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 .. FIXME: Why does this recommend to build in-tree?
52 First, configure and build LLVM. This needs to be done directly inside the
53 LLVM source tree rather than in a separate objects directory. Next, you need
54 to create a new directory somewhere in the LLVM source base. For this example,
55 we'll assume that you made ``lib/Transforms/Hello``. Finally, you must set up
56 a build script (``Makefile``) that will compile the source code for the new
57 pass. To do this, copy the following into ``Makefile``:
61 # Makefile for hello pass
63 # Path to top level of LLVM hierarchy
66 # Name of the library to build
69 # Make the shared library become a loadable module so the tools can
70 # dlopen/dlsym on the resulting library.
73 # Include the makefile implementation stuff
74 include $(LEVEL)/Makefile.common
76 This makefile specifies that all of the ``.cpp`` files in the current directory
77 are to be compiled and linked together into a shared object
78 ``$(LEVEL)/Debug+Asserts/lib/Hello.so`` that can be dynamically loaded by the
79 :program:`opt` or :program:`bugpoint` tools via their :option:`-load` options.
80 If your operating system uses a suffix other than ``.so`` (such as Windows or Mac
81 OS X), the appropriate extension will be used.
83 If you are used CMake to build LLVM, see :ref:`cmake-out-of-source-pass`.
85 Now that we have the build scripts set up, we just need to write the code for
88 .. _writing-an-llvm-pass-basiccode:
93 Now that we have a way to compile our new pass, we just have to write it.
98 #include "llvm/Pass.h"
99 #include "llvm/IR/Function.h"
100 #include "llvm/Support/raw_ostream.h"
102 Which are needed because we are writing a `Pass
103 <http://llvm.org/doxygen/classllvm_1_1Pass.html>`_, we are operating on
104 `Function <http://llvm.org/doxygen/classllvm_1_1Function.html>`_\ s, and we will
105 be doing some printing.
111 using namespace llvm;
113 ... which is required because the functions from the include files live in the
122 ... which starts out an anonymous namespace. Anonymous namespaces are to C++
123 what the "``static``" keyword is to C (at global scope). It makes the things
124 declared inside of the anonymous namespace visible only to the current file.
125 If you're not familiar with them, consult a decent C++ book for more
128 Next, we declare our pass itself:
132 struct Hello : public FunctionPass {
134 This declares a "``Hello``" class that is a subclass of :ref:`FunctionPass
135 <writing-an-llvm-pass-FunctionPass>`. The different builtin pass subclasses
136 are described in detail :ref:`later <writing-an-llvm-pass-pass-classes>`, but
137 for now, know that ``FunctionPass`` operates on a function at a time.
142 Hello() : FunctionPass(ID) {}
144 This declares pass identifier used by LLVM to identify pass. This allows LLVM
145 to avoid using expensive C++ runtime information.
149 virtual bool runOnFunction(Function &F) {
151 errs().write_escaped(F.getName()) << "\n";
154 }; // end of struct Hello
155 } // end of anonymous namespace
157 We declare a :ref:`runOnFunction <writing-an-llvm-pass-runOnFunction>` method,
158 which overrides an abstract virtual method inherited from :ref:`FunctionPass
159 <writing-an-llvm-pass-FunctionPass>`. This is where we are supposed to do our
160 thing, so we just print out our message with the name of each function.
166 We initialize pass ID here. LLVM uses ID's address to identify a pass, so
167 initialization value is not important.
171 static RegisterPass<Hello> X("hello", "Hello World Pass",
172 false /* Only looks at CFG */,
173 false /* Analysis Pass */);
175 Lastly, we :ref:`register our class <writing-an-llvm-pass-registration>`
176 ``Hello``, giving it a command line argument "``hello``", and a name "Hello
177 World Pass". The last two arguments describe its behavior: if a pass walks CFG
178 without modifying it then the third argument is set to ``true``; if a pass is
179 an analysis pass, for example dominator tree pass, then ``true`` is supplied as
182 As a whole, the ``.cpp`` file looks like:
186 #include "llvm/Pass.h"
187 #include "llvm/IR/Function.h"
188 #include "llvm/Support/raw_ostream.h"
190 using namespace llvm;
193 struct Hello : public FunctionPass {
195 Hello() : FunctionPass(ID) {}
197 virtual bool runOnFunction(Function &F) {
199 errs().write_escaped(F.getName()) << '\n';
206 static RegisterPass<Hello> X("hello", "Hello World Pass", false, false);
208 Now that it's all together, compile the file with a simple "``gmake``" command
209 in the local directory and you should get a new file
210 "``Debug+Asserts/lib/Hello.so``" under the top level directory of the LLVM
211 source tree (not in the local directory). Note that everything in this file is
212 contained in an anonymous namespace --- this reflects the fact that passes
213 are self contained units that do not need external interfaces (although they
214 can have them) to be useful.
216 Running a pass with ``opt``
217 ---------------------------
219 Now that you have a brand new shiny shared object file, we can use the
220 :program:`opt` command to run an LLVM program through your pass. Because you
221 registered your pass with ``RegisterPass``, you will be able to use the
222 :program:`opt` tool to access it, once loaded.
224 To test it, follow the example at the end of the :doc:`GettingStarted` to
225 compile "Hello World" to LLVM. We can now run the bitcode file (hello.bc) for
226 the program through our transformation like this (or course, any bitcode file
229 .. code-block:: console
231 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello < hello.bc > /dev/null
236 The :option:`-load` option specifies that :program:`opt` should load your pass
237 as a shared object, which makes "``-hello``" a valid command line argument
238 (which is one reason you need to :ref:`register your pass
239 <writing-an-llvm-pass-registration>`). Because the Hello pass does not modify
240 the program in any interesting way, we just throw away the result of
241 :program:`opt` (sending it to ``/dev/null``).
243 To see what happened to the other string you registered, try running
244 :program:`opt` with the :option:`-help` option:
246 .. code-block:: console
248 $ opt -load ../../../Debug+Asserts/lib/Hello.so -help
249 OVERVIEW: llvm .bc -> .bc modular optimizer
251 USAGE: opt [options] <input bitcode>
254 Optimizations available:
256 -globalopt - Global Variable Optimizer
257 -globalsmodref-aa - Simple mod/ref analysis for globals
258 -gvn - Global Value Numbering
259 -hello - Hello World Pass
260 -indvars - Induction Variable Simplification
261 -inline - Function Integration/Inlining
264 The pass name gets added as the information string for your pass, giving some
265 documentation to users of :program:`opt`. Now that you have a working pass,
266 you would go ahead and make it do the cool transformations you want. Once you
267 get it all working and tested, it may become useful to find out how fast your
268 pass is. The :ref:`PassManager <writing-an-llvm-pass-passmanager>` provides a
269 nice command line option (:option:`--time-passes`) that allows you to get
270 information about the execution time of your pass along with the other passes
271 you queue up. For example:
273 .. code-block:: console
275 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
279 ===============================================================================
280 ... Pass execution timing report ...
