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="#debughints">Using GDB with dynamically loaded passes</a>
88 <li><a href="#breakpoint">Setting a breakpoint in your pass</a></li>
89 <li><a href="#debugmisc">Miscellaneous Problems</a></li>
91 <li><a href="#future">Future extensions planned</a>
93 <li><a href="#SMP">Multithreaded LLVM</a></li>
94 <li><a href="#PassFunctionPass"><tt>ModulePass</tt>es requiring
95 <tt>FunctionPass</tt>es</a></li>
99 <div class="doc_author">
100 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
103 <!-- *********************************************************************** -->
104 <div class="doc_section">
105 <a name="introduction">Introduction - What is a pass?</a>
107 <!-- *********************************************************************** -->
109 <div class="doc_text">
111 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
112 passes are where most of the interesting parts of the compiler exist. Passes
113 perform the transformations and optimizations that make up the compiler, they
114 build the analysis results that are used by these transformations, and they are,
115 above all, a structuring technique for compiler code.</p>
117 <p>All LLVM passes are subclasses of the <tt><a
118 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
119 class, which implement functionality by overriding virtual methods inherited
120 from <tt>Pass</tt>. Depending on how your pass works, you should inherit from
121 the <tt><a href="#ModulePass">ModulePass</a></tt>, <tt><a
122 href="#CallGraphSCCPass">CallGraphSCCPass</a></tt>, <tt><a
123 href="#FunctionPass">FunctionPass</a></tt>, or <tt><a
124 href="#BasicBlockPass">BasicBlockPass</a></tt> classes, which gives the system
125 more information about what your pass does, and how it can be combined with
126 other passes. One of the main features of the LLVM Pass Framework is that it
127 schedules passes to run in an efficient way based on the constraints that your
128 pass meets (which are indicated by which class they derive from).</p>
130 <p>We start by showing you how to construct a pass, everything from setting up
131 the code, to compiling, loading, and executing it. After the basics are down,
132 more advanced features are discussed.</p>
136 <!-- *********************************************************************** -->
137 <div class="doc_section">
138 <a name="quickstart">Quick Start - Writing hello world</a>
140 <!-- *********************************************************************** -->
142 <div class="doc_text">
144 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
145 is designed to simply print out the name of non-external functions that exist in
146 the program being compiled. It does not modify the program at all, it just
147 inspects it. The source code and files for this pass are available in the LLVM
148 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
152 <!-- ======================================================================= -->
153 <div class="doc_subsection">
154 <a name="makefile">Setting up the build environment</a>
157 <div class="doc_text">
159 <p>First, you need to create a new directory somewhere in the LLVM source
160 base. For this example, we'll assume that you made
161 <tt>lib/Transforms/Hello</tt>. Next, you must set up a build script
162 (Makefile) that will compile the source code for the new pass. To do this,
163 copy the following into <tt>Makefile</tt>:</p>
167 # Makefile for hello pass
169 # Path to top level of LLVM heirarchy
172 # Name of the library to build
175 # Build a dynamically linkable shared object
178 # Make the shared library become a loadable module so the tools can
179 # dlopen/dlsym on the resulting library.
182 # Include the makefile implementation stuff
183 include $(LEVEL)/Makefile.common
186 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
187 directory are to be compiled and linked together into a
188 <tt>Debug/lib/Hello.so</tt> shared object that can be dynamically loaded by
189 the <tt>opt</tt> or <tt>analyze</tt> tools via their <tt>-load</tt> options.
190 If your operating system uses a suffix other than .so (such as windows or
191 Mac OS/X), the appropriate extension will be used.</p>
193 <p>Now that we have the build scripts set up, we just need to write the code for
198 <!-- ======================================================================= -->
199 <div class="doc_subsection">
200 <a name="basiccode">Basic code required</a>
203 <div class="doc_text">
205 <p>Now that we have a way to compile our new pass, we just have to write it.
209 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
210 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Function_8h-source.html">llvm/Function.h</a>"
213 <p>Which are needed because we are writing a <tt><a
214 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">Pass</a></tt>, and
215 we are operating on <tt><a
216 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Function.html">Function</a></tt>'s.</p>
220 <b>using namespace llvm;</b>
222 <p>... which is required because the functions from the include files
223 live in the llvm namespace.
