1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <title>Writing an LLVM Pass</title>
6 <link rel="stylesheet" href="llvm.css" type="text/css">
10 <div class="doc_title">
15 <li><a href="#introduction">Introduction - What is a pass?</a></li>
16 <li><a href="#quickstart">Quick Start - Writing hello world</a>
18 <li><a href="#makefile">Setting up the build environment</a></li>
19 <li><a href="#basiccode">Basic code required</a></li>
20 <li><a href="#running">Running a pass with <tt>opt</tt>
21 or <tt>analyze</tt></a></li>
23 <li><a href="#passtype">Pass classes and requirements</a>
25 <li><a href="#ImmutablePass">The <tt>ImmutablePass</tt> class</a></li>
26 <li><a href="#ModulePass">The <tt>ModulePass</tt> class</a>
28 <li><a href="#runOnModule">The <tt>runOnModule</tt> method</a></li>
30 <li><a href="#CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
32 <li><a href="#doInitialization_scc">The <tt>doInitialization(Module
33 &)</tt> method</a></li>
34 <li><a href="#runOnSCC">The <tt>runOnSCC</tt> method</a></li>
35 <li><a href="#doFinalization_scc">The <tt>doFinalization(Module
36 &)</tt> method</a></li>
38 <li><a href="#FunctionPass">The <tt>FunctionPass</tt> class</a>
40 <li><a href="#doInitialization_mod">The <tt>doInitialization(Module
41 &)</tt> method</a></li>
42 <li><a href="#runOnFunction">The <tt>runOnFunction</tt> method</a></li>
43 <li><a href="#doFinalization_mod">The <tt>doFinalization(Module
44 &)</tt> method</a></li>
46 <li><a href="#BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
48 <li><a href="#doInitialization_fn">The <tt>doInitialization(Function
49 &)</tt> method</a></li>
50 <li><a href="#runOnBasicBlock">The <tt>runOnBasicBlock</tt>
52 <li><a href="#doFinalization_fn">The <tt>doFinalization(Function
53 &)</tt> method</a></li>
55 <li><a href="#MachineFunctionPass">The <tt>MachineFunctionPass</tt>
58 <li><a href="#runOnMachineFunction">The
59 <tt>runOnMachineFunction(MachineFunction &)</tt> method</a></li>
62 <li><a href="#registration">Pass Registration</a>
64 <li><a href="#print">The <tt>print</tt> method</a></li>
66 <li><a href="#interaction">Specifying interactions between passes</a>
68 <li><a href="#getAnalysisUsage">The <tt>getAnalysisUsage</tt>
70 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
71 <li><a href="#AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a></li>
72 <li><a href="#AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a></li>
73 <li><a href="#getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a></li>
75 <li><a href="#analysisgroup">Implementing Analysis Groups</a>
77 <li><a href="#agconcepts">Analysis Group Concepts</a></li>
78 <li><a href="#registerag">Using <tt>RegisterAnalysisGroup</tt></a></li>
80 <li><a href="#passStatistics">Pass Statistics</a>
81 <li><a href="#passmanager">What PassManager does</a>
83 <li><a href="#releaseMemory">The <tt>releaseMemory</tt> method</a></li>
85 <li><a href="#debughints">Using GDB with dynamically loaded passes</a>
87 <li><a href="#breakpoint">Setting a breakpoint in your pass</a></li>
88 <li><a href="#debugmisc">Miscellaneous Problems</a></li>
90 <li><a href="#future">Future extensions planned</a>
92 <li><a href="#SMP">Multithreaded LLVM</a></li>
93 <li><a href="#PassFunctionPass"><tt>ModulePass</tt>es requiring
94 <tt>FunctionPass</tt>es</a></li>
98 <div class="doc_author">
99 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
102 <!-- *********************************************************************** -->
103 <div class="doc_section">
104 <a name="introduction">Introduction - What is a pass?</a>
106 <!-- *********************************************************************** -->
108 <div class="doc_text">
110 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
111 passes are where most of the interesting parts of the compiler exist. Passes
112 perform the transformations and optimizations that make up the compiler, they
113 build the analysis results that are used by these transformations, and they are,
114 above all, a structuring technique for compiler code.</p>
116 <p>All LLVM passes are subclasses of the <tt><a
117 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
118 class, which implement functionality by overriding virtual methods inherited
119 from <tt>Pass</tt>. Depending on how your pass works, you should inherit from
120 the <tt><a href="#ModulePass">ModulePass</a></tt>, <tt><a
121 href="#CallGraphSCCPass">CallGraphSCCPass</a></tt>, <tt><a
122 href="#FunctionPass">FunctionPass</a></tt>, or <tt><a
123 href="#BasicBlockPass">BasicBlockPass</a></tt> classes, which gives the system
124 more information about what your pass does, and how it can be combined with
125 other passes. One of the main features of the LLVM Pass Framework is that it
126 schedules passes to run in an efficient way based on the constraints that your
127 pass meets (which are indicated by which class they derive from).</p>
129 <p>We start by showing you how to construct a pass, everything from setting up
130 the code, to compiling, loading, and executing it. After the basics are down,
131 more advanced features are discussed.</p>
135 <!-- *********************************************************************** -->
136 <div class="doc_section">
137 <a name="quickstart">Quick Start - Writing hello world</a>
139 <!-- *********************************************************************** -->
141 <div class="doc_text">
143 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
144 is designed to simply print out the name of non-external functions that exist in
145 the program being compiled. It does not modify the program at all, it just
146 inspects it. The source code and files for this pass are available in the LLVM
147 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
151 <!-- ======================================================================= -->
152 <div class="doc_subsection">
153 <a name="makefile">Setting up the build environment</a>
156 <div class="doc_text">
158 <p>First, you need to create a new directory somewhere in the LLVM source
159 base. For this example, we'll assume that you made
160 <tt>lib/Transforms/Hello</tt>. Next, you must set up a build script
161 (Makefile) that will compile the source code for the new pass. To do this,
162 copy the following into <tt>Makefile</tt>:</p>
166 # Makefile for hello pass
168 # Path to top level of LLVM heirarchy
171 # Name of the library to build
174 # Build a dynamically loadable shared object
177 # Include the makefile implementation stuff
178 include $(LEVEL)/Makefile.common
181 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
182 directory are to be compiled and linked together into a
183 <tt>lib/Debug/libhello.so</tt> shared object that can be dynamically loaded by
184 the <tt>opt</tt> or <tt>analyze</tt> tools. If your operating system uses a
185 suffix other than .so (such as windows or Mac OS/X), the appropriate extension
188 <p>Now that we have the build scripts set up, we just need to write the code for
193 <!-- ======================================================================= -->
194 <div class="doc_subsection">
195 <a name="basiccode">Basic code required</a>
198 <div class="doc_text">
200 <p>Now that we have a way to compile our new pass, we just have to write it.