281 ===============================================================================
282 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
284 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
285 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bitcode Writer
286 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
287 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
288 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass
289 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
291 As you can see, our implementation above is pretty fast. The additional
292 passes listed are automatically inserted by the :program:`opt` tool to verify
293 that the LLVM emitted by your pass is still valid and well formed LLVM, which
294 hasn't been broken somehow.
296 Now that you have seen the basics of the mechanics behind passes, we can talk
297 about some more details of how they work and how to use them.
299 .. _writing-an-llvm-pass-pass-classes:
301 Pass classes and requirements
302 =============================
304 One of the first things that you should do when designing a new pass is to
305 decide what class you should subclass for your pass. The :ref:`Hello World
306 <writing-an-llvm-pass-basiccode>` example uses the :ref:`FunctionPass
307 <writing-an-llvm-pass-FunctionPass>` class for its implementation, but we did
308 not discuss why or when this should occur. Here we talk about the classes
309 available, from the most general to the most specific.
311 When choosing a superclass for your ``Pass``, you should choose the **most
312 specific** class possible, while still being able to meet the requirements
313 listed. This gives the LLVM Pass Infrastructure information necessary to
314 optimize how passes are run, so that the resultant compiler isn't unnecessarily
317 The ``ImmutablePass`` class
318 ---------------------------
320 The most plain and boring type of pass is the "`ImmutablePass
321 <http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html>`_" class. This pass
322 type is used for passes that do not have to be run, do not change state, and
323 never need to be updated. This is not a normal type of transformation or
324 analysis, but can provide information about the current compiler configuration.
326 Although this pass class is very infrequently used, it is important for
327 providing information about the current target machine being compiled for, and
328 other static information that can affect the various transformations.
330 ``ImmutablePass``\ es never invalidate other transformations, are never
331 invalidated, and are never "run".
333 .. _writing-an-llvm-pass-ModulePass:
335 The ``ModulePass`` class
336 ------------------------
338 The `ModulePass <http://llvm.org/doxygen/classllvm_1_1ModulePass.html>`_ class
339 is the most general of all superclasses that you can use. Deriving from
340 ``ModulePass`` indicates that your pass uses the entire program as a unit,
341 referring to function bodies in no predictable order, or adding and removing
342 functions. Because nothing is known about the behavior of ``ModulePass``
343 subclasses, no optimization can be done for their execution.
345 A module pass can use function level passes (e.g. dominators) using the
346 ``getAnalysis`` interface ``getAnalysis<DominatorTree>(llvm::Function *)`` to
347 provide the function to retrieve analysis result for, if the function pass does
348 not require any module or immutable passes. Note that this can only be done
349 for functions for which the analysis ran, e.g. in the case of dominators you
350 should only ask for the ``DominatorTree`` for function definitions, not
353 To write a correct ``ModulePass`` subclass, derive from ``ModulePass`` and
354 overload the ``runOnModule`` method with the following signature:
356 The ``runOnModule`` method
357 ^^^^^^^^^^^^^^^^^^^^^^^^^^
361 virtual bool runOnModule(Module &M) = 0;
363 The ``runOnModule`` method performs the interesting work of the pass. It
364 should return ``true`` if the module was modified by the transformation and
367 .. _writing-an-llvm-pass-CallGraphSCCPass:
369 The ``CallGraphSCCPass`` class
370 ------------------------------
372 The `CallGraphSCCPass
373 <http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html>`_ is used by
374 passes that need to traverse the program bottom-up on the call graph (callees
375 before callers). Deriving from ``CallGraphSCCPass`` provides some mechanics
376 for building and traversing the ``CallGraph``, but also allows the system to
377 optimize execution of ``CallGraphSCCPass``\ es. If your pass meets the
378 requirements outlined below, and doesn't meet the requirements of a
379 :ref:`FunctionPass <writing-an-llvm-pass-FunctionPass>` or :ref:`BasicBlockPass
380 <writing-an-llvm-pass-BasicBlockPass>`, you should derive from
381 ``CallGraphSCCPass``.
383 ``TODO``: explain briefly what SCC, Tarjan's algo, and B-U mean.
385 To be explicit, CallGraphSCCPass subclasses are:
387 #. ... *not allowed* to inspect or modify any ``Function``\ s other than those
388 in the current SCC and the direct callers and direct callees of the SCC.
389 #. ... *required* to preserve the current ``CallGraph`` object, updating it to
390 reflect any changes made to the program.
391 #. ... *not allowed* to add or remove SCC's from the current Module, though
392 they may change the contents of an SCC.
393 #. ... *allowed* to add or remove global variables from the current Module.
394 #. ... *allowed* to maintain state across invocations of :ref:`runOnSCC
395 <writing-an-llvm-pass-runOnSCC>` (including global data).
397 Implementing a ``CallGraphSCCPass`` is slightly tricky in some cases because it
398 has to handle SCCs with more than one node in it. All of the virtual methods
399 described below should return ``true`` if they modified the program, or
400 ``false`` if they didn't.
402 The ``doInitialization(CallGraph &)`` method
403 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
407 virtual bool doInitialization(CallGraph &CG);
409 The ``doInitialization`` method is allowed to do most of the things that
410 ``CallGraphSCCPass``\ es are not allowed to do. They can add and remove
411 functions, get pointers to functions, etc. The ``doInitialization`` method is
412 designed to do simple initialization type of stuff that does not depend on the
413 SCCs being processed. The ``doInitialization`` method call is not scheduled to
414 overlap with any other pass executions (thus it should be very fast).