232 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
233 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
234 things declared inside of the anonymous namespace only visible to the current
235 file. If you're not familiar with them, consult a decent C++ book for more
238 <p>Next, we declare our pass itself:</p>
241 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
244 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
245 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
246 The different builtin pass subclasses are described in detail <a
247 href="#passtype">later</a>, but for now, know that <a
248 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate a function at a
252 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
253 std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
256 }; <i>// end of struct Hello</i>
259 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
260 which overloads an abstract virtual method inherited from <a
261 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
262 to do our thing, so we just print out our message with the name of each
266 RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
267 } <i>// end of anonymous namespace</i>
270 <p>Lastly, we register our class <tt>Hello</tt>, giving it a command line
271 argument "<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>". There are
272 several different ways of <a href="#registration">registering your pass</a>,
273 depending on what it is to be used for. For "optimizations" we use the
274 <tt>RegisterOpt</tt> template.</p>
276 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
279 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
280 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Function_8h-source.html">llvm/Function.h</a>"
282 <b>using namespace llvm;</b>
285 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
286 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
287 std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
292 RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
296 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
297 command in the local directory and you should get a new
298 "<tt>Debug/lib/Hello.so</tt> file. Note that everything in this file is
299 contained in an anonymous namespace: this reflects the fact that passes are self
300 contained units that do not need external interfaces (although they can have
301 them) to be useful.</p>
305 <!-- ======================================================================= -->
306 <div class="doc_subsection">
307 <a name="running">Running a pass with <tt>opt</tt> or <tt>analyze</tt></a>
310 <div class="doc_text">
312 <p>Now that you have a brand new shiny shared object file, we can use the
313 <tt>opt</tt> command to run an LLVM program through your pass. Because you
314 registered your pass with the <tt>RegisterOpt</tt> template, you will be able to
315 use the <tt>opt</tt> tool to access it, once loaded.</p>
317 <p>To test it, follow the example at the end of the <a
318 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
319 LLVM. We can now run the bytecode file (<tt>hello.bc</tt>) for the program
320 through our transformation like this (or course, any bytecode file will
324 $ opt -load ../../../Debug/lib/Hello.so -hello < hello.bc > /dev/null
330 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
331 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
332 argument (which is one reason you need to <a href="#registration">register your
333 pass</a>). Because the hello pass does not modify the program in any
334 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
335 <tt>/dev/null</tt>).</p>
337 <p>To see what happened to the other string you registered, try running
338 <tt>opt</tt> with the <tt>--help</tt> option:</p>
341 $ opt -load ../../../Debug/lib/Hello.so --help
342 OVERVIEW: llvm .bc -> .bc modular optimizer
344 USAGE: opt [options] <input bytecode>
347 Optimizations available:
349 -funcresolve - Resolve Functions
350 -gcse - Global Common Subexpression Elimination
351 -globaldce - Dead Global Elimination
352 <b>-hello - Hello World Pass</b>
353 -indvars - Canonicalize Induction Variables
354 -inline - Function Integration/Inlining
355 -instcombine - Combine redundant instructions
359 <p>The pass name get added as the information string for your pass, giving some
360 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
361 would go ahead and make it do the cool transformations you want. Once you get
362 it all working and tested, it may become useful to find out how fast your pass
363 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
364 line option (<tt>--time-passes</tt>) that allows you to get information about
365 the execution time of your pass along with the other passes you queue up. For
369 $ opt -load ../../../Debug/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
373 ===============================================================================
374 ... Pass execution timing report ...
375 ===============================================================================
376 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
378 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
379 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bytecode Writer
380 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
381 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
382 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
383 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
386 <p>As you can see, our implementation above is pretty fast :). The additional
387 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
388 that the LLVM emitted by your pass is still valid and well formed LLVM, which
389 hasn't been broken somehow.</p>
391 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
392 about some more details of how they work and how to use them.</p>
396 <!-- *********************************************************************** -->
397 <div class="doc_section">
398 <a name="passtype">Pass classes and requirements</a>
400 <!-- *********************************************************************** -->
402 <div class="doc_text">
404 <p>One of the first things that you should do when designing a new pass is to
405 decide what class you should subclass for your pass. The <a
406 href="#basiccode">Hello World</a> example uses the <tt><a
407 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
408 did not discuss why or when this should occur. Here we talk about the classes
409 available, from the most general to the most specific.</p>
411 <p>When choosing a superclass for your Pass, you should choose the <b>most
412 specific</b> class possible, while still being able to meet the requirements
413 listed. This gives the LLVM Pass Infrastructure information necessary to
414 optimize how passes are run, so that the resultant compiler isn't unneccesarily
419 <!-- ======================================================================= -->
420 <div class="doc_subsection">
421 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
424 <div class="doc_text">
426 <p>The most plain and boring type of pass is the "<tt><a
427 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
428 class. This pass type is used for passes that do not have to be run, do not
429 change state, and never need to be updated. This is not a normal type of
430 transformation or analysis, but can provide information about the current
431 compiler configuration.</p>
433 <p>Although this pass class is very infrequently used, it is important for
434 providing information about the current target machine being compiled for, and
435 other static information that can affect the various transformations.</p>
437 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
438 invalidated, and are never "run".</p>
442 <!-- ======================================================================= -->
443 <div class="doc_subsection">
444 <a name="ModulePass">The <tt>ModulePass</tt> class</a>
447 <div class="doc_text">
450 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1ModulePass.html">ModulePass</a></tt>"
