1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
6 <title>Writing an LLVM Pass</title>
7 <link rel="stylesheet" href="llvm.css" type="text/css">
11 <div class="doc_title">
16 <li><a href="#introduction">Introduction - What is a pass?</a></li>
17 <li><a href="#quickstart">Quick Start - Writing hello world</a>
19 <li><a href="#makefile">Setting up the build environment</a></li>
20 <li><a href="#basiccode">Basic code required</a></li>
21 <li><a href="#running">Running a pass with <tt>opt</tt>
22 or <tt>analyze</tt></a></li>
24 <li><a href="#passtype">Pass classes and requirements</a>
26 <li><a href="#ImmutablePass">The <tt>ImmutablePass</tt> class</a></li>
27 <li><a href="#ModulePass">The <tt>ModulePass</tt> class</a>
29 <li><a href="#runOnModule">The <tt>runOnModule</tt> method</a></li>
31 <li><a href="#CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
33 <li><a href="#doInitialization_scc">The <tt>doInitialization(Module
34 &)</tt> method</a></li>
35 <li><a href="#runOnSCC">The <tt>runOnSCC</tt> method</a></li>
36 <li><a href="#doFinalization_scc">The <tt>doFinalization(Module
37 &)</tt> method</a></li>
39 <li><a href="#FunctionPass">The <tt>FunctionPass</tt> class</a>
41 <li><a href="#doInitialization_mod">The <tt>doInitialization(Module
42 &)</tt> method</a></li>
43 <li><a href="#runOnFunction">The <tt>runOnFunction</tt> method</a></li>
44 <li><a href="#doFinalization_mod">The <tt>doFinalization(Module
45 &)</tt> method</a></li>
47 <li><a href="#BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
49 <li><a href="#doInitialization_fn">The <tt>doInitialization(Function
50 &)</tt> method</a></li>
51 <li><a href="#runOnBasicBlock">The <tt>runOnBasicBlock</tt>
53 <li><a href="#doFinalization_fn">The <tt>doFinalization(Function
54 &)</tt> method</a></li>
56 <li><a href="#MachineFunctionPass">The <tt>MachineFunctionPass</tt>
59 <li><a href="#runOnMachineFunction">The
60 <tt>runOnMachineFunction(MachineFunction &)</tt> method</a></li>
63 <li><a href="#registration">Pass Registration</a>
65 <li><a href="#print">The <tt>print</tt> method</a></li>
67 <li><a href="#interaction">Specifying interactions between passes</a>
69 <li><a href="#getAnalysisUsage">The <tt>getAnalysisUsage</tt>
71 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
72 <li><a href="#AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a></li>
73 <li><a href="#AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a></li>
74 <li><a href="#getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a></li>
76 <li><a href="#analysisgroup">Implementing Analysis Groups</a>
78 <li><a href="#agconcepts">Analysis Group Concepts</a></li>
79 <li><a href="#registerag">Using <tt>RegisterAnalysisGroup</tt></a></li>
81 <li><a href="#passStatistics">Pass Statistics</a>
82 <li><a href="#passmanager">What PassManager does</a>
84 <li><a href="#releaseMemory">The <tt>releaseMemory</tt> method</a></li>
86 <li><a href="#registering">Registering dynamically loaded passes</a>
88 <li><a href="#registering_existing">Using existing registries</a></li>
89 <li><a href="#registering_new">Creating new registries</a></li>
91 <li><a href="#debughints">Using GDB with dynamically loaded passes</a>
93 <li><a href="#breakpoint">Setting a breakpoint in your pass</a></li>
94 <li><a href="#debugmisc">Miscellaneous Problems</a></li>
96 <li><a href="#future">Future extensions planned</a>
98 <li><a href="#SMP">Multithreaded LLVM</a></li>
99 <li><a href="#PassFunctionPass"><tt>ModulePass</tt>es requiring
100 <tt>FunctionPass</tt>es</a></li>
104 <div class="doc_author">
105 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a>,
106 <a href="mailto:jlaskey@apple.com">Jim Laskey</a></p>
109 <!-- *********************************************************************** -->
110 <div class="doc_section">
111 <a name="introduction">Introduction - What is a pass?</a>
113 <!-- *********************************************************************** -->
115 <div class="doc_text">
117 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
118 passes are where most of the interesting parts of the compiler exist. Passes
119 perform the transformations and optimizations that make up the compiler, they
120 build the analysis results that are used by these transformations, and they are,
121 above all, a structuring technique for compiler code.</p>
123 <p>All LLVM passes are subclasses of the <tt><a
124 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
125 class, which implement functionality by overriding virtual methods inherited
126 from <tt>Pass</tt>. Depending on how your pass works, you should inherit from
127 the <tt><a href="#ModulePass">ModulePass</a></tt>, <tt><a
128 href="#CallGraphSCCPass">CallGraphSCCPass</a></tt>, <tt><a
129 href="#FunctionPass">FunctionPass</a></tt>, or <tt><a
130 href="#BasicBlockPass">BasicBlockPass</a></tt> classes, which gives the system
131 more information about what your pass does, and how it can be combined with
132 other passes. One of the main features of the LLVM Pass Framework is that it
133 schedules passes to run in an efficient way based on the constraints that your
134 pass meets (which are indicated by which class they derive from).</p>
136 <p>We start by showing you how to construct a pass, everything from setting up
137 the code, to compiling, loading, and executing it. After the basics are down,
138 more advanced features are discussed.</p>
142 <!-- *********************************************************************** -->
143 <div class="doc_section">
144 <a name="quickstart">Quick Start - Writing hello world</a>
146 <!-- *********************************************************************** -->
148 <div class="doc_text">
150 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
151 is designed to simply print out the name of non-external functions that exist in
152 the program being compiled. It does not modify the program at all, it just
153 inspects it. The source code and files for this pass are available in the LLVM
154 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
158 <!-- ======================================================================= -->
159 <div class="doc_subsection">
160 <a name="makefile">Setting up the build environment</a>
163 <div class="doc_text">
165 <p>First, you need to create a new directory somewhere in the LLVM source
166 base. For this example, we'll assume that you made
167 <tt>lib/Transforms/Hello</tt>. Next, you must set up a build script
168 (Makefile) that will compile the source code for the new pass. To do this,
169 copy the following into <tt>Makefile</tt>:</p>
172 <div class="doc_code"><pre>
173 # Makefile for hello pass
175 # Path to top level of LLVM heirarchy
178 # Name of the library to build
181 # Build a dynamically linkable shared object
184 # Make the shared library become a loadable module so the tools can
185 # dlopen/dlsym on the resulting library.