416 .. _writing-an-llvm-pass-runOnSCC:
418 The ``runOnSCC`` method
419 ^^^^^^^^^^^^^^^^^^^^^^^
423 virtual bool runOnSCC(CallGraphSCC &SCC) = 0;
425 The ``runOnSCC`` method performs the interesting work of the pass, and should
426 return ``true`` if the module was modified by the transformation, ``false``
429 The ``doFinalization(CallGraph &)`` method
430 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
434 virtual bool doFinalization(CallGraph &CG);
436 The ``doFinalization`` method is an infrequently used method that is called
437 when the pass framework has finished calling :ref:`runOnFunction
438 <writing-an-llvm-pass-runOnFunction>` for every function in the program being
441 .. _writing-an-llvm-pass-FunctionPass:
443 The ``FunctionPass`` class
444 --------------------------
446 In contrast to ``ModulePass`` subclasses, `FunctionPass
447 <http://llvm.org/doxygen/classllvm_1_1Pass.html>`_ subclasses do have a
448 predictable, local behavior that can be expected by the system. All
449 ``FunctionPass`` execute on each function in the program independent of all of
450 the other functions in the program. ``FunctionPass``\ es do not require that
451 they are executed in a particular order, and ``FunctionPass``\ es do not modify
454 To be explicit, ``FunctionPass`` subclasses are not allowed to:
456 #. Inspect or modify a ``Function`` other than the one currently being processed.
457 #. Add or remove ``Function``\ s from the current ``Module``.
458 #. Add or remove global variables from the current ``Module``.
459 #. Maintain state across invocations of:ref:`runOnFunction
460 <writing-an-llvm-pass-runOnFunction>` (including global data).
462 Implementing a ``FunctionPass`` is usually straightforward (See the :ref:`Hello
463 World <writing-an-llvm-pass-basiccode>` pass for example).
464 ``FunctionPass``\ es may overload three virtual methods to do their work. All
465 of these methods should return ``true`` if they modified the program, or
466 ``false`` if they didn't.
468 .. _writing-an-llvm-pass-doInitialization-mod:
470 The ``doInitialization(Module &)`` method
471 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
475 virtual bool doInitialization(Module &M);
477 The ``doInitialization`` method is allowed to do most of the things that
478 ``FunctionPass``\ es are not allowed to do. They can add and remove functions,
479 get pointers to functions, etc. The ``doInitialization`` method is designed to
480 do simple initialization type of stuff that does not depend on the functions
481 being processed. The ``doInitialization`` method call is not scheduled to
482 overlap with any other pass executions (thus it should be very fast).
484 A good example of how this method should be used is the `LowerAllocations
485 <http://llvm.org/doxygen/LowerAllocations_8cpp-source.html>`_ pass. This pass
486 converts ``malloc`` and ``free`` instructions into platform dependent
487 ``malloc()`` and ``free()`` function calls. It uses the ``doInitialization``
488 method to get a reference to the ``malloc`` and ``free`` functions that it
489 needs, adding prototypes to the module if necessary.
491 .. _writing-an-llvm-pass-runOnFunction:
493 The ``runOnFunction`` method
494 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
498 virtual bool runOnFunction(Function &F) = 0;
500 The ``runOnFunction`` method must be implemented by your subclass to do the
501 transformation or analysis work of your pass. As usual, a ``true`` value
502 should be returned if the function is modified.
504 .. _writing-an-llvm-pass-doFinalization-mod:
506 The ``doFinalization(Module &)`` method
507 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
511 virtual bool doFinalization(Module &M);
513 The ``doFinalization`` method is an infrequently used method that is called
514 when the pass framework has finished calling :ref:`runOnFunction
515 <writing-an-llvm-pass-runOnFunction>` for every function in the program being
518 .. _writing-an-llvm-pass-LoopPass:
520 The ``LoopPass`` class
521 ----------------------
523 All ``LoopPass`` execute on each loop in the function independent of all of the
524 other loops in the function. ``LoopPass`` processes loops in loop nest order
525 such that outer most loop is processed last.
527 ``LoopPass`` subclasses are allowed to update loop nest using ``LPPassManager``
528 interface. Implementing a loop pass is usually straightforward.
529 ``LoopPass``\ es may overload three virtual methods to do their work. All
530 these methods should return ``true`` if they modified the program, or ``false``
533 The ``doInitialization(Loop *, LPPassManager &)`` method
534 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
538 virtual bool doInitialization(Loop *, LPPassManager &LPM);
540 The ``doInitialization`` method is designed to do simple initialization type of
541 stuff that does not depend on the functions being processed. The
542 ``doInitialization`` method call is not scheduled to overlap with any other
543 pass executions (thus it should be very fast). ``LPPassManager`` interface
544 should be used to access ``Function`` or ``Module`` level analysis information.
546 .. _writing-an-llvm-pass-runOnLoop:
548 The ``runOnLoop`` method
549 ^^^^^^^^^^^^^^^^^^^^^^^^
553 virtual bool runOnLoop(Loop *, LPPassManager &LPM) = 0;
555 The ``runOnLoop`` method must be implemented by your subclass to do the
556 transformation or analysis work of your pass. As usual, a ``true`` value
557 should be returned if the function is modified. ``LPPassManager`` interface
558 should be used to update loop nest.
560 The ``doFinalization()`` method
561 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
565 virtual bool doFinalization();
567 The ``doFinalization`` method is an infrequently used method that is called
568 when the pass framework has finished calling :ref:`runOnLoop
569 <writing-an-llvm-pass-runOnLoop>` for every loop in the program being compiled.
571 .. _writing-an-llvm-pass-RegionPass:
573 The ``RegionPass`` class
574 ------------------------
576 ``RegionPass`` is similar to :ref:`LoopPass <writing-an-llvm-pass-LoopPass>`,
577 but executes on each single entry single exit region in the function.
578 ``RegionPass`` processes regions in nested order such that the outer most
579 region is processed last.
581 ``RegionPass`` subclasses are allowed to update the region tree by using the
582 ``RGPassManager`` interface. You may overload three virtual methods of
583 ``RegionPass`` to implement your own region pass. All these methods should
584 return ``true`` if they modified the program, or ``false`` if they did not.
586 The ``doInitialization(Region *, RGPassManager &)`` method
587 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
591 virtual bool doInitialization(Region *, RGPassManager &RGM);
593 The ``doInitialization`` method is designed to do simple initialization type of
594 stuff that does not depend on the functions being processed. The
595 ``doInitialization`` method call is not scheduled to overlap with any other
596 pass executions (thus it should be very fast). ``RPPassManager`` interface
597 should be used to access ``Function`` or ``Module`` level analysis information.
