` and ``free()`` function calls. It uses the ``doInitialization``
484 method to get a reference to the ``malloc`` and ``free`` functions that it
485 needs, adding prototypes to the module if necessary.
487 .. _writing-an-llvm-pass-runOnFunction:
489 The ``runOnFunction`` method
490 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
494 virtual bool runOnFunction(Function &F) = 0;
496 The ``runOnFunction`` method must be implemented by your subclass to do the
497 transformation or analysis work of your pass. As usual, a ``true`` value
498 should be returned if the function is modified.
500 .. _writing-an-llvm-pass-doFinalization-mod:
502 The ``doFinalization(Module &)`` method
503 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
507 virtual bool doFinalization(Module &M);
509 The ``doFinalization`` method is an infrequently used method that is called
510 when the pass framework has finished calling :ref:`runOnFunction
511 <writing-an-llvm-pass-runOnFunction>` for every function in the program being
514 .. _writing-an-llvm-pass-LoopPass:
516 The ``LoopPass`` class
517 ----------------------
519 All ``LoopPass`` execute on each loop in the function independent of all of the
520 other loops in the function. ``LoopPass`` processes loops in loop nest order
521 such that outer most loop is processed last.
523 ``LoopPass`` subclasses are allowed to update loop nest using ``LPPassManager``
524 interface. Implementing a loop pass is usually straightforward.
525 ``LoopPass``\ es may overload three virtual methods to do their work. All
526 these methods should return ``true`` if they modified the program, or ``false``
529 The ``doInitialization(Loop *, LPPassManager &)`` method
530 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
534 virtual bool doInitialization(Loop *, LPPassManager &LPM);
536 The ``doInitialization`` method is designed to do simple initialization type of
537 stuff that does not depend on the functions being processed. The
538 ``doInitialization`` method call is not scheduled to overlap with any other
539 pass executions (thus it should be very fast). ``LPPassManager`` interface
540 should be used to access ``Function`` or ``Module`` level analysis information.
542 .. _writing-an-llvm-pass-runOnLoop:
544 The ``runOnLoop`` method
545 ^^^^^^^^^^^^^^^^^^^^^^^^
549 virtual bool runOnLoop(Loop *, LPPassManager &LPM) = 0;
551 The ``runOnLoop`` method must be implemented by your subclass to do the
552 transformation or analysis work of your pass. As usual, a ``true`` value
553 should be returned if the function is modified. ``LPPassManager`` interface
554 should be used to update loop nest.
556 The ``doFinalization()`` method
557 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
561 virtual bool doFinalization();
563 The ``doFinalization`` method is an infrequently used method that is called
564 when the pass framework has finished calling :ref:`runOnLoop
565 <writing-an-llvm-pass-runOnLoop>` for every loop in the program being compiled.
567 .. _writing-an-llvm-pass-RegionPass:
569 The ``RegionPass`` class
570 ------------------------
572 ``RegionPass`` is similar to :ref:`LoopPass <writing-an-llvm-pass-LoopPass>`,
573 but executes on each single entry single exit region in the function.
574 ``RegionPass`` processes regions in nested order such that the outer most
575 region is processed last.
577 ``RegionPass`` subclasses are allowed to update the region tree by using the
578 ``RGPassManager`` interface. You may overload three virtual methods of
579 ``RegionPass`` to implement your own region pass. All these methods should
580 return ``true`` if they modified the program, or ``false`` if they did not.
582 The ``doInitialization(Region *, RGPassManager &)`` method
583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
587 virtual bool doInitialization(Region *, RGPassManager &RGM);
589 The ``doInitialization`` method is designed to do simple initialization type of
590 stuff that does not depend on the functions being processed. The
591 ``doInitialization`` method call is not scheduled to overlap with any other
592 pass executions (thus it should be very fast). ``RPPassManager`` interface
593 should be used to access ``Function`` or ``Module`` level analysis information.
595 .. _writing-an-llvm-pass-runOnRegion:
597 The ``runOnRegion`` method
598 ^^^^^^^^^^^^^^^^^^^^^^^^^^
602 virtual bool runOnRegion(Region *, RGPassManager &RGM) = 0;
604 The ``runOnRegion`` method must be implemented by your subclass to do the
605 transformation or analysis work of your pass. As usual, a true value should be
606 returned if the region is modified. ``RGPassManager`` interface should be used to
609 The ``doFinalization()`` method
610 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
614 virtual bool doFinalization();
616 The ``doFinalization`` method is an infrequently used method that is called
617 when the pass framework has finished calling :ref:`runOnRegion
618 <writing-an-llvm-pass-runOnRegion>` for every region in the program being
621 .. _writing-an-llvm-pass-BasicBlockPass:
623 The ``BasicBlockPass`` class
624 ----------------------------
626 ``BasicBlockPass``\ es are just like :ref:`FunctionPass's
627 <writing-an-llvm-pass-FunctionPass>` , except that they must limit their scope
628 of inspection and modification to a single basic block at a time. As such,
629 they are **not** allowed to do any of the following:
631 #. Modify or inspect any basic blocks outside of the current one.