204 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
205 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Function_8h-source.html">llvm/Function.h</a>"
208 <p>Which are needed because we are writing a <tt><a
209 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">Pass</a></tt>, and
210 we are operating on <tt><a
211 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Function.html">Function</a></tt>'s.</p>
215 <b>using namespace llvm;</b>
217 <p>... which is required because the functions from the include files
218 live in the llvm namespace.
227 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
228 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
229 things declared inside of the anonymous namespace only visible to the current
230 file. If you're not familiar with them, consult a decent C++ book for more
233 <p>Next, we declare our pass itself:</p>
236 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
239 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
240 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
241 The different builtin pass subclasses are described in detail <a
242 href="#passtype">later</a>, but for now, know that <a
243 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate a function at a
247 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
248 std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
251 }; <i>// end of struct Hello</i>
254 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
255 which overloads an abstract virtual method inherited from <a
256 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
257 to do our thing, so we just print out our message with the name of each
261 RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
262 } <i>// end of anonymous namespace</i>
265 <p>Lastly, we register our class <tt>Hello</tt>, giving it a command line
266 argument "<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>". There are
267 several different ways of <a href="#registration">registering your pass</a>,
268 depending on what it is to be used for. For "optimizations" we use the
269 <tt>RegisterOpt</tt> template.</p>
271 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
274 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
275 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Function_8h-source.html">llvm/Function.h</a>"
277 <b>using namespace llvm;</b>
280 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
281 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
282 std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
287 RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
291 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
292 command in the local directory and you should get a new
293 "<tt>lib/Debug/libhello.so</tt> file. Note that everything in this file is
294 contained in an anonymous namespace: this reflects the fact that passes are self
295 contained units that do not need external interfaces (although they can have
296 them) to be useful.</p>
300 <!-- ======================================================================= -->
301 <div class="doc_subsection">
302 <a name="running">Running a pass with <tt>opt</tt> or <tt>analyze</tt></a>
305 <div class="doc_text">
307 <p>Now that you have a brand new shiny shared object file, we can use the
308 <tt>opt</tt> command to run an LLVM program through your pass. Because you
309 registered your pass with the <tt>RegisterOpt</tt> template, you will be able to
310 use the <tt>opt</tt> tool to access it, once loaded.</p>
312 <p>To test it, follow the example at the end of the <a
313 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
314 LLVM. We can now run the bytecode file (<tt>hello.bc</tt>) for the program
315 through our transformation like this (or course, any bytecode file will
319 $ opt -load ../../../lib/Debug/libhello.so -hello < hello.bc > /dev/null
325 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
326 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
327 argument (which is one reason you need to <a href="#registration">register your
328 pass</a>). Because the hello pass does not modify the program in any
329 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
330 <tt>/dev/null</tt>).</p>
332 <p>To see what happened to the other string you registered, try running
333 <tt>opt</tt> with the <tt>--help</tt> option:</p>
336 $ opt -load ../../../lib/Debug/libhello.so --help
337 OVERVIEW: llvm .bc -> .bc modular optimizer
339 USAGE: opt [options] <input bytecode>
342 Optimizations available:
344 -funcresolve - Resolve Functions
345 -gcse - Global Common Subexpression Elimination
346 -globaldce - Dead Global Elimination
347 <b>-hello - Hello World Pass</b>
348 -indvars - Canonicalize Induction Variables
349 -inline - Function Integration/Inlining
350 -instcombine - Combine redundant instructions
354 <p>The pass name get added as the information string for your pass, giving some
355 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
356 would go ahead and make it do the cool transformations you want. Once you get
357 it all working and tested, it may become useful to find out how fast your pass
358 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
359 line option (<tt>--time-passes</tt>) that allows you to get information about
360 the execution time of your pass along with the other passes you queue up. For
364 $ opt -load ../../../lib/Debug/libhello.so -hello -time-passes < hello.bc > /dev/null
368 ===============================================================================
369 ... Pass execution timing report ...