599 .. _writing-an-llvm-pass-runOnRegion:
601 The ``runOnRegion`` method
602 ^^^^^^^^^^^^^^^^^^^^^^^^^^
606 virtual bool runOnRegion(Region *, RGPassManager &RGM) = 0;
608 The ``runOnRegion`` method must be implemented by your subclass to do the
609 transformation or analysis work of your pass. As usual, a true value should be
610 returned if the region is modified. ``RGPassManager`` interface should be used to
613 The ``doFinalization()`` method
614 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
618 virtual bool doFinalization();
620 The ``doFinalization`` method is an infrequently used method that is called
621 when the pass framework has finished calling :ref:`runOnRegion
622 <writing-an-llvm-pass-runOnRegion>` for every region in the program being
625 .. _writing-an-llvm-pass-BasicBlockPass:
627 The ``BasicBlockPass`` class
628 ----------------------------
630 ``BasicBlockPass``\ es are just like :ref:`FunctionPass's
631 <writing-an-llvm-pass-FunctionPass>` , except that they must limit their scope
632 of inspection and modification to a single basic block at a time. As such,
633 they are **not** allowed to do any of the following:
635 #. Modify or inspect any basic blocks outside of the current one.
636 #. Maintain state across invocations of :ref:`runOnBasicBlock
637 <writing-an-llvm-pass-runOnBasicBlock>`.
638 #. Modify the control flow graph (by altering terminator instructions)
639 #. Any of the things forbidden for :ref:`FunctionPasses
640 <writing-an-llvm-pass-FunctionPass>`.
642 ``BasicBlockPass``\ es are useful for traditional local and "peephole"
643 optimizations. They may override the same :ref:`doInitialization(Module &)
644 <writing-an-llvm-pass-doInitialization-mod>` and :ref:`doFinalization(Module &)
645 <writing-an-llvm-pass-doFinalization-mod>` methods that :ref:`FunctionPass's
646 <writing-an-llvm-pass-FunctionPass>` have, but also have the following virtual
647 methods that may also be implemented:
649 The ``doInitialization(Function &)`` method
650 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
654 virtual bool doInitialization(Function &F);
656 The ``doInitialization`` method is allowed to do most of the things that
657 ``BasicBlockPass``\ es are not allowed to do, but that ``FunctionPass``\ es
658 can. The ``doInitialization`` method is designed to do simple initialization
659 that does not depend on the ``BasicBlock``\ s being processed. The
660 ``doInitialization`` method call is not scheduled to overlap with any other
661 pass executions (thus it should be very fast).
663 .. _writing-an-llvm-pass-runOnBasicBlock:
665 The ``runOnBasicBlock`` method
666 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
670 virtual bool runOnBasicBlock(BasicBlock &BB) = 0;
672 Override this function to do the work of the ``BasicBlockPass``. This function
673 is not allowed to inspect or modify basic blocks other than the parameter, and
674 are not allowed to modify the CFG. A ``true`` value must be returned if the
675 basic block is modified.
677 The ``doFinalization(Function &)`` method
678 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
682 virtual bool doFinalization(Function &F);
684 The ``doFinalization`` method is an infrequently used method that is called
685 when the pass framework has finished calling :ref:`runOnBasicBlock
686 <writing-an-llvm-pass-runOnBasicBlock>` for every ``BasicBlock`` in the program
687 being compiled. This can be used to perform per-function finalization.
689 The ``MachineFunctionPass`` class
690 ---------------------------------
692 A ``MachineFunctionPass`` is a part of the LLVM code generator that executes on
693 the machine-dependent representation of each LLVM function in the program.
695 Code generator passes are registered and initialized specially by
696 ``TargetMachine::addPassesToEmitFile`` and similar routines, so they cannot
697 generally be run from the :program:`opt` or :program:`bugpoint` commands.
699 A ``MachineFunctionPass`` is also a ``FunctionPass``, so all the restrictions
700 that apply to a ``FunctionPass`` also apply to it. ``MachineFunctionPass``\ es
701 also have additional restrictions. In particular, ``MachineFunctionPass``\ es
702 are not allowed to do any of the following:
704 #. Modify or create any LLVM IR ``Instruction``\ s, ``BasicBlock``\ s,
705 ``Argument``\ s, ``Function``\ s, ``GlobalVariable``\ s,
706 ``GlobalAlias``\ es, or ``Module``\ s.
707 #. Modify a ``MachineFunction`` other than the one currently being processed.
708 #. Maintain state across invocations of :ref:`runOnMachineFunction
709 <writing-an-llvm-pass-runOnMachineFunction>` (including global data).
711 .. _writing-an-llvm-pass-runOnMachineFunction:
713 The ``runOnMachineFunction(MachineFunction &MF)`` method
714 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
718 virtual bool runOnMachineFunction(MachineFunction &MF) = 0;
720 ``runOnMachineFunction`` can be considered the main entry point of a
721 ``MachineFunctionPass``; that is, you should override this method to do the
722 work of your ``MachineFunctionPass``.
724 The ``runOnMachineFunction`` method is called on every ``MachineFunction`` in a
725 ``Module``, so that the ``MachineFunctionPass`` may perform optimizations on
726 the machine-dependent representation of the function. If you want to get at
727 the LLVM ``Function`` for the ``MachineFunction`` you're working on, use
728 ``MachineFunction``'s ``getFunction()`` accessor method --- but remember, you
729 may not modify the LLVM ``Function`` or its contents from a
730 ``MachineFunctionPass``.
732 .. _writing-an-llvm-pass-registration:
737 In the :ref:`Hello World <writing-an-llvm-pass-basiccode>` example pass we
738 illustrated how pass registration works, and discussed some of the reasons that
739 it is used and what it does. Here we discuss how and why passes are
742 As we saw above, passes are registered with the ``RegisterPass`` template. The
743 template parameter is the name of the pass that is to be used on the command
744 line to specify that the pass should be added to a program (for example, with
745 :program:`opt` or :program:`bugpoint`). The first argument is the name of the
746 pass, which is to be used for the :option:`-help` output of programs, as well
747 as for debug output generated by the :option:`--debug-pass` option.
749 If you want your pass to be easily dumpable, you should implement the virtual
757 virtual void print(llvm::raw_ostream &O, const Module *M) const;
759 The ``print`` method must be implemented by "analyses" in order to print a
760 human readable version of the analysis results. This is useful for debugging
761 an analysis itself, as well as for other people to figure out how an analysis
762 works. Use the opt ``-analyze`` argument to invoke this method.