188 # Include the makefile implementation stuff
189 include $(LEVEL)/Makefile.common
192 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
193 directory are to be compiled and linked together into a
194 <tt>Debug/lib/Hello.so</tt> shared object that can be dynamically loaded by
195 the <tt>opt</tt> or <tt>analyze</tt> tools via their <tt>-load</tt> options.
196 If your operating system uses a suffix other than .so (such as windows or
197 Mac OS/X), the appropriate extension will be used.</p>
199 <p>Now that we have the build scripts set up, we just need to write the code for
204 <!-- ======================================================================= -->
205 <div class="doc_subsection">
206 <a name="basiccode">Basic code required</a>
209 <div class="doc_text">
211 <p>Now that we have a way to compile our new pass, we just have to write it.
214 <div class="doc_code"><pre>
215 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
216 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
219 <p>Which are needed because we are writing a <tt><a
220 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>, and
221 we are operating on <tt><a
222 href="http://llvm.org/doxygen/classllvm_1_1Function.html">Function</a></tt>'s.</p>
225 <div class="doc_code"><pre>
226 <b>using namespace llvm;</b>
228 <p>... which is required because the functions from the include files
229 live in the llvm namespace.
234 <div class="doc_code"><pre>
238 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
239 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
240 things declared inside of the anonymous namespace only visible to the current
241 file. If you're not familiar with them, consult a decent C++ book for more
244 <p>Next, we declare our pass itself:</p>
246 <div class="doc_code"><pre>
247 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
250 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
251 href="http://llvm.org/doxygen/classllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
252 The different builtin pass subclasses are described in detail <a
253 href="#passtype">later</a>, but for now, know that <a
254 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate a function at a
257 <div class="doc_code"><pre>
258 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
259 std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
262 }; <i>// end of struct Hello</i>
265 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
266 which overloads an abstract virtual method inherited from <a
267 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
268 to do our thing, so we just print out our message with the name of each
271 <div class="doc_code"><pre>
272 RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
273 } <i>// end of anonymous namespace</i>
276 <p>Lastly, we register our class <tt>Hello</tt>, giving it a command line
277 argument "<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>". There are
278 several different ways of <a href="#registration">registering your pass</a>,
279 depending on what it is to be used for. For "optimizations" we use the
280 <tt>RegisterOpt</tt> template.</p>
282 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
284 <div class="doc_code"><pre>
285 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
286 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
288 <b>using namespace llvm;</b>
291 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
292 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
293 std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
298 RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
302 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
303 command in the local directory and you should get a new
304 "<tt>Debug/lib/Hello.so</tt> file. Note that everything in this file is
305 contained in an anonymous namespace: this reflects the fact that passes are self
306 contained units that do not need external interfaces (although they can have
307 them) to be useful.</p>
311 <!-- ======================================================================= -->
312 <div class="doc_subsection">
313 <a name="running">Running a pass with <tt>opt</tt> or <tt>analyze</tt></a>
316 <div class="doc_text">
318 <p>Now that you have a brand new shiny shared object file, we can use the
319 <tt>opt</tt> command to run an LLVM program through your pass. Because you
320 registered your pass with the <tt>RegisterOpt</tt> template, you will be able to
321 use the <tt>opt</tt> tool to access it, once loaded.</p>
323 <p>To test it, follow the example at the end of the <a
324 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
325 LLVM. We can now run the bytecode file (<tt>hello.bc</tt>) for the program
326 through our transformation like this (or course, any bytecode file will
329 <div class="doc_code"><pre>
330 $ opt -load ../../../Debug/lib/Hello.so -hello < hello.bc > /dev/null
336 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
337 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
338 argument (which is one reason you need to <a href="#registration">register your
339 pass</a>). Because the hello pass does not modify the program in any
340 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
341 <tt>/dev/null</tt>).</p>
343 <p>To see what happened to the other string you registered, try running
344 <tt>opt</tt> with the <tt>--help</tt> option:</p>
346 <div class="doc_code"><pre>
347 $ opt -load ../../../Debug/lib/Hello.so --help
348 OVERVIEW: llvm .bc -> .bc modular optimizer
350 USAGE: opt [options] <input bytecode>
353 Optimizations available:
355 -funcresolve - Resolve Functions
356 -gcse - Global Common Subexpression Elimination
357 -globaldce - Dead Global Elimination
358 <b>-hello - Hello World Pass</b>
359 -indvars - Canonicalize Induction Variables
360 -inline - Function Integration/Inlining
361 -instcombine - Combine redundant instructions
365 <p>The pass name get added as the information string for your pass, giving some
366 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
367 would go ahead and make it do the cool transformations you want. Once you get
368 it all working and tested, it may become useful to find out how fast your pass
369 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
370 line option (<tt>--time-passes</tt>) that allows you to get information about
371 the execution time of your pass along with the other passes you queue up. For
374 <div class="doc_code"><pre>
375 $ opt -load ../../../Debug/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
379 ===============================================================================
380 ... Pass execution timing report ...