370 ===============================================================================
371 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
373 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
374 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bytecode Writer
375 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
376 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
377 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
378 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
381 <p>As you can see, our implementation above is pretty fast :). The additional
382 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
383 that the LLVM emitted by your pass is still valid and well formed LLVM, which
384 hasn't been broken somehow.</p>
386 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
387 about some more details of how they work and how to use them.</p>
391 <!-- *********************************************************************** -->
392 <div class="doc_section">
393 <a name="passtype">Pass classes and requirements</a>
395 <!-- *********************************************************************** -->
397 <div class="doc_text">
399 <p>One of the first things that you should do when designing a new pass is to
400 decide what class you should subclass for your pass. The <a
401 href="#basiccode">Hello World</a> example uses the <tt><a
402 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
403 did not discuss why or when this should occur. Here we talk about the classes
404 available, from the most general to the most specific.</p>
406 <p>When choosing a superclass for your Pass, you should choose the <b>most
407 specific</b> class possible, while still being able to meet the requirements
408 listed. This gives the LLVM Pass Infrastructure information necessary to
409 optimize how passes are run, so that the resultant compiler isn't unneccesarily
414 <!-- ======================================================================= -->
415 <div class="doc_subsection">
416 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
419 <div class="doc_text">
421 <p>The most plain and boring type of pass is the "<tt><a
422 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
423 class. This pass type is used for passes that do not have to be run, do not
424 change state, and never need to be updated. This is not a normal type of
425 transformation or analysis, but can provide information about the current
426 compiler configuration.</p>
428 <p>Although this pass class is very infrequently used, it is important for
429 providing information about the current target machine being compiled for, and
430 other static information that can affect the various transformations.</p>
432 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
433 invalidated, and are never "run".</p>
437 <!-- ======================================================================= -->
438 <div class="doc_subsection">
439 <a name="ModulePass">The <tt>ModulePass</tt> class</a>
442 <div class="doc_text">
445 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1ModulePass.html">ModulePass</a></tt>"
446 class is the most general of all superclasses that you can use. Deriving from
447 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
448 refering to function bodies in no predictable order, or adding and removing
449 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
450 subclasses, no optimization can be done for their execution.</p>
452 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
453 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
454 following signature:</p>
458 <!-- _______________________________________________________________________ -->
459 <div class="doc_subsubsection">
460 <a name="runOnModule">The <tt>runOnModule</tt> method</a>
463 <div class="doc_text">
466 <b>virtual bool</b> runOnModule(Module &M) = 0;
469 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass,
470 and should return true if the module was modified by the transformation, false
475 <!-- ======================================================================= -->
476 <div class="doc_subsection">
477 <a name="CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
480 <div class="doc_text">
483 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1CallGraphSCCPass.html">CallGraphSCCPass</a></tt>"
484 is used by passes that need to traverse the program bottom-up on the call graph
485 (callees before callers). Deriving from CallGraphSCCPass provides some
486 mechanics for building and traversing the CallGraph, but also allows the system
487 to optimize execution of CallGraphSCCPass's. If your pass meets the
488 requirements outlined below, and doesn't meet the requirements of a <tt><a
489 href="#FunctionPass">FunctionPass</a></tt> or <tt><a
490 href="#BasicBlockPass">BasicBlockPass</a></tt>, you should derive from
491 <tt>CallGraphSCCPass</tt>.</p>
493 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
495 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
499 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
500 the current SCC.</li>
502 <li>... <em>allowed</em> to inspect any Function's other than those in the
503 current SCC and the direct callees of the SCC.</li>
505 <li>... <em>required</em> to preserve the current CallGraph object, updating it
506 to reflect any changes made to the program.</li>
508 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
509 though they may change the contents of an SCC.</li>
511 <li>... <em>allowed</em> to add or remove global variables from the current
514 <li>... <em>allowed</em> to maintain state across invocations of
515 <a href="#runOnSCC"><tt>runOnSCC</tt></a> (including global data).</li>
518 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
519 because it has to handle SCCs with more than one node in it. All of the virtual
520 methods described below should return true if they modified the program, or
521 false if they didn't.</p>
525 <!-- _______________________________________________________________________ -->
526 <div class="doc_subsubsection">
527 <a name="doInitialization_scc">The <tt>doInitialization(Module &)</tt>
531 <div class="doc_text">
534 <b>virtual bool</b> doInitialization(Module &M);
537 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
538 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
539 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
540 is designed to do simple initialization type of stuff that does not depend on
541 the SCCs being processed. The <tt>doInitialization</tt> method call is not
542 scheduled to overlap with any other pass executions (thus it should be very
547 <!-- _______________________________________________________________________ -->
548 <div class="doc_subsubsection">
549 <a name="runOnSCC">The <tt>runOnSCC</tt> method</a>
552 <div class="doc_text">
555 <b>virtual bool</b> runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;
558 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
559 should return true if the module was modified by the transformation, false
564 <!-- _______________________________________________________________________ -->
565 <div class="doc_subsubsection">
566 <a name="doFinalization_scc">The <tt>doFinalization(Module
567 &)</tt> method</a>
570 <div class="doc_text">
573 <b>virtual bool</b> doFinalization(Module &M);
576 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
577 called when the pass framework has finished calling <a
578 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
579 program being compiled.</p>
583 <!-- ======================================================================= -->
584 <div class="doc_subsection">
585 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
588 <div class="doc_text">
590 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
591 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
592 subclasses do have a predictable, local behavior that can be expected by the
593 system. All <tt>FunctionPass</tt> execute on each function in the program
594 independent of all of the other functions in the program.