451 class is the most general of all superclasses that you can use. Deriving from
452 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
453 refering to function bodies in no predictable order, or adding and removing
454 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
455 subclasses, no optimization can be done for their execution.</p>
457 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
458 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
459 following signature:</p>
463 <!-- _______________________________________________________________________ -->
464 <div class="doc_subsubsection">
465 <a name="runOnModule">The <tt>runOnModule</tt> method</a>
468 <div class="doc_text">
471 <b>virtual bool</b> runOnModule(Module &M) = 0;
474 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
475 It should return true if the module was modified by the transformation and
480 <!-- ======================================================================= -->
481 <div class="doc_subsection">
482 <a name="CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
485 <div class="doc_text">
488 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1CallGraphSCCPass.html">CallGraphSCCPass</a></tt>"
489 is used by passes that need to traverse the program bottom-up on the call graph
490 (callees before callers). Deriving from CallGraphSCCPass provides some
491 mechanics for building and traversing the CallGraph, but also allows the system
492 to optimize execution of CallGraphSCCPass's. If your pass meets the
493 requirements outlined below, and doesn't meet the requirements of a <tt><a
494 href="#FunctionPass">FunctionPass</a></tt> or <tt><a
495 href="#BasicBlockPass">BasicBlockPass</a></tt>, you should derive from
496 <tt>CallGraphSCCPass</tt>.</p>
498 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
500 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
504 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
505 the current SCC.</li>
507 <li>... <em>allowed</em> to inspect any Function's other than those in the
508 current SCC and the direct callees of the SCC.</li>
510 <li>... <em>required</em> to preserve the current CallGraph object, updating it
511 to reflect any changes made to the program.</li>
513 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
514 though they may change the contents of an SCC.</li>
516 <li>... <em>allowed</em> to add or remove global variables from the current
519 <li>... <em>allowed</em> to maintain state across invocations of
520 <a href="#runOnSCC"><tt>runOnSCC</tt></a> (including global data).</li>
523 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
524 because it has to handle SCCs with more than one node in it. All of the virtual
525 methods described below should return true if they modified the program, or
526 false if they didn't.</p>
530 <!-- _______________________________________________________________________ -->
531 <div class="doc_subsubsection">
532 <a name="doInitialization_scc">The <tt>doInitialization(Module &)</tt>
536 <div class="doc_text">
539 <b>virtual bool</b> doInitialization(Module &M);
542 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
543 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
544 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
545 is designed to do simple initialization type of stuff that does not depend on
546 the SCCs being processed. The <tt>doInitialization</tt> method call is not
547 scheduled to overlap with any other pass executions (thus it should be very
552 <!-- _______________________________________________________________________ -->
553 <div class="doc_subsubsection">
554 <a name="runOnSCC">The <tt>runOnSCC</tt> method</a>
557 <div class="doc_text">
560 <b>virtual bool</b> runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;
563 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
564 should return true if the module was modified by the transformation, false
569 <!-- _______________________________________________________________________ -->
570 <div class="doc_subsubsection">
571 <a name="doFinalization_scc">The <tt>doFinalization(Module
572 &)</tt> method</a>
575 <div class="doc_text">
578 <b>virtual bool</b> doFinalization(Module &M);
581 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
582 called when the pass framework has finished calling <a
583 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
584 program being compiled.</p>
588 <!-- ======================================================================= -->
589 <div class="doc_subsection">
590 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
593 <div class="doc_text">
595 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
596 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
597 subclasses do have a predictable, local behavior that can be expected by the
598 system. All <tt>FunctionPass</tt> execute on each function in the program
599 independent of all of the other functions in the program.
600 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
601 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
603 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
606 <li>Modify a Function other than the one currently being processed.</li>
607 <li>Add or remove Function's from the current Module.</li>
608 <li>Add or remove global variables from the current Module.</li>
609 <li>Maintain state across invocations of
610 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
613 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
614 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
615 may overload three virtual methods to do their work. All of these methods
616 should return true if they modified the program, or false if they didn't.</p>
620 <!-- _______________________________________________________________________ -->
621 <div class="doc_subsubsection">
622 <a name="doInitialization_mod">The <tt>doInitialization(Module &)</tt>
626 <div class="doc_text">
629 <b>virtual bool</b> doInitialization(Module &M);
632 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
633 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
634 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
635 is designed to do simple initialization type of stuff that does not depend on
636 the functions being processed. The <tt>doInitialization</tt> method call is not
637 scheduled to overlap with any other pass executions (thus it should be very
640 <p>A good example of how this method should be used is the <a
641 href="http://llvm.cs.uiuc.edu/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
642 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
643 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
644 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
645 free functions that it needs, adding prototypes to the module if necessary.</p>
649 <!-- _______________________________________________________________________ -->
650 <div class="doc_subsubsection">
651 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
654 <div class="doc_text">
657 <b>virtual bool</b> runOnFunction(Function &F) = 0;
660 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
661 the transformation or analysis work of your pass. As usual, a true value should
662 be returned if the function is modified.</p>
666 <!-- _______________________________________________________________________ -->
667 <div class="doc_subsubsection">
668 <a name="doFinalization_mod">The <tt>doFinalization(Module
669 &)</tt> method</a>
672 <div class="doc_text">
675 <b>virtual bool</b> doFinalization(Module &M);
678 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
679 called when the pass framework has finished calling <a
680 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
681 program being compiled.</p>
685 <!-- ======================================================================= -->
686 <div class="doc_subsection">
687 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
690 <div class="doc_text">
692 <p><tt>BasicBlockPass</tt>'s are just like <a
693 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
694 their scope of inspection and modification to a single basic block at a time.