632 #. Maintain state across invocations of :ref:`runOnBasicBlock
633 <writing-an-llvm-pass-runOnBasicBlock>`.
634 #. Modify the control flow graph (by altering terminator instructions)
635 #. Any of the things forbidden for :ref:`FunctionPasses
636 <writing-an-llvm-pass-FunctionPass>`.
638 ``BasicBlockPass``\ es are useful for traditional local and "peephole"
639 optimizations. They may override the same :ref:`doInitialization(Module &)
640 <writing-an-llvm-pass-doInitialization-mod>` and :ref:`doFinalization(Module &)
641 <writing-an-llvm-pass-doFinalization-mod>` methods that :ref:`FunctionPass's
642 <writing-an-llvm-pass-FunctionPass>` have, but also have the following virtual
643 methods that may also be implemented:
645 The ``doInitialization(Function &)`` method
646 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
650 virtual bool doInitialization(Function &F);
652 The ``doInitialization`` method is allowed to do most of the things that
653 ``BasicBlockPass``\ es are not allowed to do, but that ``FunctionPass``\ es
654 can. The ``doInitialization`` method is designed to do simple initialization
655 that does not depend on the ``BasicBlock``\ s being processed. The
656 ``doInitialization`` method call is not scheduled to overlap with any other
657 pass executions (thus it should be very fast).
659 .. _writing-an-llvm-pass-runOnBasicBlock:
661 The ``runOnBasicBlock`` method
662 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
666 virtual bool runOnBasicBlock(BasicBlock &BB) = 0;
668 Override this function to do the work of the ``BasicBlockPass``. This function
669 is not allowed to inspect or modify basic blocks other than the parameter, and
670 are not allowed to modify the CFG. A ``true`` value must be returned if the
671 basic block is modified.
673 The ``doFinalization(Function &)`` method
674 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
678 virtual bool doFinalization(Function &F);
680 The ``doFinalization`` method is an infrequently used method that is called
681 when the pass framework has finished calling :ref:`runOnBasicBlock
682 <writing-an-llvm-pass-runOnBasicBlock>` for every ``BasicBlock`` in the program
683 being compiled. This can be used to perform per-function finalization.
685 The ``MachineFunctionPass`` class
686 ---------------------------------
688 A ``MachineFunctionPass`` is a part of the LLVM code generator that executes on
689 the machine-dependent representation of each LLVM function in the program.
691 Code generator passes are registered and initialized specially by
692 ``TargetMachine::addPassesToEmitFile`` and similar routines, so they cannot
693 generally be run from the :program:`opt` or :program:`bugpoint` commands.
695 A ``MachineFunctionPass`` is also a ``FunctionPass``, so all the restrictions
696 that apply to a ``FunctionPass`` also apply to it. ``MachineFunctionPass``\ es
697 also have additional restrictions. In particular, ``MachineFunctionPass``\ es
698 are not allowed to do any of the following:
700 #. Modify or create any LLVM IR ``Instruction``\ s, ``BasicBlock``\ s,
701 ``Argument``\ s, ``Function``\ s, ``GlobalVariable``\ s,
702 ``GlobalAlias``\ es, or ``Module``\ s.
703 #. Modify a ``MachineFunction`` other than the one currently being processed.
704 #. Maintain state across invocations of :ref:`runOnMachineFunction
705 <writing-an-llvm-pass-runOnMachineFunction>` (including global data).
707 .. _writing-an-llvm-pass-runOnMachineFunction:
709 The ``runOnMachineFunction(MachineFunction &MF)`` method
710 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
714 virtual bool runOnMachineFunction(MachineFunction &MF) = 0;
716 ``runOnMachineFunction`` can be considered the main entry point of a
717 ``MachineFunctionPass``; that is, you should override this method to do the
718 work of your ``MachineFunctionPass``.
720 The ``runOnMachineFunction`` method is called on every ``MachineFunction`` in a
721 ``Module``, so that the ``MachineFunctionPass`` may perform optimizations on
722 the machine-dependent representation of the function. If you want to get at
723 the LLVM ``Function`` for the ``MachineFunction`` you're working on, use
724 ``MachineFunction``'s ``getFunction()`` accessor method --- but remember, you
725 may not modify the LLVM ``Function`` or its contents from a
726 ``MachineFunctionPass``.