595 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
596 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
598 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
601 <li>Modify a Function other than the one currently being processed.</li>
602 <li>Add or remove Function's from the current Module.</li>
603 <li>Add or remove global variables from the current Module.</li>
604 <li>Maintain state across invocations of
605 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
608 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
609 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
610 may overload three virtual methods to do their work. All of these methods
611 should return true if they modified the program, or false if they didn't.</p>
615 <!-- _______________________________________________________________________ -->
616 <div class="doc_subsubsection">
617 <a name="doInitialization_mod">The <tt>doInitialization(Module &)</tt>
621 <div class="doc_text">
624 <b>virtual bool</b> doInitialization(Module &M);
627 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
628 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
629 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
630 is designed to do simple initialization type of stuff that does not depend on
631 the functions being processed. The <tt>doInitialization</tt> method call is not
632 scheduled to overlap with any other pass executions (thus it should be very
635 <p>A good example of how this method should be used is the <a
636 href="http://llvm.cs.uiuc.edu/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
637 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
638 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
639 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
640 free functions that it needs, adding prototypes to the module if necessary.</p>
644 <!-- _______________________________________________________________________ -->
645 <div class="doc_subsubsection">
646 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
649 <div class="doc_text">
652 <b>virtual bool</b> runOnFunction(Function &F) = 0;
655 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
656 the transformation or analysis work of your pass. As usual, a true value should
657 be returned if the function is modified.</p>
661 <!-- _______________________________________________________________________ -->
662 <div class="doc_subsubsection">
663 <a name="doFinalization_mod">The <tt>doFinalization(Module
664 &)</tt> method</a>
667 <div class="doc_text">
670 <b>virtual bool</b> doFinalization(Module &M);
673 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
674 called when the pass framework has finished calling <a
675 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
676 program being compiled.</p>
680 <!-- ======================================================================= -->
681 <div class="doc_subsection">
682 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
685 <div class="doc_text">
687 <p><tt>BasicBlockPass</tt>'s are just like <a
688 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
689 their scope of inspection and modification to a single basic block at a time.
690 As such, they are <b>not</b> allowed to do any of the following:</p>
693 <li>Modify or inspect any basic blocks outside of the current one</li>
694 <li>Maintain state across invocations of
695 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
696 <li>Modify the control flow graph (by altering terminator instructions)</li>
697 <li>Any of the things forbidden for
698 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
701 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
702 optimizations. They may override the same <a
703 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
704 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
705 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
709 <!-- _______________________________________________________________________ -->
710 <div class="doc_subsubsection">
711 <a name="doInitialization_fn">The <tt>doInitialization(Function
712 &)</tt> method</a>
715 <div class="doc_text">
718 <b>virtual bool</b> doInitialization(Function &F);
721 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
722 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
723 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
724 to do simple initialization that does not depend on the
725 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
726 scheduled to overlap with any other pass executions (thus it should be very
731 <!-- _______________________________________________________________________ -->
732 <div class="doc_subsubsection">
733 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
736 <div class="doc_text">
739 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
742 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
743 function is not allowed to inspect or modify basic blocks other than the
744 parameter, and are not allowed to modify the CFG. A true value must be returned
745 if the basic block is modified.</p>
749 <!-- _______________________________________________________________________ -->
750 <div class="doc_subsubsection">
751 <a name="doFinalization_fn">The <tt>doFinalization(Function &)</tt>
755 <div class="doc_text">
758 <b>virtual bool</b> doFinalization(Function &F);
761 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
762 called when the pass framework has finished calling <a
763 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
764 program being compiled. This can be used to perform per-function
769 <!-- ======================================================================= -->
770 <div class="doc_subsection">
771 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
774 <div class="doc_text">
776 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
777 executes on the machine-dependent representation of each LLVM function in the
778 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
779 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
780 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
781 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
784 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
785 <li>Modify a MachineFunction other than the one currently being processed.</li>
786 <li>Add or remove MachineFunctions from the current Module.</li>
787 <li>Add or remove global variables from the current Module.</li>
788 <li>Maintain state across invocations of <a
789 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
795 <!-- _______________________________________________________________________ -->
796 <div class="doc_subsubsection">
797 <a name="runOnMachineFunction">The <tt>runOnMachineFunction(MachineFunction
798 &MF)</tt> method</a>
801 <div class="doc_text">
804 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
807 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
808 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
809 work of your <tt>MachineFunctionPass</tt>.</p>
811 <p>The <tt>runOnMachineFunction</tt> method is called on every
812 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
813 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
814 representation of the function. If you want to get at the LLVM <tt>Function</tt>
815 for the <tt>MachineFunction</tt> you're working on, use
816 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
817 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
