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></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(CallGraph
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(CallGraph
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="#LoopPass">The <tt>LoopPass</tt> class</a>
48 <li><a href="#doInitialization_loop">The <tt>doInitialization(Loop *,
49 LPPassManager &)</tt> method</a></li>
50 <li><a href="#runOnLoop">The <tt>runOnLoop</tt> method</a></li>
51 <li><a href="#doFinalization_loop">The <tt>doFinalization()
54 <li><a href="#BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
56 <li><a href="#doInitialization_fn">The <tt>doInitialization(Function
57 &)</tt> method</a></li>
58 <li><a href="#runOnBasicBlock">The <tt>runOnBasicBlock</tt>
60 <li><a href="#doFinalization_fn">The <tt>doFinalization(Function
61 &)</tt> method</a></li>
63 <li><a href="#MachineFunctionPass">The <tt>MachineFunctionPass</tt>
66 <li><a href="#runOnMachineFunction">The
67 <tt>runOnMachineFunction(MachineFunction &)</tt> method</a></li>
70 <li><a href="#registration">Pass Registration</a>
72 <li><a href="#print">The <tt>print</tt> method</a></li>
74 <li><a href="#interaction">Specifying interactions between passes</a>
76 <li><a href="#getAnalysisUsage">The <tt>getAnalysisUsage</tt>
78 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
79 <li><a href="#AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a></li>
80 <li><a href="#AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a></li>
81 <li><a href="#getAnalysis">The <tt>getAnalysis<></tt> and
82 <tt>getAnalysisIfAvailable<></tt> methods</a></li>
84 <li><a href="#analysisgroup">Implementing Analysis Groups</a>
86 <li><a href="#agconcepts">Analysis Group Concepts</a></li>
87 <li><a href="#registerag">Using <tt>RegisterAnalysisGroup</tt></a></li>
89 <li><a href="#passStatistics">Pass Statistics</a>
90 <li><a href="#passmanager">What PassManager does</a>
92 <li><a href="#releaseMemory">The <tt>releaseMemory</tt> method</a></li>
94 <li><a href="#registering">Registering dynamically loaded passes</a>
96 <li><a href="#registering_existing">Using existing registries</a></li>
97 <li><a href="#registering_new">Creating new registries</a></li>
99 <li><a href="#debughints">Using GDB with dynamically loaded passes</a>
101 <li><a href="#breakpoint">Setting a breakpoint in your pass</a></li>
102 <li><a href="#debugmisc">Miscellaneous Problems</a></li>
104 <li><a href="#future">Future extensions planned</a>
106 <li><a href="#SMP">Multithreaded LLVM</a></li>
110 <div class="doc_author">
111 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
112 <a href="mailto:jlaskey@mac.com">Jim Laskey</a></p>
115 <!-- *********************************************************************** -->
116 <div class="doc_section">
117 <a name="introduction">Introduction - What is a pass?</a>
119 <!-- *********************************************************************** -->
121 <div class="doc_text">
123 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
124 passes are where most of the interesting parts of the compiler exist. Passes
125 perform the transformations and optimizations that make up the compiler, they
126 build the analysis results that are used by these transformations, and they are,
127 above all, a structuring technique for compiler code.</p>
129 <p>All LLVM passes are subclasses of the <tt><a
130 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
131 class, which implement functionality by overriding virtual methods inherited
132 from <tt>Pass</tt>. Depending on how your pass works, you should inherit from
133 the <tt><a href="#ModulePass">ModulePass</a></tt>, <tt><a
134 href="#CallGraphSCCPass">CallGraphSCCPass</a></tt>, <tt><a
135 href="#FunctionPass">FunctionPass</a></tt>, or <tt><a
136 href="#LoopPass">LoopPass</a></tt>, or <tt><a
137 href="#BasicBlockPass">BasicBlockPass</a></tt> classes, which gives the system
138 more information about what your pass does, and how it can be combined with
139 other passes. One of the main features of the LLVM Pass Framework is that it
140 schedules passes to run in an efficient way based on the constraints that your
141 pass meets (which are indicated by which class they derive from).</p>
143 <p>We start by showing you how to construct a pass, everything from setting up
144 the code, to compiling, loading, and executing it. After the basics are down,
145 more advanced features are discussed.</p>
149 <!-- *********************************************************************** -->
150 <div class="doc_section">
151 <a name="quickstart">Quick Start - Writing hello world</a>
153 <!-- *********************************************************************** -->
155 <div class="doc_text">
157 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
158 is designed to simply print out the name of non-external functions that exist in
159 the program being compiled. It does not modify the program at all, it just
160 inspects it. The source code and files for this pass are available in the LLVM
161 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
165 <!-- ======================================================================= -->
166 <div class="doc_subsection">
167 <a name="makefile">Setting up the build environment</a>
170 <div class="doc_text">
172 <p>First, you need to create a new directory somewhere in the LLVM source
173 base. For this example, we'll assume that you made
174 <tt>lib/Transforms/Hello</tt>. Next, you must set up a build script
175 (Makefile) that will compile the source code for the new pass. To do this,
176 copy the following into <tt>Makefile</tt>:</p>
179 <div class="doc_code"><pre>
180 # Makefile for hello pass
182 # Path to top level of LLVM hierarchy
185 # Name of the library to build
188 # Make the shared library become a loadable module so the tools can
189 # dlopen/dlsym on the resulting library.