381 ===============================================================================
382 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
384 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
385 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bytecode Writer
386 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
387 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
388 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
389 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
392 <p>As you can see, our implementation above is pretty fast :). The additional
393 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
394 that the LLVM emitted by your pass is still valid and well formed LLVM, which
395 hasn't been broken somehow.</p>
397 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
398 about some more details of how they work and how to use them.</p>
402 <!-- *********************************************************************** -->
403 <div class="doc_section">
404 <a name="passtype">Pass classes and requirements</a>
406 <!-- *********************************************************************** -->
408 <div class="doc_text">
410 <p>One of the first things that you should do when designing a new pass is to
411 decide what class you should subclass for your pass. The <a
412 href="#basiccode">Hello World</a> example uses the <tt><a
413 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
414 did not discuss why or when this should occur. Here we talk about the classes
415 available, from the most general to the most specific.</p>
417 <p>When choosing a superclass for your Pass, you should choose the <b>most
418 specific</b> class possible, while still being able to meet the requirements
419 listed. This gives the LLVM Pass Infrastructure information necessary to
420 optimize how passes are run, so that the resultant compiler isn't unneccesarily
425 <!-- ======================================================================= -->
426 <div class="doc_subsection">
427 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
430 <div class="doc_text">
432 <p>The most plain and boring type of pass is the "<tt><a
433 href="http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
434 class. This pass type is used for passes that do not have to be run, do not
435 change state, and never need to be updated. This is not a normal type of
436 transformation or analysis, but can provide information about the current
437 compiler configuration.</p>
439 <p>Although this pass class is very infrequently used, it is important for
440 providing information about the current target machine being compiled for, and
441 other static information that can affect the various transformations.</p>
443 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
444 invalidated, and are never "run".</p>
448 <!-- ======================================================================= -->
449 <div class="doc_subsection">
450 <a name="ModulePass">The <tt>ModulePass</tt> class</a>
453 <div class="doc_text">
456 href="http://llvm.org/doxygen/classllvm_1_1ModulePass.html">ModulePass</a></tt>"
457 class is the most general of all superclasses that you can use. Deriving from
458 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
459 refering to function bodies in no predictable order, or adding and removing
460 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
461 subclasses, no optimization can be done for their execution.</p>
463 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
464 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
465 following signature:</p>
469 <!-- _______________________________________________________________________ -->
470 <div class="doc_subsubsection">
471 <a name="runOnModule">The <tt>runOnModule</tt> method</a>
474 <div class="doc_text">
476 <div class="doc_code"><pre>
477 <b>virtual bool</b> runOnModule(Module &M) = 0;
480 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
481 It should return true if the module was modified by the transformation and
486 <!-- ======================================================================= -->
487 <div class="doc_subsection">
488 <a name="CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
491 <div class="doc_text">
494 href="http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html">CallGraphSCCPass</a></tt>"
495 is used by passes that need to traverse the program bottom-up on the call graph
496 (callees before callers). Deriving from CallGraphSCCPass provides some
497 mechanics for building and traversing the CallGraph, but also allows the system
498 to optimize execution of CallGraphSCCPass's. If your pass meets the
499 requirements outlined below, and doesn't meet the requirements of a <tt><a
500 href="#FunctionPass">FunctionPass</a></tt> or <tt><a
501 href="#BasicBlockPass">BasicBlockPass</a></tt>, you should derive from
502 <tt>CallGraphSCCPass</tt>.</p>
504 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
506 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
510 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
511 the current SCC.</li>
513 <li>... <em>allowed</em> to inspect any Function's other than those in the
514 current SCC and the direct callees of the SCC.</li>
516 <li>... <em>required</em> to preserve the current CallGraph object, updating it
517 to reflect any changes made to the program.</li>
519 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
520 though they may change the contents of an SCC.</li>
522 <li>... <em>allowed</em> to add or remove global variables from the current
525 <li>... <em>allowed</em> to maintain state across invocations of
526 <a href="#runOnSCC"><tt>runOnSCC</tt></a> (including global data).</li>
529 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
530 because it has to handle SCCs with more than one node in it. All of the virtual
531 methods described below should return true if they modified the program, or
532 false if they didn't.</p>
536 <!-- _______________________________________________________________________ -->
537 <div class="doc_subsubsection">
538 <a name="doInitialization_scc">The <tt>doInitialization(Module &)</tt>
542 <div class="doc_text">
544 <div class="doc_code"><pre>
545 <b>virtual bool</b> doInitialization(Module &M);
548 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
549 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
550 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
551 is designed to do simple initialization type of stuff that does not depend on
552 the SCCs being processed. The <tt>doInitialization</tt> method call is not
553 scheduled to overlap with any other pass executions (thus it should be very
558 <!-- _______________________________________________________________________ -->
559 <div class="doc_subsubsection">
560 <a name="runOnSCC">The <tt>runOnSCC</tt> method</a>
563 <div class="doc_text">
565 <div class="doc_code"><pre>
566 <b>virtual bool</b> runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;
569 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
570 should return true if the module was modified by the transformation, false
575 <!-- _______________________________________________________________________ -->
576 <div class="doc_subsubsection">
577 <a name="doFinalization_scc">The <tt>doFinalization(Module
578 &)</tt> method</a>
581 <div class="doc_text">
583 <div class="doc_code"><pre>
584 <b>virtual bool</b> doFinalization(Module &M);
587 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
588 called when the pass framework has finished calling <a
589 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
590 program being compiled.</p>
594 <!-- ======================================================================= -->
595 <div class="doc_subsection">
596 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
599 <div class="doc_text">
601 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
602 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
603 subclasses do have a predictable, local behavior that can be expected by the
604 system. All <tt>FunctionPass</tt> execute on each function in the program
605 independent of all of the other functions in the program.