764 The ``llvm::raw_ostream`` parameter specifies the stream to write the results
765 on, and the ``Module`` parameter gives a pointer to the top level module of the
766 program that has been analyzed. Note however that this pointer may be ``NULL``
767 in certain circumstances (such as calling the ``Pass::dump()`` from a
768 debugger), so it should only be used to enhance debug output, it should not be
771 .. _writing-an-llvm-pass-interaction:
773 Specifying interactions between passes
774 --------------------------------------
776 One of the main responsibilities of the ``PassManager`` is to make sure that
777 passes interact with each other correctly. Because ``PassManager`` tries to
778 :ref:`optimize the execution of passes <writing-an-llvm-pass-passmanager>` it
779 must know how the passes interact with each other and what dependencies exist
780 between the various passes. To track this, each pass can declare the set of
781 passes that are required to be executed before the current pass, and the passes
782 which are invalidated by the current pass.
784 Typically this functionality is used to require that analysis results are
785 computed before your pass is run. Running arbitrary transformation passes can
786 invalidate the computed analysis results, which is what the invalidation set
787 specifies. If a pass does not implement the :ref:`getAnalysisUsage
788 <writing-an-llvm-pass-getAnalysisUsage>` method, it defaults to not having any
789 prerequisite passes, and invalidating **all** other passes.
791 .. _writing-an-llvm-pass-getAnalysisUsage:
793 The ``getAnalysisUsage`` method
794 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
798 virtual void getAnalysisUsage(AnalysisUsage &Info) const;
800 By implementing the ``getAnalysisUsage`` method, the required and invalidated
801 sets may be specified for your transformation. The implementation should fill
802 in the `AnalysisUsage
803 <http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html>`_ object with
804 information about which passes are required and not invalidated. To do this, a
805 pass may call any of the following methods on the ``AnalysisUsage`` object:
807 The ``AnalysisUsage::addRequired<>`` and ``AnalysisUsage::addRequiredTransitive<>`` methods
808 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
810 If your pass requires a previous pass to be executed (an analysis for example),
811 it can use one of these methods to arrange for it to be run before your pass.
812 LLVM has many different types of analyses and passes that can be required,
813 spanning the range from ``DominatorSet`` to ``BreakCriticalEdges``. Requiring
814 ``BreakCriticalEdges``, for example, guarantees that there will be no critical
815 edges in the CFG when your pass has been run.
817 Some analyses chain to other analyses to do their job. For example, an
818 `AliasAnalysis <AliasAnalysis>` implementation is required to :ref:`chain
819 <aliasanalysis-chaining>` to other alias analysis passes. In cases where
820 analyses chain, the ``addRequiredTransitive`` method should be used instead of
821 the ``addRequired`` method. This informs the ``PassManager`` that the
822 transitively required pass should be alive as long as the requiring pass is.
824 The ``AnalysisUsage::addPreserved<>`` method
825 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
827 One of the jobs of the ``PassManager`` is to optimize how and when analyses are
828 run. In particular, it attempts to avoid recomputing data unless it needs to.
829 For this reason, passes are allowed to declare that they preserve (i.e., they
830 don't invalidate) an existing analysis if it's available. For example, a
831 simple constant folding pass would not modify the CFG, so it can't possibly
832 affect the results of dominator analysis. By default, all passes are assumed
833 to invalidate all others.
835 The ``AnalysisUsage`` class provides several methods which are useful in
836 certain circumstances that are related to ``addPreserved``. In particular, the
837 ``setPreservesAll`` method can be called to indicate that the pass does not
838 modify the LLVM program at all (which is true for analyses), and the
839 ``setPreservesCFG`` method can be used by transformations that change
840 instructions in the program but do not modify the CFG or terminator
841 instructions (note that this property is implicitly set for
842 :ref:`BasicBlockPass <writing-an-llvm-pass-BasicBlockPass>`\ es).
844 ``addPreserved`` is particularly useful for transformations like
845 ``BreakCriticalEdges``. This pass knows how to update a small set of loop and
846 dominator related analyses if they exist, so it can preserve them, despite the
847 fact that it hacks on the CFG.
849 Example implementations of ``getAnalysisUsage``
850 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
854 // This example modifies the program, but does not modify the CFG
855 void LICM::getAnalysisUsage(AnalysisUsage &AU) const {
856 AU.setPreservesCFG();
857 AU.addRequired<LoopInfo>();
860 .. _writing-an-llvm-pass-getAnalysis:
862 The ``getAnalysis<>`` and ``getAnalysisIfAvailable<>`` methods
863 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
865 The ``Pass::getAnalysis<>`` method is automatically inherited by your class,
866 providing you with access to the passes that you declared that you required
867 with the :ref:`getAnalysisUsage <writing-an-llvm-pass-getAnalysisUsage>`
868 method. It takes a single template argument that specifies which pass class
869 you want, and returns a reference to that pass. For example:
873 bool LICM::runOnFunction(Function &F) {
874 LoopInfo &LI = getAnalysis<LoopInfo>();
878 This method call returns a reference to the pass desired. You may get a
879 runtime assertion failure if you attempt to get an analysis that you did not
880 declare as required in your :ref:`getAnalysisUsage
881 <writing-an-llvm-pass-getAnalysisUsage>` implementation. This method can be
882 called by your ``run*`` method implementation, or by any other local method
883 invoked by your ``run*`` method.
885 A module level pass can use function level analysis info using this interface.
890 bool ModuleLevelPass::runOnModule(Module &M) {
892 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
896 In above example, ``runOnFunction`` for ``DominatorTree`` is called by pass
897 manager before returning a reference to the desired pass.
899 If your pass is capable of updating analyses if they exist (e.g.,
900 ``BreakCriticalEdges``, as described above), you can use the
901 ``getAnalysisIfAvailable`` method, which returns a pointer to the analysis if
902 it is active. For example:
906 if (DominatorSet *DS = getAnalysisIfAvailable<DominatorSet>()) {
907 // A DominatorSet is active. This code will update it.
910 Implementing Analysis Groups
911 ----------------------------
913 Now that we understand the basics of how passes are defined, how they are used,
914 and how they are required from other passes, it's time to get a little bit
915 fancier. All of the pass relationships that we have seen so far are very
916 simple: one pass depends on one other specific pass to be run before it can
917 run. For many applications, this is great, for others, more flexibility is
920 In particular, some analyses are defined such that there is a single simple
921 interface to the analysis results, but multiple ways of calculating them.
922 Consider alias analysis for example. The most trivial alias analysis returns
923 "may alias" for any alias query. The most sophisticated analysis a
924 flow-sensitive, context-sensitive interprocedural analysis that can take a
925 significant amount of time to execute (and obviously, there is a lot of room
926 between these two extremes for other implementations). To cleanly support
927 situations like this, the LLVM Pass Infrastructure supports the notion of
930 Analysis Group Concepts
931 ^^^^^^^^^^^^^^^^^^^^^^^
933 An Analysis Group is a single simple interface that may be implemented by
934 multiple different passes. Analysis Groups can be given human readable names
935 just like passes, but unlike passes, they need not derive from the ``Pass``
936 class. An analysis group may have one or more implementations, one of which is
937 the "default" implementation.