728 .. _writing-an-llvm-pass-registration:
733 In the :ref:`Hello World <writing-an-llvm-pass-basiccode>` example pass we
734 illustrated how pass registration works, and discussed some of the reasons that
735 it is used and what it does. Here we discuss how and why passes are
738 As we saw above, passes are registered with the ``RegisterPass`` template. The
739 template parameter is the name of the pass that is to be used on the command
740 line to specify that the pass should be added to a program (for example, with
741 :program:`opt` or :program:`bugpoint`). The first argument is the name of the
742 pass, which is to be used for the :option:`-help` output of programs, as well
743 as for debug output generated by the :option:`--debug-pass` option.
745 If you want your pass to be easily dumpable, you should implement the virtual
753 virtual void print(llvm::raw_ostream &O, const Module *M) const;
755 The ``print`` method must be implemented by "analyses" in order to print a
756 human readable version of the analysis results. This is useful for debugging
757 an analysis itself, as well as for other people to figure out how an analysis
758 works. Use the opt ``-analyze`` argument to invoke this method.
760 The ``llvm::raw_ostream`` parameter specifies the stream to write the results
761 on, and the ``Module`` parameter gives a pointer to the top level module of the
762 program that has been analyzed. Note however that this pointer may be ``NULL``
763 in certain circumstances (such as calling the ``Pass::dump()`` from a
764 debugger), so it should only be used to enhance debug output, it should not be
767 .. _writing-an-llvm-pass-interaction:
769 Specifying interactions between passes
770 --------------------------------------
772 One of the main responsibilities of the ``PassManager`` is to make sure that
773 passes interact with each other correctly. Because ``PassManager`` tries to
774 :ref:`optimize the execution of passes <writing-an-llvm-pass-passmanager>` it
775 must know how the passes interact with each other and what dependencies exist
776 between the various passes. To track this, each pass can declare the set of
777 passes that are required to be executed before the current pass, and the passes
778 which are invalidated by the current pass.
780 Typically this functionality is used to require that analysis results are
781 computed before your pass is run. Running arbitrary transformation passes can
782 invalidate the computed analysis results, which is what the invalidation set
783 specifies. If a pass does not implement the :ref:`getAnalysisUsage
784 <writing-an-llvm-pass-getAnalysisUsage>` method, it defaults to not having any
785 prerequisite passes, and invalidating **all** other passes.
787 .. _writing-an-llvm-pass-getAnalysisUsage:
789 The ``getAnalysisUsage`` method
790 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
794 virtual void getAnalysisUsage(AnalysisUsage &Info) const;
796 By implementing the ``getAnalysisUsage`` method, the required and invalidated
797 sets may be specified for your transformation. The implementation should fill
798 in the `AnalysisUsage
799 <http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html>`_ object with
800 information about which passes are required and not invalidated. To do this, a
801 pass may call any of the following methods on the ``AnalysisUsage`` object:
803 The ``AnalysisUsage::addRequired<>`` and ``AnalysisUsage::addRequiredTransitive<>`` methods
804 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
806 If your pass requires a previous pass to be executed (an analysis for example),
807 it can use one of these methods to arrange for it to be run before your pass.
808 LLVM has many different types of analyses and passes that can be required,
809 spanning the range from ``DominatorSet`` to ``BreakCriticalEdges``. Requiring
810 ``BreakCriticalEdges``, for example, guarantees that there will be no critical
811 edges in the CFG when your pass has been run.
813 Some analyses chain to other analyses to do their job. For example, an
814 `AliasAnalysis <AliasAnalysis>` implementation is required to :ref:`chain
815 <aliasanalysis-chaining>` to other alias analysis passes. In cases where
816 analyses chain, the ``addRequiredTransitive`` method should be used instead of
817 the ``addRequired`` method. This informs the ``PassManager`` that the
818 transitively required pass should be alive as long as the requiring pass is.
820 The ``AnalysisUsage::addPreserved<>`` method
821 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
823 One of the jobs of the ``PassManager`` is to optimize how and when analyses are
824 run. In particular, it attempts to avoid recomputing data unless it needs to.
825 For this reason, passes are allowed to declare that they preserve (i.e., they
826 don't invalidate) an existing analysis if it's available. For example, a
827 simple constant folding pass would not modify the CFG, so it can't possibly
828 affect the results of dominator analysis. By default, all passes are assumed
829 to invalidate all others.