695 As such, they are <b>not</b> allowed to do any of the following:</p>
698 <li>Modify or inspect any basic blocks outside of the current one</li>
699 <li>Maintain state across invocations of
700 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
701 <li>Modify the control flow graph (by altering terminator instructions)</li>
702 <li>Any of the things forbidden for
703 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
706 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
707 optimizations. They may override the same <a
708 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
709 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
710 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
714 <!-- _______________________________________________________________________ -->
715 <div class="doc_subsubsection">
716 <a name="doInitialization_fn">The <tt>doInitialization(Function
717 &)</tt> method</a>
720 <div class="doc_text">
723 <b>virtual bool</b> doInitialization(Function &F);
726 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
727 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
728 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
729 to do simple initialization that does not depend on the
730 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
731 scheduled to overlap with any other pass executions (thus it should be very
736 <!-- _______________________________________________________________________ -->
737 <div class="doc_subsubsection">
738 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
741 <div class="doc_text">
744 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
747 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
748 function is not allowed to inspect or modify basic blocks other than the
749 parameter, and are not allowed to modify the CFG. A true value must be returned
750 if the basic block is modified.</p>
754 <!-- _______________________________________________________________________ -->
755 <div class="doc_subsubsection">
756 <a name="doFinalization_fn">The <tt>doFinalization(Function &)</tt>
760 <div class="doc_text">
763 <b>virtual bool</b> doFinalization(Function &F);
766 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
767 called when the pass framework has finished calling <a
768 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
769 program being compiled. This can be used to perform per-function
774 <!-- ======================================================================= -->
775 <div class="doc_subsection">
776 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
779 <div class="doc_text">
781 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
782 executes on the machine-dependent representation of each LLVM function in the
783 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
784 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
785 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
786 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
789 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
790 <li>Modify a MachineFunction other than the one currently being processed.</li>
791 <li>Add or remove MachineFunctions from the current Module.</li>
792 <li>Add or remove global variables from the current Module.</li>
793 <li>Maintain state across invocations of <a
794 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
800 <!-- _______________________________________________________________________ -->
801 <div class="doc_subsubsection">
802 <a name="runOnMachineFunction">The <tt>runOnMachineFunction(MachineFunction
803 &MF)</tt> method</a>
806 <div class="doc_text">
809 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
812 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
813 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
814 work of your <tt>MachineFunctionPass</tt>.</p>
816 <p>The <tt>runOnMachineFunction</tt> method is called on every
817 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
818 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
819 representation of the function. If you want to get at the LLVM <tt>Function</tt>
820 for the <tt>MachineFunction</tt> you're working on, use
821 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
822 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
823 <tt>MachineFunctionPass</tt>.</p>
827 <!-- *********************************************************************** -->
828 <div class="doc_section">
829 <a name="registration">Pass registration</a>
831 <!-- *********************************************************************** -->
833 <div class="doc_text">
835 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
836 pass registration works, and discussed some of the reasons that it is used and
837 what it does. Here we discuss how and why passes are registered.</p>
839 <p>Passes can be registered in several different ways. Depending on the general
840 classification of the pass, you should use one of the following templates to
841 register the pass:</p>
844 <li><b><tt>RegisterOpt</tt></b> - This template should be used when you are
845 registering a pass that logically should be available for use in the
846 '<tt>opt</tt>' utility.</li>
848 <li><b><tt>RegisterAnalysis</tt></b> - This template should be used when you are
849 registering a pass that logically should be available for use in the
850 '<tt>analyze</tt>' utility.</li>
852 <li><b><tt>RegisterPass</tt></b> - This is the generic form of the
853 <tt>Register*</tt> templates that should be used if you want your pass listed by
854 multiple or no utilities. This template takes an extra third argument that
855 specifies which tools it should be listed in. See the <a
856 href="http://llvm.cs.uiuc.edu/doxygen/PassSupport_8h-source.html">PassSupport.h</a>
857 file for more information.</li>
861 <p>Regardless of how you register your pass, you must specify at least two
862 parameters. The first parameter is the name of the pass that is to be used on
863 the command line to specify that the pass should be added to a program (for
864 example <tt>opt</tt> or <tt>analyze</tt>). The second argument is the name of
865 the pass, which is to be used for the <tt>--help</tt> output of programs, as
866 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
868 <p>If a pass is registered to be used by the <tt>analyze</tt> utility, you
869 should implement the virtual <tt>print</tt> method:</p>
873 <!-- _______________________________________________________________________ -->
874 <div class="doc_subsubsection">
875 <a name="print">The <tt>print</tt> method</a>
878 <div class="doc_text">
881 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
884 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
885 a human readable version of the analysis results. This is useful for debugging
886 an analysis itself, as well as for other people to figure out how an analysis
887 works. The <tt>analyze</tt> tool uses this method to generate its output.</p>
889 <p>The <tt>ostream</tt> parameter specifies the stream to write the results on,
890 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
891 program that has been analyzed. Note however that this pointer may be null in
892 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
893 debugger), so it should only be used to enhance debug output, it should not be