818 <tt>MachineFunctionPass</tt>.</p>
822 <!-- *********************************************************************** -->
823 <div class="doc_section">
824 <a name="registration">Pass registration</a>
826 <!-- *********************************************************************** -->
828 <div class="doc_text">
830 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
831 pass registration works, and discussed some of the reasons that it is used and
832 what it does. Here we discuss how and why passes are registered.</p>
834 <p>Passes can be registered in several different ways. Depending on the general
835 classification of the pass, you should use one of the following templates to
836 register the pass:</p>
839 <li><b><tt>RegisterOpt</tt></b> - This template should be used when you are
840 registering a pass that logically should be available for use in the
841 '<tt>opt</tt>' utility.</li>
843 <li><b><tt>RegisterAnalysis</tt></b> - This template should be used when you are
844 registering a pass that logically should be available for use in the
845 '<tt>analyze</tt>' utility.</li>
847 <li><b><tt>RegisterPass</tt></b> - This is the generic form of the
848 <tt>Register*</tt> templates that should be used if you want your pass listed by
849 multiple or no utilities. This template takes an extra third argument that
850 specifies which tools it should be listed in. See the <a
851 href="http://llvm.cs.uiuc.edu/doxygen/PassSupport_8h-source.html">PassSupport.h</a>
852 file for more information.</li>
856 <p>Regardless of how you register your pass, you must specify at least two
857 parameters. The first parameter is the name of the pass that is to be used on
858 the command line to specify that the pass should be added to a program (for
859 example <tt>opt</tt> or <tt>analyze</tt>). The second argument is the name of
860 the pass, which is to be used for the <tt>--help</tt> output of programs, as
861 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
863 <p>If a pass is registered to be used by the <tt>analyze</tt> utility, you
864 should implement the virtual <tt>print</tt> method:</p>
868 <!-- _______________________________________________________________________ -->
869 <div class="doc_subsubsection">
870 <a name="print">The <tt>print</tt> method</a>
873 <div class="doc_text">
876 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
879 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
880 a human readable version of the analysis results. This is useful for debugging
881 an analysis itself, as well as for other people to figure out how an analysis
882 works. The <tt>analyze</tt> tool uses this method to generate its output.</p>
884 <p>The <tt>ostream</tt> parameter specifies the stream to write the results on,
885 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
886 program that has been analyzed. Note however that this pointer may be null in
887 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
888 debugger), so it should only be used to enhance debug output, it should not be
893 <!-- *********************************************************************** -->
894 <div class="doc_section">
895 <a name="interaction">Specifying interactions between passes</a>
897 <!-- *********************************************************************** -->
899 <div class="doc_text">
901 <p>One of the main responsibilities of the <tt>PassManager</tt> is the make sure
902 that passes interact with each other correctly. Because <tt>PassManager</tt>
903 tries to <a href="#passmanager">optimize the execution of passes</a> it must
904 know how the passes interact with each other and what dependencies exist between
905 the various passes. To track this, each pass can declare the set of passes that
906 are required to be executed before the current pass, and the passes which are
907 invalidated by the current pass.</p>
909 <p>Typically this functionality is used to require that analysis results are
910 computed before your pass is run. Running arbitrary transformation passes can
911 invalidate the computed analysis results, which is what the invalidation set
912 specifies. If a pass does not implement the <tt><a
913 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
914 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
918 <!-- _______________________________________________________________________ -->
919 <div class="doc_subsubsection">
920 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
923 <div class="doc_text">
926 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
929 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
930 invalidated sets may be specified for your transformation. The implementation
931 should fill in the <tt><a
932 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
933 object with information about which passes are required and not invalidated. To
934 do this, a pass may call any of the following methods on the AnalysisUsage
938 <!-- _______________________________________________________________________ -->
939 <div class="doc_subsubsection">
940 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
943 <div class="doc_text">
945 If your pass requires a previous pass to be executed (an analysis for example),
946 it can use one of these methods to arrange for it to be run before your pass.
947 LLVM has many different types of analyses and passes that can be required,
948 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
949 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
950 be no critical edges in the CFG when your pass has been run.
954 Some analyses chain to other analyses to do their job. For example, an <a
955 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
956 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
957 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
958 used instead of the <tt>addRequired</tt> method. This informs the PassManager
959 that the transitively required pass should be alive as long as the requiring
964 <!-- _______________________________________________________________________ -->
965 <div class="doc_subsubsection">
966 <a name="AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a>
969 <div class="doc_text">
971 One of the jobs of the PassManager is to optimize how and when analyses are run.
972 In particular, it attempts to avoid recomputing data unless it needs to. For
973 this reason, passes are allowed to declare that they preserve (i.e., they don't
974 invalidate) an existing analysis if it's available. For example, a simple
975 constant folding pass would not modify the CFG, so it can't possibly affect the
976 results of dominator analysis. By default, all passes are assumed to invalidate
981 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
982 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
983 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
984 not modify the LLVM program at all (which is true for analyses), and the
985 <tt>setPreservesCFG</tt> method can be used by transformations that change
986 instructions in the program but do not modify the CFG or terminator instructions
987 (note that this property is implicitly set for <a
988 href="#BasicBlockPass">BasicBlockPass</a>'s).
992 <tt>addPreserved</tt> is particularly useful for transformations like
993 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
994 and dominator related analyses if they exist, so it can preserve them, despite
995 the fact that it hacks on the CFG.