606 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
607 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
609 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
612 <li>Modify a Function other than the one currently being processed.</li>
613 <li>Add or remove Function's from the current Module.</li>
614 <li>Add or remove global variables from the current Module.</li>
615 <li>Maintain state across invocations of
616 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
619 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
620 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
621 may overload three virtual methods to do their work. All of these methods
622 should return true if they modified the program, or false if they didn't.</p>
626 <!-- _______________________________________________________________________ -->
627 <div class="doc_subsubsection">
628 <a name="doInitialization_mod">The <tt>doInitialization(Module &)</tt>
632 <div class="doc_text">
634 <div class="doc_code"><pre>
635 <b>virtual bool</b> doInitialization(Module &M);
638 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
639 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
640 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
641 is designed to do simple initialization type of stuff that does not depend on
642 the functions being processed. The <tt>doInitialization</tt> method call is not
643 scheduled to overlap with any other pass executions (thus it should be very
646 <p>A good example of how this method should be used is the <a
647 href="http://llvm.org/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
648 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
649 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
650 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
651 free functions that it needs, adding prototypes to the module if necessary.</p>
655 <!-- _______________________________________________________________________ -->
656 <div class="doc_subsubsection">
657 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
660 <div class="doc_text">
662 <div class="doc_code"><pre>
663 <b>virtual bool</b> runOnFunction(Function &F) = 0;
666 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
667 the transformation or analysis work of your pass. As usual, a true value should
668 be returned if the function is modified.</p>
672 <!-- _______________________________________________________________________ -->
673 <div class="doc_subsubsection">
674 <a name="doFinalization_mod">The <tt>doFinalization(Module
675 &)</tt> method</a>
678 <div class="doc_text">
680 <div class="doc_code"><pre>
681 <b>virtual bool</b> doFinalization(Module &M);
684 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
685 called when the pass framework has finished calling <a
686 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
687 program being compiled.</p>
691 <!-- ======================================================================= -->
692 <div class="doc_subsection">
693 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
696 <div class="doc_text">
698 <p><tt>BasicBlockPass</tt>'s are just like <a
699 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
700 their scope of inspection and modification to a single basic block at a time.
701 As such, they are <b>not</b> allowed to do any of the following:</p>
704 <li>Modify or inspect any basic blocks outside of the current one</li>
705 <li>Maintain state across invocations of
706 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
707 <li>Modify the control flow graph (by altering terminator instructions)</li>
708 <li>Any of the things forbidden for
709 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
712 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
713 optimizations. They may override the same <a
714 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
715 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
716 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
720 <!-- _______________________________________________________________________ -->
721 <div class="doc_subsubsection">
722 <a name="doInitialization_fn">The <tt>doInitialization(Function
723 &)</tt> method</a>
726 <div class="doc_text">
728 <div class="doc_code"><pre>
729 <b>virtual bool</b> doInitialization(Function &F);
732 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
733 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
734 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
735 to do simple initialization that does not depend on the
736 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
737 scheduled to overlap with any other pass executions (thus it should be very
742 <!-- _______________________________________________________________________ -->
743 <div class="doc_subsubsection">
744 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
747 <div class="doc_text">
749 <div class="doc_code"><pre>
750 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
753 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
754 function is not allowed to inspect or modify basic blocks other than the
755 parameter, and are not allowed to modify the CFG. A true value must be returned
756 if the basic block is modified.</p>
760 <!-- _______________________________________________________________________ -->
761 <div class="doc_subsubsection">
762 <a name="doFinalization_fn">The <tt>doFinalization(Function &)</tt>
766 <div class="doc_text">
768 <div class="doc_code"><pre>
769 <b>virtual bool</b> doFinalization(Function &F);
772 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
773 called when the pass framework has finished calling <a
774 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
775 program being compiled. This can be used to perform per-function
780 <!-- ======================================================================= -->
781 <div class="doc_subsection">
782 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
785 <div class="doc_text">
787 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
788 executes on the machine-dependent representation of each LLVM function in the
789 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
790 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
791 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
792 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
795 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
796 <li>Modify a MachineFunction other than the one currently being processed.</li>
797 <li>Add or remove MachineFunctions from the current Module.</li>
798 <li>Add or remove global variables from the current Module.</li>
799 <li>Maintain state across invocations of <a
800 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
806 <!-- _______________________________________________________________________ -->
807 <div class="doc_subsubsection">
808 <a name="runOnMachineFunction">The <tt>runOnMachineFunction(MachineFunction
809 &MF)</tt> method</a>
812 <div class="doc_text">
814 <div class="doc_code"><pre>
815 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
818 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
819 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
820 work of your <tt>MachineFunctionPass</tt>.</p>
822 <p>The <tt>runOnMachineFunction</tt> method is called on every
823 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
824 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
825 representation of the function. If you want to get at the LLVM <tt>Function</tt>
826 for the <tt>MachineFunction</tt> you're working on, use
827 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
828 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
829 <tt>MachineFunctionPass</tt>.</p>
833 <!-- *********************************************************************** -->
834 <div class="doc_section">
835 <a name="registration">Pass registration</a>
837 <!-- *********************************************************************** -->
839 <div class="doc_text">
841 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
842 pass registration works, and discussed some of the reasons that it is used and
843 what it does. Here we discuss how and why passes are registered.</p>
845 <p>Passes can be registered in several different ways. Depending on the general
846 classification of the pass, you should use one of the following templates to
847 register the pass:</p>
850 <li><b><tt>RegisterOpt</tt></b> - This template should be used when you are
851 registering a pass that logically should be available for use in the
852 '<tt>opt</tt>' utility.</li>
854 <li><b><tt>RegisterAnalysis</tt></b> - This template should be used when you are
855 registering a pass that logically should be available for use in the
856 '<tt>analyze</tt>' utility.</li>
858 <li><b><tt>RegisterPass</tt></b> - This is the generic form of the
859 <tt>Register*</tt> templates that should be used if you want your pass listed by
860 multiple or no utilities. This template takes an extra third argument that
861 specifies which tools it should be listed in. See the <a
862 href="http://llvm.org/doxygen/PassSupport_8h-source.html">PassSupport.h</a>
863 file for more information.</li>
867 <p>Regardless of how you register your pass, you must specify at least two
868 parameters. The first parameter is the name of the pass that is to be used on
869 the command line to specify that the pass should be added to a program (for
870 example <tt>opt</tt> or <tt>analyze</tt>). The second argument is the name of
871 the pass, which is to be used for the <tt>--help</tt> output of programs, as
872 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
874 <p>If a pass is registered to be used by the <tt>analyze</tt> utility, you
875 should implement the virtual <tt>print</tt> method:</p>
879 <!-- _______________________________________________________________________ -->
880 <div class="doc_subsubsection">
881 <a name="print">The <tt>print</tt> method</a>
884 <div class="doc_text">
886 <div class="doc_code"><pre>
887 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
890 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
891 a human readable version of the analysis results. This is useful for debugging
892 an analysis itself, as well as for other people to figure out how an analysis
893 works. The <tt>analyze</tt> tool uses this method to generate its output.</p>
895 <p>The <tt>ostream</tt> parameter specifies the stream to write the results on,
896 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
897 program that has been analyzed. Note however that this pointer may be null in
898 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
899 debugger), so it should only be used to enhance debug output, it should not be
904 <!-- *********************************************************************** -->
905 <div class="doc_section">
906 <a name="interaction">Specifying interactions between passes</a>
908 <!-- *********************************************************************** -->
910 <div class="doc_text">
912 <p>One of the main responsibilities of the <tt>PassManager</tt> is the make sure
913 that passes interact with each other correctly. Because <tt>PassManager</tt>
914 tries to <a href="#passmanager">optimize the execution of passes</a> it must
915 know how the passes interact with each other and what dependencies exist between
916 the various passes. To track this, each pass can declare the set of passes that
917 are required to be executed before the current pass, and the passes which are
918 invalidated by the current pass.</p>
920 <p>Typically this functionality is used to require that analysis results are
921 computed before your pass is run. Running arbitrary transformation passes can
922 invalidate the computed analysis results, which is what the invalidation set
923 specifies. If a pass does not implement the <tt><a
924 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
925 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
929 <!-- _______________________________________________________________________ -->
930 <div class="doc_subsubsection">
931 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
934 <div class="doc_text">
936 <div class="doc_code"><pre>
937 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
940 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
941 invalidated sets may be specified for your transformation. The implementation
942 should fill in the <tt><a
943 href="http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
944 object with information about which passes are required and not invalidated. To
945 do this, a pass may call any of the following methods on the AnalysisUsage
949 <!-- _______________________________________________________________________ -->
950 <div class="doc_subsubsection">
951 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
954 <div class="doc_text">
956 If your pass requires a previous pass to be executed (an analysis for example),
957 it can use one of these methods to arrange for it to be run before your pass.