898 <!-- *********************************************************************** -->
899 <div class="doc_section">
900 <a name="interaction">Specifying interactions between passes</a>
902 <!-- *********************************************************************** -->
904 <div class="doc_text">
906 <p>One of the main responsibilities of the <tt>PassManager</tt> is the make sure
907 that passes interact with each other correctly. Because <tt>PassManager</tt>
908 tries to <a href="#passmanager">optimize the execution of passes</a> it must
909 know how the passes interact with each other and what dependencies exist between
910 the various passes. To track this, each pass can declare the set of passes that
911 are required to be executed before the current pass, and the passes which are
912 invalidated by the current pass.</p>
914 <p>Typically this functionality is used to require that analysis results are
915 computed before your pass is run. Running arbitrary transformation passes can
916 invalidate the computed analysis results, which is what the invalidation set
917 specifies. If a pass does not implement the <tt><a
918 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
919 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
923 <!-- _______________________________________________________________________ -->
924 <div class="doc_subsubsection">
925 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
928 <div class="doc_text">
931 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
934 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
935 invalidated sets may be specified for your transformation. The implementation
936 should fill in the <tt><a
937 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
938 object with information about which passes are required and not invalidated. To
939 do this, a pass may call any of the following methods on the AnalysisUsage
943 <!-- _______________________________________________________________________ -->
944 <div class="doc_subsubsection">
945 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
948 <div class="doc_text">
950 If your pass requires a previous pass to be executed (an analysis for example),
951 it can use one of these methods to arrange for it to be run before your pass.
952 LLVM has many different types of analyses and passes that can be required,
953 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
954 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
955 be no critical edges in the CFG when your pass has been run.
959 Some analyses chain to other analyses to do their job. For example, an <a
960 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
961 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
962 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
963 used instead of the <tt>addRequired</tt> method. This informs the PassManager
964 that the transitively required pass should be alive as long as the requiring
969 <!-- _______________________________________________________________________ -->
970 <div class="doc_subsubsection">
971 <a name="AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a>
974 <div class="doc_text">
976 One of the jobs of the PassManager is to optimize how and when analyses are run.
977 In particular, it attempts to avoid recomputing data unless it needs to. For
978 this reason, passes are allowed to declare that they preserve (i.e., they don't
979 invalidate) an existing analysis if it's available. For example, a simple
980 constant folding pass would not modify the CFG, so it can't possibly affect the
981 results of dominator analysis. By default, all passes are assumed to invalidate
986 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
987 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
988 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
989 not modify the LLVM program at all (which is true for analyses), and the
990 <tt>setPreservesCFG</tt> method can be used by transformations that change
991 instructions in the program but do not modify the CFG or terminator instructions
992 (note that this property is implicitly set for <a
993 href="#BasicBlockPass">BasicBlockPass</a>'s).
997 <tt>addPreserved</tt> is particularly useful for transformations like
998 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
999 and dominator related analyses if they exist, so it can preserve them, despite
1000 the fact that it hacks on the CFG.
1004 <!-- _______________________________________________________________________ -->
1005 <div class="doc_subsubsection">
1006 <a name="AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a>
1009 <div class="doc_text">
1012 <i>// This is an example implementation from an analysis, which does not modify
1013 // the program at all, yet has a prerequisite.</i>
1014 <b>void</b> <a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1015 AU.setPreservesAll();
1016 AU.addRequired<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1023 <i>// This example modifies the program, but does not modify the CFG</i>
1024 <b>void</b> <a href="http://llvm.cs.uiuc.edu/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1025 AU.setPreservesCFG();
1026 AU.addRequired<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1032 <!-- _______________________________________________________________________ -->
1033 <div class="doc_subsubsection">
1034 <a name="getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a>
1037 <div class="doc_text">
1039 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1040 your class, providing you with access to the passes that you declared that you
1041 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
1042 method. It takes a single template argument that specifies which pass class you
1043 want, and returns a reference to that pass. For example:</p>
1046 bool LICM::runOnFunction(Function &F) {
1047 LoopInfo &LI = getAnalysis<LoopInfo>();
1052 <p>This method call returns a reference to the pass desired. You may get a
1053 runtime assertion failure if you attempt to get an analysis that you did not
1054 declare as required in your <a
1055 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
1056 method can be called by your <tt>run*</tt> method implementation, or by any
1057 other local method invoked by your <tt>run*</tt> method.</p>
1060 If your pass is capable of updating analyses if they exist (e.g.,
1061 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1062 <tt>getAnalysisToUpdate</tt> method, which returns a pointer to the analysis if
1063 it is active. For example:</p>
1067 if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
1068 <i>// A DominatorSet is active. This code will update it.</i>
1075 <!-- *********************************************************************** -->
1076 <div class="doc_section">
1077 <a name="analysisgroup">Implementing Analysis Groups</a>
1079 <!-- *********************************************************************** -->
1081 <div class="doc_text">
1083 <p>Now that we understand the basics of how passes are defined, how the are
1084 used, and how they are required from other passes, it's time to get a little bit
1085 fancier. All of the pass relationships that we have seen so far are very
1086 simple: one pass depends on one other specific pass to be run before it can run.