192 # Include the makefile implementation stuff
193 include $(LEVEL)/Makefile.common
196 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
197 directory are to be compiled and linked together into a
198 <tt>Debug+Asserts/lib/Hello.so</tt> shared object that can be dynamically loaded by
199 the <tt>opt</tt> or <tt>bugpoint</tt> tools via their <tt>-load</tt> options.
200 If your operating system uses a suffix other than .so (such as windows or
201 Mac OS/X), the appropriate extension will be used.</p>
203 <p>Now that we have the build scripts set up, we just need to write the code for
208 <!-- ======================================================================= -->
209 <div class="doc_subsection">
210 <a name="basiccode">Basic code required</a>
213 <div class="doc_text">
215 <p>Now that we have a way to compile our new pass, we just have to write it.
218 <div class="doc_code"><pre>
219 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
220 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
221 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
224 <p>Which are needed because we are writing a <tt><a
225 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>,
226 we are operating on <tt><a
227 href="http://llvm.org/doxygen/classllvm_1_1Function.html">Function</a></tt>'s,
228 and we will be doing some printing.</p>
231 <div class="doc_code"><pre>
232 <b>using namespace llvm;</b>
234 <p>... which is required because the functions from the include files
235 live in the llvm namespace.
240 <div class="doc_code"><pre>
244 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
245 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
246 things declared inside of the anonymous namespace only visible to the current
247 file. If you're not familiar with them, consult a decent C++ book for more
250 <p>Next, we declare our pass itself:</p>
252 <div class="doc_code"><pre>
253 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
256 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
257 href="http://llvm.org/doxygen/classllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
258 The different builtin pass subclasses are described in detail <a
259 href="#passtype">later</a>, but for now, know that <a
260 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate a function at a
263 <div class="doc_code"><pre>
265 Hello() : FunctionPass(&ID) {}
268 <p> This declares pass identifier used by LLVM to identify pass. This allows LLVM to
269 avoid using expensive C++ runtime information.</p>
271 <div class="doc_code"><pre>
272 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
273 errs() << "<i>Hello: </i>" << F.getName() << "\n";
276 }; <i>// end of struct Hello</i>
279 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
280 which overloads an abstract virtual method inherited from <a
281 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
282 to do our thing, so we just print out our message with the name of each
285 <div class="doc_code"><pre>
289 <p> We initialize pass ID here. LLVM uses ID's address to identify pass so
290 initialization value is not important.</p>
292 <div class="doc_code"><pre>
293 RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>",
294 false /* Only looks at CFG */,
295 false /* Analysis Pass */);
296 } <i>// end of anonymous namespace</i>
299 <p>Lastly, we <a href="#registration">register our class</a> <tt>Hello</tt>,
300 giving it a command line
301 argument "<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>".