999 <!-- _______________________________________________________________________ -->
1000 <div class="doc_subsubsection">
1001 <a name="AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a>
1004 <div class="doc_text">
1007 <i>// This is an example implementation from an analysis, which does not modify
1008 // the program at all, yet has a prerequisite.</i>
1009 <b>void</b> <a href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1010 AU.setPreservesAll();
1011 AU.addRequired<<a href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1018 <i>// This example modifies the program, but does not modify the CFG</i>
1019 <b>void</b> <a href="http://llvm.cs.uiuc.edu/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1020 AU.setPreservesCFG();
1021 AU.addRequired<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1027 <!-- _______________________________________________________________________ -->
1028 <div class="doc_subsubsection">
1029 <a name="getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a>
1032 <div class="doc_text">
1034 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1035 your class, providing you with access to the passes that you declared that you
1036 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
1037 method. It takes a single template argument that specifies which pass class you
1038 want, and returns a reference to that pass. For example:</p>
1041 bool LICM::runOnFunction(Function &F) {
1042 LoopInfo &LI = getAnalysis<LoopInfo>();
1047 <p>This method call returns a reference to the pass desired. You may get a
1048 runtime assertion failure if you attempt to get an analysis that you did not
1049 declare as required in your <a
1050 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
1051 method can be called by your <tt>run*</tt> method implementation, or by any
1052 other local method invoked by your <tt>run*</tt> method.</p>
1055 If your pass is capable of updating analyses if they exist (e.g.,
1056 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1057 <tt>getAnalysisToUpdate</tt> method, which returns a pointer to the analysis if
1058 it is active. For example:</p>
1062 if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
1063 <i>// A DominatorSet is active. This code will update it.</i>
1070 <!-- *********************************************************************** -->
1071 <div class="doc_section">
1072 <a name="analysisgroup">Implementing Analysis Groups</a>
1074 <!-- *********************************************************************** -->
1076 <div class="doc_text">
1078 <p>Now that we understand the basics of how passes are defined, how the are
1079 used, and how they are required from other passes, it's time to get a little bit
1080 fancier. All of the pass relationships that we have seen so far are very
1081 simple: one pass depends on one other specific pass to be run before it can run.
1082 For many applications, this is great, for others, more flexibility is
1085 <p>In particular, some analyses are defined such that there is a single simple
1086 interface to the analysis results, but multiple ways of calculating them.
1087 Consider alias analysis for example. The most trivial alias analysis returns
1088 "may alias" for any alias query. The most sophisticated analysis a
1089 flow-sensitive, context-sensitive interprocedural analysis that can take a
1090 significant amount of time to execute (and obviously, there is a lot of room
1091 between these two extremes for other implementations). To cleanly support
1092 situations like this, the LLVM Pass Infrastructure supports the notion of
1093 Analysis Groups.</p>
1097 <!-- _______________________________________________________________________ -->
1098 <div class="doc_subsubsection">
1099 <a name="agconcepts">Analysis Group Concepts</a>
1102 <div class="doc_text">
1104 <p>An Analysis Group is a single simple interface that may be implemented by
1105 multiple different passes. Analysis Groups can be given human readable names
1106 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1107 class. An analysis group may have one or more implementations, one of which is
1108 the "default" implementation.</p>
1110 <p>Analysis groups are used by client passes just like other passes are: the
1111 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1112 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
1113 scans the available passes to see if any implementations of the analysis group
1114 are available. If none is available, the default implementation is created for
1115 the pass to use. All standard rules for <A href="#interaction">interaction
1116 between passes</a> still apply.</p>
1118 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
1119 passes, all analysis group implementations must be registered, and must use the
1120 <A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
1121 implementation pool. Also, a default implementation of the interface
1122 <b>must</b> be registered with <A
1123 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
1125 <p>As a concrete example of an Analysis Group in action, consider the <a
1126 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1127 analysis group. The default implementation of the alias analysis interface (the
1129 href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1130 pass) just does a few simple checks that don't require significant analysis to
1131 compute (such as: two different globals can never alias each other, etc).
1132 Passes that use the <tt><a
1133 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1134 interface (for example the <tt><a
1135 href="http://llvm.cs.uiuc.edu/doxygen/structGCSE.html">gcse</a></tt> pass), do
1136 not care which implementation of alias analysis is actually provided, they just
1137 use the designated interface.</p>
1139 <p>From the user's perspective, commands work just like normal. Issuing the
1140 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1141 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1142 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1143 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1144 hypothetical example) instead.</p>
1148 <!-- _______________________________________________________________________ -->
1149 <div class="doc_subsubsection">
1150 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1153 <div class="doc_text">
1155 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1156 group itself as well as add pass implementations to the analysis group. First,
1157 an analysis should be registered, with a human readable name provided for it.
1158 Unlike registration of passes, there is no command line argument to be specified
1159 for the Analysis Group Interface itself, because it is "abstract":</p>
1162 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1165 <p>Once the analysis is registered, passes can declare that they are valid
1166 implementations of the interface by using the following code:</p>
1170 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1171 RegisterOpt<FancyAA>
1172 B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
1174 //<i> Declare that we implement the AliasAnalysis interface</i>
1175 RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, FancyAA> C;
1179 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1180 uses the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
1181 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1182 analysis group. Every implementation of an analysis group should join using
1183 this template. A single pass may join multiple different analysis groups with
1188 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1189 RegisterOpt<<a href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1190 D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
1192 //<i> Declare that we implement the AliasAnalysis interface</i>
1193 RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <a href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>, <b>true</b>> E;
1197 <p>Here we show how the default implementation is specified (using the extra
1198 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1199 one default implementation available at all times for an Analysis Group to be
1200 used. Here we declare that the <tt><a
1201 href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1202 pass is the default implementation for the interface.</p>
1206 <!-- *********************************************************************** -->
1207 <div class="doc_section">
1208 <a name="passStatistics">Pass Statistics</a>
1210 <!-- *********************************************************************** -->
1212 <div class="doc_text">
1214 href="http://llvm.cs.uiuc.edu/doxygen/Statistic_8h-source.html"><tt>Statistic</tt></a>
1215 class, is designed to be an easy way to expose various success
1216 metrics from passes. These statistics are printed at the end of a
1217 run, when the -stats command line option is enabled on the command
1218 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1223 <!-- *********************************************************************** -->
1224 <div class="doc_section">
1225 <a name="passmanager">What PassManager does</a>
1227 <!-- *********************************************************************** -->
1229 <div class="doc_text">
1232 href="http://llvm.cs.uiuc.edu/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1234 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1PassManager.html">class</a>
1235 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1236 are set up correctly, and then schedules passes to run efficiently. All of the
1237 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1240 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1241 time of a series of passes:</p>
1244 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1245 recomputing analysis results as much as possible. This means keeping track of
1246 which analyses are available already, which analyses get invalidated, and which
1247 analyses are needed to be run for a pass. An important part of work is that the
1248 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1249 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1250 results as soon as they are no longer needed.</li>
1252 <li><b>Pipeline the execution of passes on the program</b> - The
1253 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1254 of a series of passes by pipelining the passes together. This means that, given
1255 a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1256 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1257 the first function, then all of the <a
1258 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1259 etc... until the entire program has been run through the passes.