831 The ``AnalysisUsage`` class provides several methods which are useful in
832 certain circumstances that are related to ``addPreserved``. In particular, the
833 ``setPreservesAll`` method can be called to indicate that the pass does not
834 modify the LLVM program at all (which is true for analyses), and the
835 ``setPreservesCFG`` method can be used by transformations that change
836 instructions in the program but do not modify the CFG or terminator
837 instructions (note that this property is implicitly set for
838 :ref:`BasicBlockPass <writing-an-llvm-pass-BasicBlockPass>`\ es).
840 ``addPreserved`` is particularly useful for transformations like
841 ``BreakCriticalEdges``. This pass knows how to update a small set of loop and
842 dominator related analyses if they exist, so it can preserve them, despite the
843 fact that it hacks on the CFG.
845 Example implementations of ``getAnalysisUsage``
846 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
850 // This example modifies the program, but does not modify the CFG
851 void LICM::getAnalysisUsage(AnalysisUsage &AU) const {
852 AU.setPreservesCFG();
853 AU.addRequired<LoopInfoWrapperPass>();
856 .. _writing-an-llvm-pass-getAnalysis:
858 The ``getAnalysis<>`` and ``getAnalysisIfAvailable<>`` methods
859 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
861 The ``Pass::getAnalysis<>`` method is automatically inherited by your class,
862 providing you with access to the passes that you declared that you required
863 with the :ref:`getAnalysisUsage <writing-an-llvm-pass-getAnalysisUsage>`
864 method. It takes a single template argument that specifies which pass class
865 you want, and returns a reference to that pass. For example:
869 bool LICM::runOnFunction(Function &F) {
870 LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
874 This method call returns a reference to the pass desired. You may get a
875 runtime assertion failure if you attempt to get an analysis that you did not
876 declare as required in your :ref:`getAnalysisUsage
877 <writing-an-llvm-pass-getAnalysisUsage>` implementation. This method can be
878 called by your ``run*`` method implementation, or by any other local method
879 invoked by your ``run*`` method.
881 A module level pass can use function level analysis info using this interface.
886 bool ModuleLevelPass::runOnModule(Module &M) {
888 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
892 In above example, ``runOnFunction`` for ``DominatorTree`` is called by pass
893 manager before returning a reference to the desired pass.
895 If your pass is capable of updating analyses if they exist (e.g.,
896 ``BreakCriticalEdges``, as described above), you can use the
897 ``getAnalysisIfAvailable`` method, which returns a pointer to the analysis if
898 it is active. For example:
902 if (DominatorSet *DS = getAnalysisIfAvailable<DominatorSet>()) {
903 // A DominatorSet is active. This code will update it.
906 Implementing Analysis Groups
907 ----------------------------
909 Now that we understand the basics of how passes are defined, how they are used,
910 and how they are required from other passes, it's time to get a little bit
911 fancier. All of the pass relationships that we have seen so far are very
912 simple: one pass depends on one other specific pass to be run before it can
913 run. For many applications, this is great, for others, more flexibility is
916 In particular, some analyses are defined such that there is a single simple
917 interface to the analysis results, but multiple ways of calculating them.
918 Consider alias analysis for example. The most trivial alias analysis returns
919 "may alias" for any alias query. The most sophisticated analysis a
920 flow-sensitive, context-sensitive interprocedural analysis that can take a
921 significant amount of time to execute (and obviously, there is a lot of room
922 between these two extremes for other implementations). To cleanly support
923 situations like this, the LLVM Pass Infrastructure supports the notion of
926 Analysis Group Concepts
927 ^^^^^^^^^^^^^^^^^^^^^^^
929 An Analysis Group is a single simple interface that may be implemented by
930 multiple different passes. Analysis Groups can be given human readable names
931 just like passes, but unlike passes, they need not derive from the ``Pass``
932 class. An analysis group may have one or more implementations, one of which is
933 the "default" implementation.
935 Analysis groups are used by client passes just like other passes are: the
936 ``AnalysisUsage::addRequired()`` and ``Pass::getAnalysis()`` methods. In order
937 to resolve this requirement, the :ref:`PassManager
938 <writing-an-llvm-pass-passmanager>` scans the available passes to see if any
939 implementations of the analysis group are available. If none is available, the
940 default implementation is created for the pass to use. All standard rules for
941 :ref:`interaction between passes <writing-an-llvm-pass-interaction>` still
944 Although :ref:`Pass Registration <writing-an-llvm-pass-registration>` is
945 optional for normal passes, all analysis group implementations must be
946 registered, and must use the :ref:`INITIALIZE_AG_PASS
947 <writing-an-llvm-pass-RegisterAnalysisGroup>` template to join the
948 implementation pool. Also, a default implementation of the interface **must**
949 be registered with :ref:`RegisterAnalysisGroup
950 <writing-an-llvm-pass-RegisterAnalysisGroup>`.