939 Analysis groups are used by client passes just like other passes are: the
940 ``AnalysisUsage::addRequired()`` and ``Pass::getAnalysis()`` methods. In order
941 to resolve this requirement, the :ref:`PassManager
942 <writing-an-llvm-pass-passmanager>` scans the available passes to see if any
943 implementations of the analysis group are available. If none is available, the
944 default implementation is created for the pass to use. All standard rules for
945 :ref:`interaction between passes <writing-an-llvm-pass-interaction>` still
948 Although :ref:`Pass Registration <writing-an-llvm-pass-registration>` is
949 optional for normal passes, all analysis group implementations must be
950 registered, and must use the :ref:`INITIALIZE_AG_PASS
951 <writing-an-llvm-pass-RegisterAnalysisGroup>` template to join the
952 implementation pool. Also, a default implementation of the interface **must**
953 be registered with :ref:`RegisterAnalysisGroup
954 <writing-an-llvm-pass-RegisterAnalysisGroup>`.
956 As a concrete example of an Analysis Group in action, consider the
957 `AliasAnalysis <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_
958 analysis group. The default implementation of the alias analysis interface
959 (the `basicaa <http://llvm.org/doxygen/structBasicAliasAnalysis.html>`_ pass)
960 just does a few simple checks that don't require significant analysis to
961 compute (such as: two different globals can never alias each other, etc).
962 Passes that use the `AliasAnalysis
963 <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_ interface (for
964 example the `gcse <http://llvm.org/doxygen/structGCSE.html>`_ pass), do not
965 care which implementation of alias analysis is actually provided, they just use
966 the designated interface.
968 From the user's perspective, commands work just like normal. Issuing the
969 command ``opt -gcse ...`` will cause the ``basicaa`` class to be instantiated
970 and added to the pass sequence. Issuing the command ``opt -somefancyaa -gcse
971 ...`` will cause the ``gcse`` pass to use the ``somefancyaa`` alias analysis
972 (which doesn't actually exist, it's just a hypothetical example) instead.
974 .. _writing-an-llvm-pass-RegisterAnalysisGroup:
976 Using ``RegisterAnalysisGroup``
977 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
979 The ``RegisterAnalysisGroup`` template is used to register the analysis group
980 itself, while the ``INITIALIZE_AG_PASS`` is used to add pass implementations to
981 the analysis group. First, an analysis group should be registered, with a
982 human readable name provided for it. Unlike registration of passes, there is
983 no command line argument to be specified for the Analysis Group Interface
984 itself, because it is "abstract":
988 static RegisterAnalysisGroup<AliasAnalysis> A("Alias Analysis");
990 Once the analysis is registered, passes can declare that they are valid
991 implementations of the interface by using the following code:
996 // Declare that we implement the AliasAnalysis interface
997 INITIALIZE_AG_PASS(FancyAA, AliasAnalysis , "somefancyaa",
998 "A more complex alias analysis implementation",
999 false, // Is CFG Only?
1000 true, // Is Analysis?
1001 false); // Is default Analysis Group implementation?
1004 This just shows a class ``FancyAA`` that uses the ``INITIALIZE_AG_PASS`` macro
1005 both to register and to "join" the `AliasAnalysis
1006 <http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html>`_ analysis group.
1007 Every implementation of an analysis group should join using this macro.
1012 // Declare that we implement the AliasAnalysis interface
1013 INITIALIZE_AG_PASS(BasicAA, AliasAnalysis, "basicaa",
1014 "Basic Alias Analysis (default AA impl)",
1015 false, // Is CFG Only?
1016 true, // Is Analysis?
1017 true); // Is default Analysis Group implementation?
1020 Here we show how the default implementation is specified (using the final
1021 argument to the ``INITIALIZE_AG_PASS`` template). There must be exactly one
1022 default implementation available at all times for an Analysis Group to be used.
1023 Only default implementation can derive from ``ImmutablePass``. Here we declare
1024 that the `BasicAliasAnalysis
1025 <http://llvm.org/doxygen/structBasicAliasAnalysis.html>`_ pass is the default
1026 implementation for the interface.
1031 The `Statistic <http://llvm.org/doxygen/Statistic_8h-source.html>`_ class is
1032 designed to be an easy way to expose various success metrics from passes.
1033 These statistics are printed at the end of a run, when the :option:`-stats`
1034 command line option is enabled on the command line. See the :ref:`Statistics
1035 section <Statistic>` in the Programmer's Manual for details.
1037 .. _writing-an-llvm-pass-passmanager:
1039 What PassManager does
1040 ---------------------
1042 The `PassManager <http://llvm.org/doxygen/PassManager_8h-source.html>`_ `class
1043 <http://llvm.org/doxygen/classllvm_1_1PassManager.html>`_ takes a list of
1044 passes, ensures their :ref:`prerequisites <writing-an-llvm-pass-interaction>`
1045 are set up correctly, and then schedules passes to run efficiently. All of the
1046 LLVM tools that run passes use the PassManager for execution of these passes.
1048 The PassManager does two main things to try to reduce the execution time of a
1051 #. **Share analysis results.** The ``PassManager`` attempts to avoid
1052 recomputing analysis results as much as possible. This means keeping track
1053 of which analyses are available already, which analyses get invalidated, and
1054 which analyses are needed to be run for a pass. An important part of work
1055 is that the ``PassManager`` tracks the exact lifetime of all analysis
1056 results, allowing it to :ref:`free memory
1057 <writing-an-llvm-pass-releaseMemory>` allocated to holding analysis results
1058 as soon as they are no longer needed.
1060 #. **Pipeline the execution of passes on the program.** The ``PassManager``
1061 attempts to get better cache and memory usage behavior out of a series of
1062 passes by pipelining the passes together. This means that, given a series
1063 of consecutive :ref:`FunctionPass <writing-an-llvm-pass-FunctionPass>`, it
1064 will execute all of the :ref:`FunctionPass
1065 <writing-an-llvm-pass-FunctionPass>` on the first function, then all of the
1066 :ref:`FunctionPasses <writing-an-llvm-pass-FunctionPass>` on the second
1067 function, etc... until the entire program has been run through the passes.