958 LLVM has many different types of analyses and passes that can be required,
959 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
960 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
961 be no critical edges in the CFG when your pass has been run.
965 Some analyses chain to other analyses to do their job. For example, an <a
966 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
967 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
968 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
969 used instead of the <tt>addRequired</tt> method. This informs the PassManager
970 that the transitively required pass should be alive as long as the requiring
975 <!-- _______________________________________________________________________ -->
976 <div class="doc_subsubsection">
977 <a name="AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a>
980 <div class="doc_text">
982 One of the jobs of the PassManager is to optimize how and when analyses are run.
983 In particular, it attempts to avoid recomputing data unless it needs to. For
984 this reason, passes are allowed to declare that they preserve (i.e., they don't
985 invalidate) an existing analysis if it's available. For example, a simple
986 constant folding pass would not modify the CFG, so it can't possibly affect the
987 results of dominator analysis. By default, all passes are assumed to invalidate
992 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
993 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
994 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
995 not modify the LLVM program at all (which is true for analyses), and the
996 <tt>setPreservesCFG</tt> method can be used by transformations that change
997 instructions in the program but do not modify the CFG or terminator instructions
998 (note that this property is implicitly set for <a
999 href="#BasicBlockPass">BasicBlockPass</a>'s).
1003 <tt>addPreserved</tt> is particularly useful for transformations like
1004 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
1005 and dominator related analyses if they exist, so it can preserve them, despite
1006 the fact that it hacks on the CFG.
1010 <!-- _______________________________________________________________________ -->
1011 <div class="doc_subsubsection">
1012 <a name="AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a>
1015 <div class="doc_text">
1017 <div class="doc_code"><pre>
1018 <i>// This is an example implementation from an analysis, which does not modify
1019 // the program at all, yet has a prerequisite.</i>
1020 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1021 AU.setPreservesAll();
1022 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1028 <div class="doc_code"><pre>
1029 <i>// This example modifies the program, but does not modify the CFG</i>
1030 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1031 AU.setPreservesCFG();
1032 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1038 <!-- _______________________________________________________________________ -->
1039 <div class="doc_subsubsection">
1040 <a name="getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a>
1043 <div class="doc_text">
1045 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1046 your class, providing you with access to the passes that you declared that you
1047 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
1048 method. It takes a single template argument that specifies which pass class you
1049 want, and returns a reference to that pass. For example:</p>
1051 <div class="doc_code"><pre>
1052 bool LICM::runOnFunction(Function &F) {
1053 LoopInfo &LI = getAnalysis<LoopInfo>();
1058 <p>This method call returns a reference to the pass desired. You may get a
1059 runtime assertion failure if you attempt to get an analysis that you did not
1060 declare as required in your <a
1061 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
1062 method can be called by your <tt>run*</tt> method implementation, or by any
1063 other local method invoked by your <tt>run*</tt> method.</p>
1066 If your pass is capable of updating analyses if they exist (e.g.,
1067 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1068 <tt>getAnalysisToUpdate</tt> method, which returns a pointer to the analysis if
1069 it is active. For example:</p>
1071 <div class="doc_code"><pre>
1073 if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
1074 <i>// A DominatorSet is active. This code will update it.</i>
1081 <!-- *********************************************************************** -->
1082 <div class="doc_section">
1083 <a name="analysisgroup">Implementing Analysis Groups</a>
1085 <!-- *********************************************************************** -->
1087 <div class="doc_text">
1089 <p>Now that we understand the basics of how passes are defined, how the are
1090 used, and how they are required from other passes, it's time to get a little bit
1091 fancier. All of the pass relationships that we have seen so far are very
1092 simple: one pass depends on one other specific pass to be run before it can run.
1093 For many applications, this is great, for others, more flexibility is
1096 <p>In particular, some analyses are defined such that there is a single simple
1097 interface to the analysis results, but multiple ways of calculating them.