952 As a concrete example of an Analysis Group in action, consider the
953 `AliasAnalysis <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_
954 analysis group. The default implementation of the alias analysis interface
955 (the `basicaa <http://llvm.org/doxygen/structBasicAliasAnalysis.html>`_ pass)
956 just does a few simple checks that don't require significant analysis to
957 compute (such as: two different globals can never alias each other, etc).
958 Passes that use the `AliasAnalysis
959 <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_ interface (for
960 example the `gcse <http://llvm.org/doxygen/structGCSE.html>`_ pass), do not
961 care which implementation of alias analysis is actually provided, they just use
962 the designated interface.
964 From the user's perspective, commands work just like normal. Issuing the
965 command ``opt -gcse ...`` will cause the ``basicaa`` class to be instantiated
966 and added to the pass sequence. Issuing the command ``opt -somefancyaa -gcse
967 ...`` will cause the ``gcse`` pass to use the ``somefancyaa`` alias analysis
968 (which doesn't actually exist, it's just a hypothetical example) instead.
970 .. _writing-an-llvm-pass-RegisterAnalysisGroup:
972 Using ``RegisterAnalysisGroup``
973 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
975 The ``RegisterAnalysisGroup`` template is used to register the analysis group
976 itself, while the ``INITIALIZE_AG_PASS`` is used to add pass implementations to
977 the analysis group. First, an analysis group should be registered, with a
978 human readable name provided for it. Unlike registration of passes, there is
979 no command line argument to be specified for the Analysis Group Interface
980 itself, because it is "abstract":
984 static RegisterAnalysisGroup<AliasAnalysis> A("Alias Analysis");
986 Once the analysis is registered, passes can declare that they are valid
987 implementations of the interface by using the following code:
992 // Declare that we implement the AliasAnalysis interface
993 INITIALIZE_AG_PASS(FancyAA, AliasAnalysis , "somefancyaa",
994 "A more complex alias analysis implementation",
995 false, // Is CFG Only?
996 true, // Is Analysis?
997 false); // Is default Analysis Group implementation?
1000 This just shows a class ``FancyAA`` that uses the ``INITIALIZE_AG_PASS`` macro
1001 both to register and to "join" the `AliasAnalysis
1002 <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_ analysis group.
1003 Every implementation of an analysis group should join using this macro.
1008 // Declare that we implement the AliasAnalysis interface
1009 INITIALIZE_AG_PASS(BasicAA, AliasAnalysis, "basicaa",
1010 "Basic Alias Analysis (default AA impl)",
1011 false, // Is CFG Only?
1012 true, // Is Analysis?
1013 true); // Is default Analysis Group implementation?
1016 Here we show how the default implementation is specified (using the final
1017 argument to the ``INITIALIZE_AG_PASS`` template). There must be exactly one
1018 default implementation available at all times for an Analysis Group to be used.
1019 Only default implementation can derive from ``ImmutablePass``. Here we declare
1020 that the `BasicAliasAnalysis
1021 <http://llvm.org/doxygen/structBasicAliasAnalysis.html>`_ pass is the default
1022 implementation for the interface.
1027 The `Statistic <http://llvm.org/doxygen/Statistic_8h-source.html>`_ class is
1028 designed to be an easy way to expose various success metrics from passes.
1029 These statistics are printed at the end of a run, when the :option:`-stats`
1030 command line option is enabled on the command line. See the :ref:`Statistics
1031 section <Statistic>` in the Programmer's Manual for details.
1033 .. _writing-an-llvm-pass-passmanager:
1035 What PassManager does
1036 ---------------------
1038 The `PassManager <http://llvm.org/doxygen/PassManager_8h-source.html>`_ `class
1039 <http://llvm.org/doxygen/classllvm_1_1PassManager.html>`_ takes a list of
1040 passes, ensures their :ref:`prerequisites <writing-an-llvm-pass-interaction>`
1041 are set up correctly, and then schedules passes to run efficiently. All of the
1042 LLVM tools that run passes use the PassManager for execution of these passes.
1044 The PassManager does two main things to try to reduce the execution time of a
1047 #. **Share analysis results.** The ``PassManager`` attempts to avoid
1048 recomputing analysis results as much as possible. This means keeping track
1049 of which analyses are available already, which analyses get invalidated, and
1050 which analyses are needed to be run for a pass. An important part of work
1051 is that the ``PassManager`` tracks the exact lifetime of all analysis
1052 results, allowing it to :ref:`free memory
1053 <writing-an-llvm-pass-releaseMemory>` allocated to holding analysis results
1054 as soon as they are no longer needed.