1087 For many applications, this is great, for others, more flexibility is
1090 <p>In particular, some analyses are defined such that there is a single simple
1091 interface to the analysis results, but multiple ways of calculating them.
1092 Consider alias analysis for example. The most trivial alias analysis returns
1093 "may alias" for any alias query. The most sophisticated analysis a
1094 flow-sensitive, context-sensitive interprocedural analysis that can take a
1095 significant amount of time to execute (and obviously, there is a lot of room
1096 between these two extremes for other implementations). To cleanly support
1097 situations like this, the LLVM Pass Infrastructure supports the notion of
1098 Analysis Groups.</p>
1102 <!-- _______________________________________________________________________ -->
1103 <div class="doc_subsubsection">
1104 <a name="agconcepts">Analysis Group Concepts</a>
1107 <div class="doc_text">
1109 <p>An Analysis Group is a single simple interface that may be implemented by
1110 multiple different passes. Analysis Groups can be given human readable names
1111 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1112 class. An analysis group may have one or more implementations, one of which is
1113 the "default" implementation.</p>
1115 <p>Analysis groups are used by client passes just like other passes are: the
1116 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1117 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
1118 scans the available passes to see if any implementations of the analysis group
1119 are available. If none is available, the default implementation is created for
1120 the pass to use. All standard rules for <A href="#interaction">interaction
1121 between passes</a> still apply.</p>
1123 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
1124 passes, all analysis group implementations must be registered, and must use the
1125 <A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
1126 implementation pool. Also, a default implementation of the interface
1127 <b>must</b> be registered with <A
1128 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
1130 <p>As a concrete example of an Analysis Group in action, consider the <a
1131 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1132 analysis group. The default implementation of the alias analysis interface (the
1134 href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1135 pass) just does a few simple checks that don't require significant analysis to
1136 compute (such as: two different globals can never alias each other, etc).
1137 Passes that use the <tt><a
1138 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1139 interface (for example the <tt><a
1140 href="http://llvm.cs.uiuc.edu/doxygen/structGCSE.html">gcse</a></tt> pass), do
1141 not care which implementation of alias analysis is actually provided, they just
1142 use the designated interface.</p>
1144 <p>From the user's perspective, commands work just like normal. Issuing the
1145 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1146 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1147 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1148 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1149 hypothetical example) instead.</p>
1153 <!-- _______________________________________________________________________ -->
1154 <div class="doc_subsubsection">
1155 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1158 <div class="doc_text">
1160 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1161 group itself as well as add pass implementations to the analysis group. First,
1162 an analysis should be registered, with a human readable name provided for it.
1163 Unlike registration of passes, there is no command line argument to be specified
1164 for the Analysis Group Interface itself, because it is "abstract":</p>
1167 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1170 <p>Once the analysis is registered, passes can declare that they are valid
1171 implementations of the interface by using the following code:</p>
1175 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1176 RegisterOpt<FancyAA>
1177 B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
1179 //<i> Declare that we implement the AliasAnalysis interface</i>
1180 RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, FancyAA> C;
1184 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1185 uses the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
1186 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1187 analysis group. Every implementation of an analysis group should join using
1188 this template. A single pass may join multiple different analysis groups with
1193 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1194 RegisterOpt<<a href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1195 D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
1197 //<i> Declare that we implement the AliasAnalysis interface</i>
1198 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;
1202 <p>Here we show how the default implementation is specified (using the extra
1203 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1204 one default implementation available at all times for an Analysis Group to be
1205 used. Here we declare that the <tt><a
1206 href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1207 pass is the default implementation for the interface.</p>
1211 <!-- *********************************************************************** -->
1212 <div class="doc_section">
1213 <a name="passStatistics">Pass Statistics</a>
1215 <!-- *********************************************************************** -->
1217 <div class="doc_text">
1219 href="http://llvm.cs.uiuc.edu/doxygen/Statistic_8h-source.html"><tt>Statistic</tt></a>
1220 class, is designed to be an easy way to expose various success
1221 metrics from passes. These statistics are printed at the end of a
1222 run, when the -stats command line option is enabled on the command
1223 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1228 <!-- *********************************************************************** -->
1229 <div class="doc_section">
1230 <a name="passmanager">What PassManager does</a>
1232 <!-- *********************************************************************** -->
1234 <div class="doc_text">
1237 href="http://llvm.cs.uiuc.edu/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1239 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1PassManager.html">class</a>
1240 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1241 are set up correctly, and then schedules passes to run efficiently. All of the
1242 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1245 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1246 time of a series of passes:</p>
1249 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1250 recomputing analysis results as much as possible. This means keeping track of
1251 which analyses are available already, which analyses get invalidated, and which
1252 analyses are needed to be run for a pass. An important part of work is that the
1253 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1254 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1255 results as soon as they are no longer needed.</li>
1257 <li><b>Pipeline the execution of passes on the program</b> - The
1258 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1259 of a series of passes by pipelining the passes together. This means that, given
1260 a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1261 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1262 the first function, then all of the <a
1263 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1264 etc... until the entire program has been run through the passes.