1069 This improves the cache behavior of the compiler, because it is only
1070 touching the LLVM program representation for a single function at a time,
1071 instead of traversing the entire program. It reduces the memory consumption
1072 of compiler, because, for example, only one `DominatorSet
1073 <http://llvm.org/doxygen/classllvm_1_1DominatorSet.html>`_ needs to be
1074 calculated at a time. This also makes it possible to implement some
1075 :ref:`interesting enhancements <writing-an-llvm-pass-SMP>` in the future.
1077 The effectiveness of the ``PassManager`` is influenced directly by how much
1078 information it has about the behaviors of the passes it is scheduling. For
1079 example, the "preserved" set is intentionally conservative in the face of an
1080 unimplemented :ref:`getAnalysisUsage <writing-an-llvm-pass-getAnalysisUsage>`
1081 method. Not implementing when it should be implemented will have the effect of
1082 not allowing any analysis results to live across the execution of your pass.
1084 The ``PassManager`` class exposes a ``--debug-pass`` command line options that
1085 is useful for debugging pass execution, seeing how things work, and diagnosing
1086 when you should be preserving more analyses than you currently are. (To get
1087 information about all of the variants of the ``--debug-pass`` option, just type
1088 "``opt -help-hidden``").
1090 By using the --debug-pass=Structure option, for example, we can see how our
1091 :ref:`Hello World <writing-an-llvm-pass-basiccode>` pass interacts with other
1092 passes. Lets try it out with the gcse and licm passes:
1094 .. code-block:: console
1096 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1098 Function Pass Manager
1099 Dominator Set Construction
1100 Immediate Dominators Construction
1101 Global Common Subexpression Elimination
1102 -- Immediate Dominators Construction
1103 -- Global Common Subexpression Elimination
1104 Natural Loop Construction
1105 Loop Invariant Code Motion
1106 -- Natural Loop Construction
1107 -- Loop Invariant Code Motion
1109 -- Dominator Set Construction
1114 This output shows us when passes are constructed and when the analysis results
1115 are known to be dead (prefixed with "``--``"). Here we see that GCSE uses
1116 dominator and immediate dominator information to do its job. The LICM pass
1117 uses natural loop information, which uses dominator sets, but not immediate
1118 dominators. Because immediate dominators are no longer useful after the GCSE
1119 pass, it is immediately destroyed. The dominator sets are then reused to
1120 compute natural loop information, which is then used by the LICM pass.
1122 After the LICM pass, the module verifier runs (which is automatically added by
1123 the :program:`opt` tool), which uses the dominator set to check that the
1124 resultant LLVM code is well formed. After it finishes, the dominator set
1125 information is destroyed, after being computed once, and shared by three
1128 Lets see how this changes when we run the :ref:`Hello World
1129 <writing-an-llvm-pass-basiccode>` pass in between the two passes:
1131 .. code-block:: console
1133 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1135 Function Pass Manager
1136 Dominator Set Construction
1137 Immediate Dominators Construction
1138 Global Common Subexpression Elimination
1139 -- Dominator Set Construction
1140 -- Immediate Dominators Construction
1141 -- Global Common Subexpression Elimination
1144 Dominator Set Construction
1145 Natural Loop Construction
1146 Loop Invariant Code Motion
1147 -- Natural Loop Construction
1148 -- Loop Invariant Code Motion
1150 -- Dominator Set Construction
1158 Here we see that the :ref:`Hello World <writing-an-llvm-pass-basiccode>` pass
1159 has killed the Dominator Set pass, even though it doesn't modify the code at
1160 all! To fix this, we need to add the following :ref:`getAnalysisUsage
1161 <writing-an-llvm-pass-getAnalysisUsage>` method to our pass:
1165 // We don't modify the program, so we preserve all analyses
1166 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1167 AU.setPreservesAll();
1170 Now when we run our pass, we get this output:
1172 .. code-block:: console
1174 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1175 Pass Arguments: -gcse -hello -licm
1177 Function Pass Manager
1178 Dominator Set Construction
1179 Immediate Dominators Construction
1180 Global Common Subexpression Elimination
1181 -- Immediate Dominators Construction
1182 -- Global Common Subexpression Elimination
1185 Natural Loop Construction
1186 Loop Invariant Code Motion
1187 -- Loop Invariant Code Motion
1188 -- Natural Loop Construction
1190 -- Dominator Set Construction
1198 Which shows that we don't accidentally invalidate dominator information
1199 anymore, and therefore do not have to compute it twice.
1201 .. _writing-an-llvm-pass-releaseMemory:
1203 The ``releaseMemory`` method
1204 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1208 virtual void releaseMemory();
1210 The ``PassManager`` automatically determines when to compute analysis results,
1211 and how long to keep them around for. Because the lifetime of the pass object
1212 itself is effectively the entire duration of the compilation process, we need
1213 some way to free analysis results when they are no longer useful. The
1214 ``releaseMemory`` virtual method is the way to do this.
1216 If you are writing an analysis or any other pass that retains a significant
1217 amount of state (for use by another pass which "requires" your pass and uses
1218 the :ref:`getAnalysis <writing-an-llvm-pass-getAnalysis>` method) you should
1219 implement ``releaseMemory`` to, well, release the memory allocated to maintain
1220 this internal state. This method is called after the ``run*`` method for the
1221 class, before the next call of ``run*`` in your pass.
1223 Registering dynamically loaded passes
1224 =====================================
1226 *Size matters* when constructing production quality tools using LLVM, both for
1227 the purposes of distribution, and for regulating the resident code size when
1228 running on the target system. Therefore, it becomes desirable to selectively
1229 use some passes, while omitting others and maintain the flexibility to change
1230 configurations later on. You want to be able to do all this, and, provide
1231 feedback to the user. This is where pass registration comes into play.
1233 The fundamental mechanisms for pass registration are the
1234 ``MachinePassRegistry`` class and subclasses of ``MachinePassRegistryNode``.
1236 An instance of ``MachinePassRegistry`` is used to maintain a list of
1237 ``MachinePassRegistryNode`` objects. This instance maintains the list and
1238 communicates additions and deletions to the command line interface.
1240 An instance of ``MachinePassRegistryNode`` subclass is used to maintain
1241 information provided about a particular pass. This information includes the
1242 command line name, the command help string and the address of the function used
1243 to create an instance of the pass. A global static constructor of one of these
1244 instances *registers* with a corresponding ``MachinePassRegistry``, the static
1245 destructor *unregisters*. Thus a pass that is statically linked in the tool
1246 will be registered at start up. A dynamically loaded pass will register on
1247 load and unregister at unload.