1098 Consider alias analysis for example. The most trivial alias analysis returns
1099 "may alias" for any alias query. The most sophisticated analysis a
1100 flow-sensitive, context-sensitive interprocedural analysis that can take a
1101 significant amount of time to execute (and obviously, there is a lot of room
1102 between these two extremes for other implementations). To cleanly support
1103 situations like this, the LLVM Pass Infrastructure supports the notion of
1104 Analysis Groups.</p>
1108 <!-- _______________________________________________________________________ -->
1109 <div class="doc_subsubsection">
1110 <a name="agconcepts">Analysis Group Concepts</a>
1113 <div class="doc_text">
1115 <p>An Analysis Group is a single simple interface that may be implemented by
1116 multiple different passes. Analysis Groups can be given human readable names
1117 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1118 class. An analysis group may have one or more implementations, one of which is
1119 the "default" implementation.</p>
1121 <p>Analysis groups are used by client passes just like other passes are: the
1122 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1123 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
1124 scans the available passes to see if any implementations of the analysis group
1125 are available. If none is available, the default implementation is created for
1126 the pass to use. All standard rules for <A href="#interaction">interaction
1127 between passes</a> still apply.</p>
1129 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
1130 passes, all analysis group implementations must be registered, and must use the
1131 <A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
1132 implementation pool. Also, a default implementation of the interface
1133 <b>must</b> be registered with <A
1134 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
1136 <p>As a concrete example of an Analysis Group in action, consider the <a
1137 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1138 analysis group. The default implementation of the alias analysis interface (the
1140 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1141 pass) just does a few simple checks that don't require significant analysis to
1142 compute (such as: two different globals can never alias each other, etc).
1143 Passes that use the <tt><a
1144 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1145 interface (for example the <tt><a
1146 href="http://llvm.org/doxygen/structGCSE.html">gcse</a></tt> pass), do
1147 not care which implementation of alias analysis is actually provided, they just
1148 use the designated interface.</p>
1150 <p>From the user's perspective, commands work just like normal. Issuing the
1151 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1152 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1153 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1154 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1155 hypothetical example) instead.</p>
1159 <!-- _______________________________________________________________________ -->
1160 <div class="doc_subsubsection">
1161 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1164 <div class="doc_text">
1166 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1167 group itself as well as add pass implementations to the analysis group. First,
1168 an analysis should be registered, with a human readable name provided for it.
1169 Unlike registration of passes, there is no command line argument to be specified
1170 for the Analysis Group Interface itself, because it is "abstract":</p>
1172 <div class="doc_code"><pre>
1173 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1176 <p>Once the analysis is registered, passes can declare that they are valid
1177 implementations of the interface by using the following code:</p>
1179 <div class="doc_code"><pre>
1181 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1182 RegisterOpt<FancyAA>
1183 B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
1185 //<i> Declare that we implement the AliasAnalysis interface</i>
1186 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, FancyAA> C;
1190 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1191 uses the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
1192 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1193 analysis group. Every implementation of an analysis group should join using
1194 this template. A single pass may join multiple different analysis groups with
1197 <div class="doc_code"><pre>
1199 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1200 RegisterOpt<<a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1201 D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
1203 //<i> Declare that we implement the AliasAnalysis interface</i>
1204 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>, <b>true</b>> E;
1208 <p>Here we show how the default implementation is specified (using the extra
1209 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1210 one default implementation available at all times for an Analysis Group to be
1211 used. Here we declare that the <tt><a
1212 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1213 pass is the default implementation for the interface.</p>
1217 <!-- *********************************************************************** -->
1218 <div class="doc_section">
1219 <a name="passStatistics">Pass Statistics</a>
1221 <!-- *********************************************************************** -->
1223 <div class="doc_text">
1225 href="http://llvm.org/doxygen/Statistic_8h-source.html"><tt>Statistic</tt></a>
1226 class is designed to be an easy way to expose various success
1227 metrics from passes. These statistics are printed at the end of a
1228 run, when the -stats command line option is enabled on the command
1229 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1234 <!-- *********************************************************************** -->
1235 <div class="doc_section">
1236 <a name="passmanager">What PassManager does</a>
1238 <!-- *********************************************************************** -->
1240 <div class="doc_text">
1243 href="http://llvm.org/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1245 href="http://llvm.org/doxygen/classllvm_1_1PassManager.html">class</a>
1246 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1247 are set up correctly, and then schedules passes to run efficiently. All of the
1248 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1251 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1252 time of a series of passes:</p>
1255 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1256 recomputing analysis results as much as possible. This means keeping track of
1257 which analyses are available already, which analyses get invalidated, and which
1258 analyses are needed to be run for a pass. An important part of work is that the
1259 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1260 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1261 results as soon as they are no longer needed.</li>
1263 <li><b>Pipeline the execution of passes on the program</b> - The
1264 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1265 of a series of passes by pipelining the passes together. This means that, given
1266 a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1267 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1268 the first function, then all of the <a
1269 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1270 etc... until the entire program has been run through the passes.