1266 <p>This improves the cache behavior of the compiler, because it is only touching
1267 the LLVM program representation for a single function at a time, instead of
1268 traversing the entire program. It reduces the memory consumption of compiler,
1269 because, for example, only one <a
1270 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1271 needs to be calculated at a time. This also makes it possible some <a
1272 href="#SMP">interesting enhancements</a> in the future.</p></li>
1276 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1277 much information it has about the behaviors of the passes it is scheduling. For
1278 example, the "preserved" set is intentionally conservative in the face of an
1279 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1280 Not implementing when it should be implemented will have the effect of not
1281 allowing any analysis results to live across the execution of your pass.</p>
1283 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1284 options that is useful for debugging pass execution, seeing how things work, and
1285 diagnosing when you should be preserving more analyses than you currently are
1286 (To get information about all of the variants of the <tt>--debug-pass</tt>
1287 option, just type '<tt>opt --help-hidden</tt>').</p>
1289 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1290 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1291 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1294 $ opt -load ../../../Debug/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1296 Function Pass Manager
1297 Dominator Set Construction
1298 Immediate Dominators Construction
1299 Global Common Subexpression Elimination
1300 -- Immediate Dominators Construction
1301 -- Global Common Subexpression Elimination
1302 Natural Loop Construction
1303 Loop Invariant Code Motion
1304 -- Natural Loop Construction
1305 -- Loop Invariant Code Motion
1307 -- Dominator Set Construction
1313 <p>This output shows us when passes are constructed and when the analysis
1314 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1315 GCSE uses dominator and immediate dominator information to do its job. The LICM
1316 pass uses natural loop information, which uses dominator sets, but not immediate
1317 dominators. Because immediate dominators are no longer useful after the GCSE
1318 pass, it is immediately destroyed. The dominator sets are then reused to
1319 compute natural loop information, which is then used by the LICM pass.</p>
1321 <p>After the LICM pass, the module verifier runs (which is automatically added
1322 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1323 resultant LLVM code is well formed. After it finishes, the dominator set
1324 information is destroyed, after being computed once, and shared by three
1327 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1328 World</a> pass in between the two passes:</p>
1331 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1333 Function Pass Manager
1334 Dominator Set Construction
1335 Immediate Dominators Construction
1336 Global Common Subexpression Elimination
1337 <b>-- Dominator Set Construction</b>
1338 -- Immediate Dominators Construction
1339 -- Global Common Subexpression Elimination
1340 <b> Hello World Pass
1342 Dominator Set Construction</b>
1343 Natural Loop Construction
1344 Loop Invariant Code Motion
1345 -- Natural Loop Construction
1346 -- Loop Invariant Code Motion
1348 -- Dominator Set Construction
1357 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1358 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1359 we need to add the following <a
1360 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1363 <i>// We don't modify the program, so we preserve all analyses</i>
1364 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1365 AU.setPreservesAll();
1369 <p>Now when we run our pass, we get this output:</p>
1372 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1373 Pass Arguments: -gcse -hello -licm
1375 Function Pass Manager
1376 Dominator Set Construction
1377 Immediate Dominators Construction
1378 Global Common Subexpression Elimination
1379 -- Immediate Dominators Construction
1380 -- Global Common Subexpression Elimination
1383 Natural Loop Construction
1384 Loop Invariant Code Motion
1385 -- Loop Invariant Code Motion
1386 -- Natural Loop Construction
1388 -- Dominator Set Construction
1397 <p>Which shows that we don't accidentally invalidate dominator information
1398 anymore, and therefore do not have to compute it twice.</p>
1402 <!-- _______________________________________________________________________ -->
1403 <div class="doc_subsubsection">
1404 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1407 <div class="doc_text">
1410 <b>virtual void</b> releaseMemory();
1413 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1414 results, and how long to keep them around for. Because the lifetime of the pass
1415 object itself is effectively the entire duration of the compilation process, we
1416 need some way to free analysis results when they are no longer useful. The
1417 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1419 <p>If you are writing an analysis or any other pass that retains a significant
1420 amount of state (for use by another pass which "requires" your pass and uses the
1421 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1422 <tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
1423 internal state. This method is called after the <tt>run*</tt> method for the
1424 class, before the next call of <tt>run*</tt> in your pass.</p>
1428 <!-- *********************************************************************** -->
1429 <div class="doc_section">
1430 <a name="debughints">Using GDB with dynamically loaded passes</a>
1432 <!-- *********************************************************************** -->
1434 <div class="doc_text">
1436 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1437 should be. First of all, you can't set a breakpoint in a shared object that has
1438 not been loaded yet, and second of all there are problems with inlined functions
1439 in shared objects. Here are some suggestions to debugging your pass with
1442 <p>For sake of discussion, I'm going to assume that you are debugging a
1443 transformation invoked by <tt>opt</tt>, although nothing described here depends
1448 <!-- _______________________________________________________________________ -->
1449 <div class="doc_subsubsection">
1450 <a name="breakpoint">Setting a breakpoint in your pass</a>
1453 <div class="doc_text">
1455 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1460 Copyright 2000 Free Software Foundation, Inc.