1056 #. **Pipeline the execution of passes on the program.** The ``PassManager``
1057 attempts to get better cache and memory usage behavior out of a series of
1058 passes by pipelining the passes together. This means that, given a series
1059 of consecutive :ref:`FunctionPass <writing-an-llvm-pass-FunctionPass>`, it
1060 will execute all of the :ref:`FunctionPass
1061 <writing-an-llvm-pass-FunctionPass>` on the first function, then all of the
1062 :ref:`FunctionPasses <writing-an-llvm-pass-FunctionPass>` on the second
1063 function, etc... until the entire program has been run through the passes.
1065 This improves the cache behavior of the compiler, because it is only
1066 touching the LLVM program representation for a single function at a time,
1067 instead of traversing the entire program. It reduces the memory consumption
1068 of compiler, because, for example, only one `DominatorSet
1069 <http://llvm.org/doxygen/classllvm_1_1DominatorSet.html>`_ needs to be
1070 calculated at a time. This also makes it possible to implement some
1071 :ref:`interesting enhancements <writing-an-llvm-pass-SMP>` in the future.
1073 The effectiveness of the ``PassManager`` is influenced directly by how much
1074 information it has about the behaviors of the passes it is scheduling. For
1075 example, the "preserved" set is intentionally conservative in the face of an
1076 unimplemented :ref:`getAnalysisUsage <writing-an-llvm-pass-getAnalysisUsage>`
1077 method. Not implementing when it should be implemented will have the effect of
1078 not allowing any analysis results to live across the execution of your pass.
1080 The ``PassManager`` class exposes a ``--debug-pass`` command line options that
1081 is useful for debugging pass execution, seeing how things work, and diagnosing
1082 when you should be preserving more analyses than you currently are. (To get
1083 information about all of the variants of the ``--debug-pass`` option, just type
1084 "``opt -help-hidden``").
1086 By using the --debug-pass=Structure option, for example, we can see how our
1087 :ref:`Hello World <writing-an-llvm-pass-basiccode>` pass interacts with other
1088 passes. Lets try it out with the gcse and licm passes:
1090 .. code-block:: console
1092 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1094 Function Pass Manager
1095 Dominator Set Construction
1096 Immediate Dominators Construction
1097 Global Common Subexpression Elimination
1098 -- Immediate Dominators Construction
1099 -- Global Common Subexpression Elimination
1100 Natural Loop Construction
1101 Loop Invariant Code Motion
1102 -- Natural Loop Construction
1103 -- Loop Invariant Code Motion
1105 -- Dominator Set Construction
1110 This output shows us when passes are constructed and when the analysis results
1111 are known to be dead (prefixed with "``--``"). Here we see that GCSE uses
1112 dominator and immediate dominator information to do its job. The LICM pass
1113 uses natural loop information, which uses dominator sets, but not immediate
1114 dominators. Because immediate dominators are no longer useful after the GCSE
1115 pass, it is immediately destroyed. The dominator sets are then reused to
1116 compute natural loop information, which is then used by the LICM pass.
1118 After the LICM pass, the module verifier runs (which is automatically added by
1119 the :program:`opt` tool), which uses the dominator set to check that the
1120 resultant LLVM code is well formed. After it finishes, the dominator set
1121 information is destroyed, after being computed once, and shared by three
1124 Lets see how this changes when we run the :ref:`Hello World
1125 <writing-an-llvm-pass-basiccode>` pass in between the two passes:
1127 .. code-block:: console
1129 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1131 Function Pass Manager
1132 Dominator Set Construction
1133 Immediate Dominators Construction
1134 Global Common Subexpression Elimination
1135 -- Dominator Set Construction
1136 -- Immediate Dominators Construction
1137 -- Global Common Subexpression Elimination
1140 Dominator Set Construction
1141 Natural Loop Construction
1142 Loop Invariant Code Motion
1143 -- Natural Loop Construction
1144 -- Loop Invariant Code Motion
1146 -- Dominator Set Construction
1154 Here we see that the :ref:`Hello World <writing-an-llvm-pass-basiccode>` pass
1155 has killed the Dominator Set pass, even though it doesn't modify the code at
1156 all! To fix this, we need to add the following :ref:`getAnalysisUsage
1157 <writing-an-llvm-pass-getAnalysisUsage>` method to our pass:
1161 // We don't modify the program, so we preserve all analyses
1162 void getAnalysisUsage(AnalysisUsage &AU) const override {
1163 AU.setPreservesAll();
1166 Now when we run our pass, we get this output:
1168 .. code-block:: console
1170 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1171 Pass Arguments: -gcse -hello -licm
1173 Function Pass Manager
1174 Dominator Set Construction
1175 Immediate Dominators Construction
1176 Global Common Subexpression Elimination
1177 -- Immediate Dominators Construction
1178 -- Global Common Subexpression Elimination
1181 Natural Loop Construction
1182 Loop Invariant Code Motion
1183 -- Loop Invariant Code Motion
1184 -- Natural Loop Construction
1186 -- Dominator Set Construction
1194 Which shows that we don't accidentally invalidate dominator information
1195 anymore, and therefore do not have to compute it twice.