1261 <p>This improves the cache behavior of the compiler, because it is only touching
1262 the LLVM program representation for a single function at a time, instead of
1263 traversing the entire program. It reduces the memory consumption of compiler,
1264 because, for example, only one <a
1265 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1266 needs to be calculated at a time. This also makes it possible some <a
1267 href="#SMP">interesting enhancements</a> in the future.</p></li>
1271 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1272 much information it has about the behaviors of the passes it is scheduling. For
1273 example, the "preserved" set is intentionally conservative in the face of an
1274 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1275 Not implementing when it should be implemented will have the effect of not
1276 allowing any analysis results to live across the execution of your pass.</p>
1278 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1279 options that is useful for debugging pass execution, seeing how things work, and
1280 diagnosing when you should be preserving more analyses than you currently are
1281 (To get information about all of the variants of the <tt>--debug-pass</tt>
1282 option, just type '<tt>opt --help-hidden</tt>').</p>
1284 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1285 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1286 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1289 $ opt -load ../../../lib/Debug/libhello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1291 Function Pass Manager
1292 Dominator Set Construction
1293 Immediate Dominators Construction
1294 Global Common Subexpression Elimination
1295 -- Immediate Dominators Construction
1296 -- Global Common Subexpression Elimination
1297 Natural Loop Construction
1298 Loop Invariant Code Motion
1299 -- Natural Loop Construction
1300 -- Loop Invariant Code Motion
1302 -- Dominator Set Construction
1308 <p>This output shows us when passes are constructed and when the analysis
1309 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1310 GCSE uses dominator and immediate dominator information to do its job. The LICM
1311 pass uses natural loop information, which uses dominator sets, but not immediate
1312 dominators. Because immediate dominators are no longer useful after the GCSE
1313 pass, it is immediately destroyed. The dominator sets are then reused to
1314 compute natural loop information, which is then used by the LICM pass.</p>
1316 <p>After the LICM pass, the module verifier runs (which is automatically added
1317 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1318 resultant LLVM code is well formed. After it finishes, the dominator set
1319 information is destroyed, after being computed once, and shared by three
1322 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1323 World</a> pass in between the two passes:</p>
1326 $ opt -load ../../../lib/Debug/libhello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1328 Function Pass Manager
1329 Dominator Set Construction
1330 Immediate Dominators Construction
1331 Global Common Subexpression Elimination
1332 <b>-- Dominator Set Construction</b>
1333 -- Immediate Dominators Construction
1334 -- Global Common Subexpression Elimination
1335 <b> Hello World Pass
1337 Dominator Set Construction</b>
1338 Natural Loop Construction
1339 Loop Invariant Code Motion
1340 -- Natural Loop Construction
1341 -- Loop Invariant Code Motion
1343 -- Dominator Set Construction
1352 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1353 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1354 we need to add the following <a
1355 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1358 <i>// We don't modify the program, so we preserve all analyses</i>
1359 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1360 AU.setPreservesAll();
1364 <p>Now when we run our pass, we get this output:</p>
1367 $ opt -load ../../../lib/Debug/libhello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1368 Pass Arguments: -gcse -hello -licm
1370 Function Pass Manager
1371 Dominator Set Construction
1372 Immediate Dominators Construction
1373 Global Common Subexpression Elimination
1374 -- Immediate Dominators Construction
1375 -- Global Common Subexpression Elimination
1378 Natural Loop Construction
1379 Loop Invariant Code Motion
1380 -- Loop Invariant Code Motion
1381 -- Natural Loop Construction
1383 -- Dominator Set Construction
1392 <p>Which shows that we don't accidentally invalidate dominator information
1393 anymore, and therefore do not have to compute it twice.</p>
1397 <!-- _______________________________________________________________________ -->
1398 <div class="doc_subsubsection">
1399 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1402 <div class="doc_text">
1405 <b>virtual void</b> releaseMemory();
1408 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1409 results, and how long to keep them around for. Because the lifetime of the pass
1410 object itself is effectively the entire duration of the compilation process, we
1411 need some way to free analysis results when they are no longer useful. The
1412 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1414 <p>If you are writing an analysis or any other pass that retains a significant
1415 amount of state (for use by another pass which "requires" your pass and uses the
1416 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1417 <tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
1418 internal state. This method is called after the <tt>run*</tt> method for the
1419 class, before the next call of <tt>run*</tt> in your pass.</p>
1423 <!-- *********************************************************************** -->
1424 <div class="doc_section">
1425 <a name="debughints">Using GDB with dynamically loaded passes</a>
1427 <!-- *********************************************************************** -->
1429 <div class="doc_text">
1431 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1432 should be. First of all, you can't set a breakpoint in a shared object that has
1433 not been loaded yet, and second of all there are problems with inlined functions
1434 in shared objects. Here are some suggestions to debugging your pass with
1437 <p>For sake of discussion, I'm going to assume that you are debugging a
1438 transformation invoked by <tt>opt</tt>, although nothing described here depends
1443 <!-- _______________________________________________________________________ -->
1444 <div class="doc_subsubsection">
1445 <a name="breakpoint">Setting a breakpoint in your pass</a>
1448 <div class="doc_text">
1450 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1455 Copyright 2000 Free Software Foundation, Inc.