1272 <p>This improves the cache behavior of the compiler, because it is only touching
1273 the LLVM program representation for a single function at a time, instead of
1274 traversing the entire program. It reduces the memory consumption of compiler,
1275 because, for example, only one <a
1276 href="http://llvm.org/doxygen/classllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1277 needs to be calculated at a time. This also makes it possible some <a
1278 href="#SMP">interesting enhancements</a> in the future.</p></li>
1282 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1283 much information it has about the behaviors of the passes it is scheduling. For
1284 example, the "preserved" set is intentionally conservative in the face of an
1285 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1286 Not implementing when it should be implemented will have the effect of not
1287 allowing any analysis results to live across the execution of your pass.</p>
1289 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1290 options that is useful for debugging pass execution, seeing how things work, and
1291 diagnosing when you should be preserving more analyses than you currently are
1292 (To get information about all of the variants of the <tt>--debug-pass</tt>
1293 option, just type '<tt>opt --help-hidden</tt>').</p>
1295 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1296 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1297 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1299 <div class="doc_code"><pre>
1300 $ opt -load ../../../Debug/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1302 Function Pass Manager
1303 Dominator Set Construction
1304 Immediate Dominators Construction
1305 Global Common Subexpression Elimination
1306 -- Immediate Dominators Construction
1307 -- Global Common Subexpression Elimination
1308 Natural Loop Construction
1309 Loop Invariant Code Motion
1310 -- Natural Loop Construction
1311 -- Loop Invariant Code Motion
1313 -- Dominator Set Construction
1319 <p>This output shows us when passes are constructed and when the analysis
1320 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1321 GCSE uses dominator and immediate dominator information to do its job. The LICM
1322 pass uses natural loop information, which uses dominator sets, but not immediate
1323 dominators. Because immediate dominators are no longer useful after the GCSE
1324 pass, it is immediately destroyed. The dominator sets are then reused to
1325 compute natural loop information, which is then used by the LICM pass.</p>
1327 <p>After the LICM pass, the module verifier runs (which is automatically added
1328 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1329 resultant LLVM code is well formed. After it finishes, the dominator set
1330 information is destroyed, after being computed once, and shared by three
1333 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1334 World</a> pass in between the two passes:</p>
1336 <div class="doc_code"><pre>
1337 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1339 Function Pass Manager
1340 Dominator Set Construction
1341 Immediate Dominators Construction
1342 Global Common Subexpression Elimination
1343 <b>-- Dominator Set Construction</b>
1344 -- Immediate Dominators Construction
1345 -- Global Common Subexpression Elimination
1346 <b> Hello World Pass
1348 Dominator Set Construction</b>
1349 Natural Loop Construction
1350 Loop Invariant Code Motion
1351 -- Natural Loop Construction
1352 -- Loop Invariant Code Motion
1354 -- Dominator Set Construction
1363 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1364 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1365 we need to add the following <a
1366 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1368 <div class="doc_code"><pre>
1369 <i>// We don't modify the program, so we preserve all analyses</i>
1370 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1371 AU.setPreservesAll();
1375 <p>Now when we run our pass, we get this output:</p>
1377 <div class="doc_code"><pre>
1378 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1379 Pass Arguments: -gcse -hello -licm
1381 Function Pass Manager
1382 Dominator Set Construction
1383 Immediate Dominators Construction
1384 Global Common Subexpression Elimination
1385 -- Immediate Dominators Construction
1386 -- Global Common Subexpression Elimination
1389 Natural Loop Construction
1390 Loop Invariant Code Motion
1391 -- Loop Invariant Code Motion
1392 -- Natural Loop Construction
1394 -- Dominator Set Construction
1403 <p>Which shows that we don't accidentally invalidate dominator information
1404 anymore, and therefore do not have to compute it twice.</p>
1408 <!-- _______________________________________________________________________ -->
1409 <div class="doc_subsubsection">
1410 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1413 <div class="doc_text">
1415 <div class="doc_code"><pre>
1416 <b>virtual void</b> releaseMemory();
1419 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1420 results, and how long to keep them around for. Because the lifetime of the pass
1421 object itself is effectively the entire duration of the compilation process, we
1422 need some way to free analysis results when they are no longer useful. The
1423 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1425 <p>If you are writing an analysis or any other pass that retains a significant
1426 amount of state (for use by another pass which "requires" your pass and uses the
1427 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1428 <tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
1429 internal state. This method is called after the <tt>run*</tt> method for the
1430 class, before the next call of <tt>run*</tt> in your pass.</p>
1434 <!-- *********************************************************************** -->
1435 <div class="doc_section">
1436 <a name="registering">Registering dynamically loaded passes</a>
1438 <!-- *********************************************************************** -->
1440 <div class="doc_text">
1442 <p><i>Size matters</i> when constructing production quality tools using llvm,
1443 both for the purposes of distribution, and for regulating the resident code size
1444 when running on the target system. Therefore, it becomes desirable to
1445 selectively use some passes, while omitting others and maintain the flexibility
1446 to change configurations later on. You want to be able to do all this, and,
1447 provide feedback to the user. This is where pass registration comes into
1450 <p>The fundamental mechanisms for pass registration are the
1451 <tt>MachinePassRegistry</tt> class and subclasses of
1452 <tt>MachinePassRegistryNode</tt>.</p>
1454 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1455 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1456 communicates additions and deletions to the command line interface.</p>
1458 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1459 information provided about a particular pass. This information includes the
1460 command line name, the command help string and the address of the function used
1461 to create an instance of the pass. A global static constructor of one of these
1462 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1463 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1464 in the tool will be registered at start up. A dynamically loaded pass will
1465 register on load and unregister at unload.</p>
1469 <!-- _______________________________________________________________________ -->
1470 <div class="doc_subsection">
1471 <a name="registering_existing">Using existing registries</a>
1474 <div class="doc_text">
1476 <p>There are predefined registries to track instruction scheduling
1477 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1478 machine passes. Here we will describe how to <i>register</i> a register
1479 allocator machine pass.</p>
1481 <p>Implement your register allocator machine pass. In your register allocator
1482 .cpp file add the following include;</p>
1484 <div class="doc_code"><pre>
1485 #include ""llvm/CodeGen/RegAllocRegistry.h""
1488 <p>Also in your register allocator .cpp file, define a creator function in the
1491 <div class="doc_code"><pre>
1492 FunctionPass *createMyRegisterAllocator() {
1493 return new MyRegisterAllocator();
1497 <p>Note that the signature of this function should match the type of
1498 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1499 "installing" declaration, in the form;</p>
1501 <div class="doc_code"><pre>
1502 static RegisterRegAlloc myRegAlloc("myregalloc",
1503 " my register allocator help string",
1504 createMyRegisterAllocator);
1507 <p>Note the two spaces prior to the help string produces a tidy result on the
1510 <div class="doc_code"><pre>
1513 -regalloc - Register allocator to use: (default = linearscan)
1514 =linearscan - linear scan register allocator
1515 =local - local register allocator
1516 =simple - simple register allocator
1517 =myregalloc - my register allocator help string
1521 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1522 an option. Registering instruction schedulers is similar except use the
1523 <tt>RegisterRegAlloc</tt> class. Note that the
1524 <tt>RegisterRegAlloc::FunctionPassCtor</tt> is significantly different from
1525 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1527 <p>To force the load/linking of your register allocator into the llc/lli tools,
1528 add your creator function's global declaration to "Passes.h" and add a "pseudo"
1529 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1534 <!-- _______________________________________________________________________ -->
1535 <div class="doc_subsection">
1536 <a name="registering_new">Creating new registries</a>
1539 <div class="doc_text">
1541 <p>The easiest way to get started is to clone one of the existing registries; we
1542 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1543 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1545 <p>Then you need to declare the registry. Example: if your pass registry is
1546 <tt>RegisterMyPasses</tt> then define;</p>
1548 <div class="doc_code"><pre>
1549 MachinePassRegistry RegisterMyPasses::Registry;
1552 <p>And finally, declare the command line option for your passes. Example:</p>
1554 <div class="doc_code"><pre>
1555 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1556 RegisterPassParser<RegisterMyPasses> >
1558 cl::init(&createDefaultMyPass),
1559 cl::desc("my pass option help"));
1562 <p>Here the command option is "mypass", with createDefaultMyPass as the default
1567 <!-- *********************************************************************** -->
1568 <div class="doc_section">
1569 <a name="debughints">Using GDB with dynamically loaded passes</a>
1571 <!-- *********************************************************************** -->
1573 <div class="doc_text">
1575 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1576 should be. First of all, you can't set a breakpoint in a shared object that has
1577 not been loaded yet, and second of all there are problems with inlined functions
1578 in shared objects. Here are some suggestions to debugging your pass with
1581 <p>For sake of discussion, I'm going to assume that you are debugging a
1582 transformation invoked by <tt>opt</tt>, although nothing described here depends
1587 <!-- _______________________________________________________________________ -->
1588 <div class="doc_subsubsection">
1589 <a name="breakpoint">Setting a breakpoint in your pass</a>
1592 <div class="doc_text">
1594 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1596 <div class="doc_code"><pre>