1197 .. _writing-an-llvm-pass-releaseMemory:
1199 The ``releaseMemory`` method
1200 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1204 virtual void releaseMemory();
1206 The ``PassManager`` automatically determines when to compute analysis results,
1207 and how long to keep them around for. Because the lifetime of the pass object
1208 itself is effectively the entire duration of the compilation process, we need
1209 some way to free analysis results when they are no longer useful. The
1210 ``releaseMemory`` virtual method is the way to do this.
1212 If you are writing an analysis or any other pass that retains a significant
1213 amount of state (for use by another pass which "requires" your pass and uses
1214 the :ref:`getAnalysis <writing-an-llvm-pass-getAnalysis>` method) you should
1215 implement ``releaseMemory`` to, well, release the memory allocated to maintain
1216 this internal state. This method is called after the ``run*`` method for the
1217 class, before the next call of ``run*`` in your pass.
1219 Registering dynamically loaded passes
1220 =====================================
1222 *Size matters* when constructing production quality tools using LLVM, both for
1223 the purposes of distribution, and for regulating the resident code size when
1224 running on the target system. Therefore, it becomes desirable to selectively
1225 use some passes, while omitting others and maintain the flexibility to change
1226 configurations later on. You want to be able to do all this, and, provide
1227 feedback to the user. This is where pass registration comes into play.
1229 The fundamental mechanisms for pass registration are the
1230 ``MachinePassRegistry`` class and subclasses of ``MachinePassRegistryNode``.
1232 An instance of ``MachinePassRegistry`` is used to maintain a list of
1233 ``MachinePassRegistryNode`` objects. This instance maintains the list and
1234 communicates additions and deletions to the command line interface.
1236 An instance of ``MachinePassRegistryNode`` subclass is used to maintain
1237 information provided about a particular pass. This information includes the
1238 command line name, the command help string and the address of the function used
1239 to create an instance of the pass. A global static constructor of one of these
1240 instances *registers* with a corresponding ``MachinePassRegistry``, the static
1241 destructor *unregisters*. Thus a pass that is statically linked in the tool
1242 will be registered at start up. A dynamically loaded pass will register on
1243 load and unregister at unload.
1245 Using existing registries
1246 -------------------------
1248 There are predefined registries to track instruction scheduling
1249 (``RegisterScheduler``) and register allocation (``RegisterRegAlloc``) machine
1250 passes. Here we will describe how to *register* a register allocator machine
1253 Implement your register allocator machine pass. In your register allocator
1254 ``.cpp`` file add the following include:
1258 #include "llvm/CodeGen/RegAllocRegistry.h"
1260 Also in your register allocator ``.cpp`` file, define a creator function in the
1265 FunctionPass *createMyRegisterAllocator() {
1266 return new MyRegisterAllocator();
1269 Note that the signature of this function should match the type of
1270 ``RegisterRegAlloc::FunctionPassCtor``. In the same file add the "installing"
1271 declaration, in the form:
1275 static RegisterRegAlloc myRegAlloc("myregalloc",
1276 "my register allocator help string",
1277 createMyRegisterAllocator);
1279 Note the two spaces prior to the help string produces a tidy result on the
1280 :option:`-help` query.
1282 .. code-block:: console
1286 -regalloc - Register allocator to use (default=linearscan)
1287 =linearscan - linear scan register allocator
1288 =local - local register allocator
1289 =simple - simple register allocator
1290 =myregalloc - my register allocator help string
1293 And that's it. The user is now free to use ``-regalloc=myregalloc`` as an
1294 option. Registering instruction schedulers is similar except use the
1295 ``RegisterScheduler`` class. Note that the
1296 ``RegisterScheduler::FunctionPassCtor`` is significantly different from
1297 ``RegisterRegAlloc::FunctionPassCtor``.
1299 To force the load/linking of your register allocator into the
1300 :program:`llc`/:program:`lli` tools, add your creator function's global
1301 declaration to ``Passes.h`` and add a "pseudo" call line to
1302 ``llvm/Codegen/LinkAllCodegenComponents.h``.