1461 GDB is free software, covered by the GNU General Public License, and you are
1462 welcome to change it and/or distribute copies of it under certain conditions.
1463 Type "show copying" to see the conditions.
1464 There is absolutely no warranty for GDB. Type "show warranty" for details.
1465 This GDB was configured as "sparc-sun-solaris2.6"...
1469 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1470 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1471 (the shared object isn't loaded until runtime), we must execute the process, and
1472 have it stop before it invokes our pass, but after it has loaded the shared
1473 object. The most foolproof way of doing this is to set a breakpoint in
1474 <tt>PassManager::run</tt> and then run the process with the arguments you
1478 (gdb) <b>break PassManager::run</b>
1479 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1480 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]</b>
1481 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
1482 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1483 70 bool PassManager::run(Module &M) { return PM->run(M); }
1487 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1488 now free to set breakpoints in your pass so that you can trace through execution
1489 or do other standard debugging stuff.</p>
1493 <!-- _______________________________________________________________________ -->
1494 <div class="doc_subsubsection">
1495 <a name="debugmisc">Miscellaneous Problems</a>
1498 <div class="doc_text">
1500 <p>Once you have the basics down, there are a couple of problems that GDB has,
1501 some with solutions, some without.</p>
1504 <li>Inline functions have bogus stack information. In general, GDB does a
1505 pretty good job getting stack traces and stepping through inline functions.
1506 When a pass is dynamically loaded however, it somehow completely loses this
1507 capability. The only solution I know of is to de-inline a function (move it
1508 from the body of a class to a .cpp file).</li>
1510 <li>Restarting the program breaks breakpoints. After following the information
1511 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1512 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1513 and you start getting errors about breakpoints being unsettable. The only way I
1514 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1515 already set in your pass, run the program, and re-set the breakpoints once
1516 execution stops in <tt>PassManager::run</tt>.</li>
1520 <p>Hopefully these tips will help with common case debugging situations. If
1521 you'd like to contribute some tips of your own, just contact <a
1522 href="mailto:sabre@nondot.org">Chris</a>.</p>
1526 <!-- *********************************************************************** -->
1527 <div class="doc_section">
1528 <a name="future">Future extensions planned</a>
1530 <!-- *********************************************************************** -->
1532 <div class="doc_text">
1534 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1535 some nifty stuff, there are things we'd like to add in the future. Here is
1536 where we are going:</p>
1540 <!-- _______________________________________________________________________ -->
1541 <div class="doc_subsubsection">
1542 <a name="SMP">Multithreaded LLVM</a>
1545 <div class="doc_text">
1547 <p>Multiple CPU machines are becoming more common and compilation can never be
1548 fast enough: obviously we should allow for a multithreaded compiler. Because of
1549 the semantics defined for passes above (specifically they cannot maintain state
1550 across invocations of their <tt>run*</tt> methods), a nice clean way to
1551 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1552 to create multiple instances of each pass object, and allow the separate
1553 instances to be hacking on different parts of the program at the same time.</p>
1555 <p>This implementation would prevent each of the passes from having to implement
1556 multithreaded constructs, requiring only the LLVM core to have locking in a few
1557 places (for global resources). Although this is a simple extension, we simply
1558 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1559 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1563 <!-- _______________________________________________________________________ -->
1564 <div class="doc_subsubsection">
1565 <a name="PassFunctionPass"><tt>ModulePass</tt>es requiring <tt>FunctionPass</tt>es</a>
1568 <div class="doc_text">
1570 <p>Currently it is illegal for a <a href="#ModulePass"><tt>ModulePass</tt></a>
1571 to require a <a href="#FunctionPass"><tt>FunctionPass</tt></a>. This is because
1572 there is only one instance of the <a
1573 href="#FunctionPass"><tt>FunctionPass</tt></a> object ever created, thus nowhere
1574 to store information for all of the functions in the program at the same time.
1575 Although this has come up a couple of times before, this has always been worked
1576 around by factoring one big complicated pass into a global and an
1577 interprocedural part, both of which are distinct. In the future, it would be
1578 nice to have this though.</p>
1580 <p>Note that it is no problem for a <a
1581 href="#FunctionPass"><tt>FunctionPass</tt></a> to require the results of a <a
1582 href="#ModulePass"><tt>ModulePass</tt></a>, only the other way around.</p>
1586 <!-- *********************************************************************** -->
1589 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1590 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1591 <a href="http://validator.w3.org/check/referer"><img
1592 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!" /></a>
1594 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1595 <a href="http://llvm.cs.uiuc.edu">The LLVM Compiler Infrastructure</a><br>
1596 Last modified: $Date$