1599 Copyright 2000 Free Software Foundation, Inc.
1600 GDB is free software, covered by the GNU General Public License, and you are
1601 welcome to change it and/or distribute copies of it under certain conditions.
1602 Type "show copying" to see the conditions.
1603 There is absolutely no warranty for GDB. Type "show warranty" for details.
1604 This GDB was configured as "sparc-sun-solaris2.6"...
1608 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1609 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1610 (the shared object isn't loaded until runtime), we must execute the process, and
1611 have it stop before it invokes our pass, but after it has loaded the shared
1612 object. The most foolproof way of doing this is to set a breakpoint in
1613 <tt>PassManager::run</tt> and then run the process with the arguments you
1616 <div class="doc_code"><pre>
1617 (gdb) <b>break PassManager::run</b>
1618 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1619 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]</b>
1620 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
1621 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1622 70 bool PassManager::run(Module &M) { return PM->run(M); }
1626 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1627 now free to set breakpoints in your pass so that you can trace through execution
1628 or do other standard debugging stuff.</p>
1632 <!-- _______________________________________________________________________ -->
1633 <div class="doc_subsubsection">
1634 <a name="debugmisc">Miscellaneous Problems</a>
1637 <div class="doc_text">
1639 <p>Once you have the basics down, there are a couple of problems that GDB has,
1640 some with solutions, some without.</p>
1643 <li>Inline functions have bogus stack information. In general, GDB does a
1644 pretty good job getting stack traces and stepping through inline functions.
1645 When a pass is dynamically loaded however, it somehow completely loses this
1646 capability. The only solution I know of is to de-inline a function (move it
1647 from the body of a class to a .cpp file).</li>
1649 <li>Restarting the program breaks breakpoints. After following the information
1650 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1651 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1652 and you start getting errors about breakpoints being unsettable. The only way I
1653 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1654 already set in your pass, run the program, and re-set the breakpoints once
1655 execution stops in <tt>PassManager::run</tt>.</li>
1659 <p>Hopefully these tips will help with common case debugging situations. If
1660 you'd like to contribute some tips of your own, just contact <a
1661 href="mailto:sabre@nondot.org">Chris</a>.</p>
1665 <!-- *********************************************************************** -->
1666 <div class="doc_section">
1667 <a name="future">Future extensions planned</a>
1669 <!-- *********************************************************************** -->
1671 <div class="doc_text">
1673 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1674 some nifty stuff, there are things we'd like to add in the future. Here is
1675 where we are going:</p>
1679 <!-- _______________________________________________________________________ -->
1680 <div class="doc_subsubsection">
1681 <a name="SMP">Multithreaded LLVM</a>
1684 <div class="doc_text">
1686 <p>Multiple CPU machines are becoming more common and compilation can never be
1687 fast enough: obviously we should allow for a multithreaded compiler. Because of
1688 the semantics defined for passes above (specifically they cannot maintain state
1689 across invocations of their <tt>run*</tt> methods), a nice clean way to
1690 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1691 to create multiple instances of each pass object, and allow the separate
1692 instances to be hacking on different parts of the program at the same time.</p>
1694 <p>This implementation would prevent each of the passes from having to implement
1695 multithreaded constructs, requiring only the LLVM core to have locking in a few
1696 places (for global resources). Although this is a simple extension, we simply
1697 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1698 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1702 <!-- _______________________________________________________________________ -->
1703 <div class="doc_subsubsection">
1704 <a name="PassFunctionPass"><tt>ModulePass</tt>es requiring <tt>FunctionPass</tt>es</a>
1707 <div class="doc_text">
1709 <p>Currently it is illegal for a <a href="#ModulePass"><tt>ModulePass</tt></a>
1710 to require a <a href="#FunctionPass"><tt>FunctionPass</tt></a>. This is because
1711 there is only one instance of the <a
1712 href="#FunctionPass"><tt>FunctionPass</tt></a> object ever created, thus nowhere
1713 to store information for all of the functions in the program at the same time.
1714 Although this has come up a couple of times before, this has always been worked
1715 around by factoring one big complicated pass into a global and an
1716 interprocedural part, both of which are distinct. In the future, it would be
1717 nice to have this though.</p>
1719 <p>Note that it is no problem for a <a
1720 href="#FunctionPass"><tt>FunctionPass</tt></a> to require the results of a <a
1721 href="#ModulePass"><tt>ModulePass</tt></a>, only the other way around.</p>
1725 <!-- *********************************************************************** -->
1728 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1729 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1730 <a href="http://validator.w3.org/check/referer"><img
1731 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!" /></a>
1733 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1734 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1735 Last modified: $Date$