1304 Creating new registries
1305 -----------------------
1307 The easiest way to get started is to clone one of the existing registries; we
1308 recommend ``llvm/CodeGen/RegAllocRegistry.h``. The key things to modify are
1309 the class name and the ``FunctionPassCtor`` type.
1311 Then you need to declare the registry. Example: if your pass registry is
1312 ``RegisterMyPasses`` then define:
1316 MachinePassRegistry RegisterMyPasses::Registry;
1318 And finally, declare the command line option for your passes. Example:
1322 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1323 RegisterPassParser<RegisterMyPasses> >
1325 cl::init(&createDefaultMyPass),
1326 cl::desc("my pass option help"));
1328 Here the command option is "``mypass``", with ``createDefaultMyPass`` as the
1331 Using GDB with dynamically loaded passes
1332 ----------------------------------------
1334 Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1335 should be. First of all, you can't set a breakpoint in a shared object that
1336 has not been loaded yet, and second of all there are problems with inlined
1337 functions in shared objects. Here are some suggestions to debugging your pass
1340 For sake of discussion, I'm going to assume that you are debugging a
1341 transformation invoked by :program:`opt`, although nothing described here
1344 Setting a breakpoint in your pass
1345 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1347 First thing you do is start gdb on the opt process:
1349 .. code-block:: console
1353 Copyright 2000 Free Software Foundation, Inc.
1354 GDB is free software, covered by the GNU General Public License, and you are
1355 welcome to change it and/or distribute copies of it under certain conditions.
1356 Type "show copying" to see the conditions.
1357 There is absolutely no warranty for GDB. Type "show warranty" for details.
1358 This GDB was configured as "sparc-sun-solaris2.6"...
1361 Note that :program:`opt` has a lot of debugging information in it, so it takes
1362 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1363 (the shared object isn't loaded until runtime), we must execute the process,
1364 and have it stop before it invokes our pass, but after it has loaded the shared
1365 object. The most foolproof way of doing this is to set a breakpoint in
1366 ``PassManager::run`` and then run the process with the arguments you want:
1368 .. code-block:: console
1370 $ (gdb) break llvm::PassManager::run
1371 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1372 (gdb) run test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1373 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1374 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1375 70 bool PassManager::run(Module &M) { return PM->run(M); }
1378 Once the :program:`opt` stops in the ``PassManager::run`` method you are now
1379 free to set breakpoints in your pass so that you can trace through execution or
1380 do other standard debugging stuff.
1382 Miscellaneous Problems
1383 ^^^^^^^^^^^^^^^^^^^^^^
1385 Once you have the basics down, there are a couple of problems that GDB has,
1386 some with solutions, some without.
1388 * Inline functions have bogus stack information. In general, GDB does a pretty
1389 good job getting stack traces and stepping through inline functions. When a
1390 pass is dynamically loaded however, it somehow completely loses this
1391 capability. The only solution I know of is to de-inline a function (move it
1392 from the body of a class to a ``.cpp`` file).
1394 * Restarting the program breaks breakpoints. After following the information
1395 above, you have succeeded in getting some breakpoints planted in your pass.
1396 Nex thing you know, you restart the program (i.e., you type "``run``" again),
1397 and you start getting errors about breakpoints being unsettable. The only
1398 way I have found to "fix" this problem is to delete the breakpoints that are
1399 already set in your pass, run the program, and re-set the breakpoints once
1400 execution stops in ``PassManager::run``.
1402 Hopefully these tips will help with common case debugging situations. If you'd
1403 like to contribute some tips of your own, just contact `Chris
1404 <mailto:sabre@nondot.org>`_.
1406 Future extensions planned
1407 -------------------------
1409 Although the LLVM Pass Infrastructure is very capable as it stands, and does
1410 some nifty stuff, there are things we'd like to add in the future. Here is
1413 .. _writing-an-llvm-pass-SMP:
1418 Multiple CPU machines are becoming more common and compilation can never be
1419 fast enough: obviously we should allow for a multithreaded compiler. Because
1420 of the semantics defined for passes above (specifically they cannot maintain
1421 state across invocations of their ``run*`` methods), a nice clean way to
1422 implement a multithreaded compiler would be for the ``PassManager`` class to
1423 create multiple instances of each pass object, and allow the separate instances
1424 to be hacking on different parts of the program at the same time.
1426 This implementation would prevent each of the passes from having to implement
1427 multithreaded constructs, requiring only the LLVM core to have locking in a few
1428 places (for global resources). Although this is a simple extension, we simply
1429 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1430 Despite that, we have kept the LLVM passes SMP ready, and you should too.