1249 Using existing registries
1250 -------------------------
1252 There are predefined registries to track instruction scheduling
1253 (``RegisterScheduler``) and register allocation (``RegisterRegAlloc``) machine
1254 passes. Here we will describe how to *register* a register allocator machine
1257 Implement your register allocator machine pass. In your register allocator
1258 ``.cpp`` file add the following include:
1262 #include "llvm/CodeGen/RegAllocRegistry.h"
1264 Also in your register allocator ``.cpp`` file, define a creator function in the
1269 FunctionPass *createMyRegisterAllocator() {
1270 return new MyRegisterAllocator();
1273 Note that the signature of this function should match the type of
1274 ``RegisterRegAlloc::FunctionPassCtor``. In the same file add the "installing"
1275 declaration, in the form:
1279 static RegisterRegAlloc myRegAlloc("myregalloc",
1280 "my register allocator help string",
1281 createMyRegisterAllocator);
1283 Note the two spaces prior to the help string produces a tidy result on the
1284 :option:`-help` query.
1286 .. code-block:: console
1290 -regalloc - Register allocator to use (default=linearscan)
1291 =linearscan - linear scan register allocator
1292 =local - local register allocator
1293 =simple - simple register allocator
1294 =myregalloc - my register allocator help string
1297 And that's it. The user is now free to use ``-regalloc=myregalloc`` as an
1298 option. Registering instruction schedulers is similar except use the
1299 ``RegisterScheduler`` class. Note that the
1300 ``RegisterScheduler::FunctionPassCtor`` is significantly different from
1301 ``RegisterRegAlloc::FunctionPassCtor``.
1303 To force the load/linking of your register allocator into the
1304 :program:`llc`/:program:`lli` tools, add your creator function's global
1305 declaration to ``Passes.h`` and add a "pseudo" call line to
1306 ``llvm/Codegen/LinkAllCodegenComponents.h``.
1308 Creating new registries
1309 -----------------------
1311 The easiest way to get started is to clone one of the existing registries; we
1312 recommend ``llvm/CodeGen/RegAllocRegistry.h``. The key things to modify are
1313 the class name and the ``FunctionPassCtor`` type.
1315 Then you need to declare the registry. Example: if your pass registry is
1316 ``RegisterMyPasses`` then define:
1320 MachinePassRegistry RegisterMyPasses::Registry;
1322 And finally, declare the command line option for your passes. Example:
1326 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1327 RegisterPassParser<RegisterMyPasses> >
1329 cl::init(&createDefaultMyPass),
1330 cl::desc("my pass option help"));
1332 Here the command option is "``mypass``", with ``createDefaultMyPass`` as the
1335 Using GDB with dynamically loaded passes
1336 ----------------------------------------
1338 Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1339 should be. First of all, you can't set a breakpoint in a shared object that
1340 has not been loaded yet, and second of all there are problems with inlined
1341 functions in shared objects. Here are some suggestions to debugging your pass
1344 For sake of discussion, I'm going to assume that you are debugging a
1345 transformation invoked by :program:`opt`, although nothing described here
1348 Setting a breakpoint in your pass
1349 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1351 First thing you do is start gdb on the opt process:
1353 .. code-block:: console
1357 Copyright 2000 Free Software Foundation, Inc.
1358 GDB is free software, covered by the GNU General Public License, and you are
1359 welcome to change it and/or distribute copies of it under certain conditions.
1360 Type "show copying" to see the conditions.
1361 There is absolutely no warranty for GDB. Type "show warranty" for details.
1362 This GDB was configured as "sparc-sun-solaris2.6"...
1365 Note that :program:`opt` has a lot of debugging information in it, so it takes
1366 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1367 (the shared object isn't loaded until runtime), we must execute the process,
1368 and have it stop before it invokes our pass, but after it has loaded the shared
1369 object. The most foolproof way of doing this is to set a breakpoint in
1370 ``PassManager::run`` and then run the process with the arguments you want:
1372 .. code-block:: console
1374 $ (gdb) break llvm::PassManager::run
1375 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1376 (gdb) run test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1377 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1378 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1379 70 bool PassManager::run(Module &M) { return PM->run(M); }
1382 Once the :program:`opt` stops in the ``PassManager::run`` method you are now
1383 free to set breakpoints in your pass so that you can trace through execution or
1384 do other standard debugging stuff.
1386 Miscellaneous Problems
1387 ^^^^^^^^^^^^^^^^^^^^^^
1389 Once you have the basics down, there are a couple of problems that GDB has,
1390 some with solutions, some without.
1392 * Inline functions have bogus stack information. In general, GDB does a pretty
1393 good job getting stack traces and stepping through inline functions. When a
1394 pass is dynamically loaded however, it somehow completely loses this
1395 capability. The only solution I know of is to de-inline a function (move it
1396 from the body of a class to a ``.cpp`` file).
1398 * Restarting the program breaks breakpoints. After following the information
1399 above, you have succeeded in getting some breakpoints planted in your pass.
1400 Nex thing you know, you restart the program (i.e., you type "``run``" again),
1401 and you start getting errors about breakpoints being unsettable. The only
1402 way I have found to "fix" this problem is to delete the breakpoints that are
1403 already set in your pass, run the program, and re-set the breakpoints once
1404 execution stops in ``PassManager::run``.
1406 Hopefully these tips will help with common case debugging situations. If you'd
1407 like to contribute some tips of your own, just contact `Chris
1408 <mailto:sabre@nondot.org>`_.
1410 Future extensions planned
1411 -------------------------
1413 Although the LLVM Pass Infrastructure is very capable as it stands, and does
1414 some nifty stuff, there are things we'd like to add in the future. Here is
1417 .. _writing-an-llvm-pass-SMP:
1422 Multiple CPU machines are becoming more common and compilation can never be
1423 fast enough: obviously we should allow for a multithreaded compiler. Because
1424 of the semantics defined for passes above (specifically they cannot maintain
1425 state across invocations of their ``run*`` methods), a nice clean way to
1426 implement a multithreaded compiler would be for the ``PassManager`` class to
1427 create multiple instances of each pass object, and allow the separate instances
1428 to be hacking on different parts of the program at the same time.
1430 This implementation would prevent each of the passes from having to implement
1431 multithreaded constructs, requiring only the LLVM core to have locking in a few
1432 places (for global resources). Although this is a simple extension, we simply
1433 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1434 Despite that, we have kept the LLVM passes SMP ready, and you should too.