1456 GDB is free software, covered by the GNU General Public License, and you are
1457 welcome to change it and/or distribute copies of it under certain conditions.
1458 Type "show copying" to see the conditions.
1459 There is absolutely no warranty for GDB. Type "show warranty" for details.
1460 This GDB was configured as "sparc-sun-solaris2.6"...
1464 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1465 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1466 (the shared object isn't loaded until runtime), we must execute the process, and
1467 have it stop before it invokes our pass, but after it has loaded the shared
1468 object. The most foolproof way of doing this is to set a breakpoint in
1469 <tt>PassManager::run</tt> and then run the process with the arguments you
1473 (gdb) <b>break PassManager::run</b>
1474 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1475 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/lib/Debug/[libname].so -[passoption]</b>
1476 Starting program: opt test.bc -load $(LLVMTOP)/llvm/lib/Debug/[libname].so -[passoption]
1477 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1478 70 bool PassManager::run(Module &M) { return PM->run(M); }
1482 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1483 now free to set breakpoints in your pass so that you can trace through execution
1484 or do other standard debugging stuff.</p>
1488 <!-- _______________________________________________________________________ -->
1489 <div class="doc_subsubsection">
1490 <a name="debugmisc">Miscellaneous Problems</a>
1493 <div class="doc_text">
1495 <p>Once you have the basics down, there are a couple of problems that GDB has,
1496 some with solutions, some without.</p>
1499 <li>Inline functions have bogus stack information. In general, GDB does a
1500 pretty good job getting stack traces and stepping through inline functions.
1501 When a pass is dynamically loaded however, it somehow completely loses this
1502 capability. The only solution I know of is to de-inline a function (move it
1503 from the body of a class to a .cpp file).</li>
1505 <li>Restarting the program breaks breakpoints. After following the information
1506 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1507 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1508 and you start getting errors about breakpoints being unsettable. The only way I
1509 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1510 already set in your pass, run the program, and re-set the breakpoints once
1511 execution stops in <tt>PassManager::run</tt>.</li>
1515 <p>Hopefully these tips will help with common case debugging situations. If
1516 you'd like to contribute some tips of your own, just contact <a
1517 href="mailto:sabre@nondot.org">Chris</a>.</p>
1521 <!-- *********************************************************************** -->
1522 <div class="doc_section">
1523 <a name="future">Future extensions planned</a>
1525 <!-- *********************************************************************** -->
1527 <div class="doc_text">
1529 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1530 some nifty stuff, there are things we'd like to add in the future. Here is
1531 where we are going:</p>
1535 <!-- _______________________________________________________________________ -->
1536 <div class="doc_subsubsection">
1537 <a name="SMP">Multithreaded LLVM</a>
1540 <div class="doc_text">
1542 <p>Multiple CPU machines are becoming more common and compilation can never be
1543 fast enough: obviously we should allow for a multithreaded compiler. Because of
1544 the semantics defined for passes above (specifically they cannot maintain state
1545 across invocations of their <tt>run*</tt> methods), a nice clean way to
1546 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1547 to create multiple instances of each pass object, and allow the separate
1548 instances to be hacking on different parts of the program at the same time.</p>
1550 <p>This implementation would prevent each of the passes from having to implement
1551 multithreaded constructs, requiring only the LLVM core to have locking in a few
1552 places (for global resources). Although this is a simple extension, we simply
1553 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1554 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1558 <!-- _______________________________________________________________________ -->
1559 <div class="doc_subsubsection">
1560 <a name="PassFunctionPass"><tt>ModulePass</tt>es requiring <tt>FunctionPass</tt>es</a>
1563 <div class="doc_text">
1565 <p>Currently it is illegal for a <a href="#ModulePass"><tt>ModulePass</tt></a>
1566 to require a <a href="#FunctionPass"><tt>FunctionPass</tt></a>. This is because
1567 there is only one instance of the <a
1568 href="#FunctionPass"><tt>FunctionPass</tt></a> object ever created, thus nowhere
1569 to store information for all of the functions in the program at the same time.
1570 Although this has come up a couple of times before, this has always been worked
1571 around by factoring one big complicated pass into a global and an
1572 interprocedural part, both of which are distinct. In the future, it would be
1573 nice to have this though.</p>
1575 <p>Note that it is no problem for a <a
1576 href="#FunctionPass"><tt>FunctionPass</tt></a> to require the results of a <a
1577 href="#ModulePass"><tt>ModulePass</tt></a>, only the other way around.</p>
1581 <!-- *********************************************************************** -->
1584 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1585 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1586 <a href="http://validator.w3.org/check/referer"><img
1587 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!" /></a>
1589 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1590 <a href="http://llvm.cs.uiuc.edu">The LLVM Compiler Infrastructure</a><br>
1591 Last modified: $Date$