302 Last two RegisterPass arguments are optional. Their default value is false.
303 If a pass walks CFG without modifying it then third argument is set to true.
304 If a pass is an analysis pass, for example dominator tree pass, then true
305 is supplied as fourth argument. </p>
307 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
309 <div class="doc_code"><pre>
310 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
311 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
312 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
314 <b>using namespace llvm;</b>
317 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
320 Hello() : FunctionPass(&ID) {}
322 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
323 errs() << "<i>Hello: </i>" << F.getName() << "\n";
329 RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
333 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
334 command in the local directory and you should get a new
335 "<tt>Debug+Asserts/lib/Hello.so</tt> file. Note that everything in this file is
336 contained in an anonymous namespace: this reflects the fact that passes are self
337 contained units that do not need external interfaces (although they can have
338 them) to be useful.</p>
342 <!-- ======================================================================= -->
343 <div class="doc_subsection">
344 <a name="running">Running a pass with <tt>opt</tt></a>
347 <div class="doc_text">
349 <p>Now that you have a brand new shiny shared object file, we can use the
350 <tt>opt</tt> command to run an LLVM program through your pass. Because you
351 registered your pass with the <tt>RegisterPass</tt> template, you will be able to
352 use the <tt>opt</tt> tool to access it, once loaded.</p>
354 <p>To test it, follow the example at the end of the <a
355 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
356 LLVM. We can now run the bitcode file (<tt>hello.bc</tt>) for the program
357 through our transformation like this (or course, any bitcode file will
360 <div class="doc_code"><pre>
361 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello < hello.bc > /dev/null
367 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
368 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
369 argument (which is one reason you need to <a href="#registration">register your
370 pass</a>). Because the hello pass does not modify the program in any
371 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
372 <tt>/dev/null</tt>).</p>
374 <p>To see what happened to the other string you registered, try running
375 <tt>opt</tt> with the <tt>-help</tt> option:</p>
377 <div class="doc_code"><pre>
378 $ opt -load ../../../Debug+Asserts/lib/Hello.so -help
379 OVERVIEW: llvm .bc -> .bc modular optimizer
381 USAGE: opt [options] <input bitcode>
384 Optimizations available:
386 -funcresolve - Resolve Functions
387 -gcse - Global Common Subexpression Elimination
388 -globaldce - Dead Global Elimination
389 <b>-hello - Hello World Pass</b>
390 -indvars - Canonicalize Induction Variables
391 -inline - Function Integration/Inlining
392 -instcombine - Combine redundant instructions
396 <p>The pass name get added as the information string for your pass, giving some
397 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
398 would go ahead and make it do the cool transformations you want. Once you get
399 it all working and tested, it may become useful to find out how fast your pass
400 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
401 line option (<tt>--time-passes</tt>) that allows you to get information about
402 the execution time of your pass along with the other passes you queue up. For
405 <div class="doc_code"><pre>
406 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
410 ===============================================================================
411 ... Pass execution timing report ...
412 ===============================================================================
413 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
415 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
416 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bitcode Writer
417 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
418 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
419 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
420 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
423 <p>As you can see, our implementation above is pretty fast :). The additional
424 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
425 that the LLVM emitted by your pass is still valid and well formed LLVM, which
426 hasn't been broken somehow.</p>
428 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
429 about some more details of how they work and how to use them.</p>
433 <!-- *********************************************************************** -->
434 <div class="doc_section">
435 <a name="passtype">Pass classes and requirements</a>
437 <!-- *********************************************************************** -->
439 <div class="doc_text">
441 <p>One of the first things that you should do when designing a new pass is to
442 decide what class you should subclass for your pass. The <a
443 href="#basiccode">Hello World</a> example uses the <tt><a
444 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
445 did not discuss why or when this should occur. Here we talk about the classes
446 available, from the most general to the most specific.</p>
448 <p>When choosing a superclass for your Pass, you should choose the <b>most
449 specific</b> class possible, while still being able to meet the requirements
450 listed. This gives the LLVM Pass Infrastructure information necessary to
451 optimize how passes are run, so that the resultant compiler isn't unnecessarily
456 <!-- ======================================================================= -->
457 <div class="doc_subsection">
458 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
461 <div class="doc_text">
463 <p>The most plain and boring type of pass is the "<tt><a
464 href="http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
465 class. This pass type is used for passes that do not have to be run, do not
466 change state, and never need to be updated. This is not a normal type of
467 transformation or analysis, but can provide information about the current
468 compiler configuration.</p>
470 <p>Although this pass class is very infrequently used, it is important for
471 providing information about the current target machine being compiled for, and
472 other static information that can affect the various transformations.</p>
474 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
475 invalidated, and are never "run".</p>
479 <!-- ======================================================================= -->
480 <div class="doc_subsection">
481 <a name="ModulePass">The <tt>ModulePass</tt> class</a>
484 <div class="doc_text">
487 href="http://llvm.org/doxygen/classllvm_1_1ModulePass.html">ModulePass</a></tt>"
488 class is the most general of all superclasses that you can use. Deriving from
489 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
490 referring to function bodies in no predictable order, or adding and removing
491 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
492 subclasses, no optimization can be done for their execution.</p>
494 <p>A module pass can use function level passes (e.g. dominators) using
495 the getAnalysis interface
496 <tt>getAnalysis<DominatorTree>(llvm::Function *)</tt> to provide the
497 function to retrieve analysis result for, if the function pass does not require
498 any module or immutable passes. Note that this can only be done for functions for which the
499 analysis ran, e.g. in the case of dominators you should only ask for the
500 DominatorTree for function definitions, not declarations.</p>
502 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
503 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
504 following signature:</p>
508 <!-- _______________________________________________________________________ -->
509 <div class="doc_subsubsection">
510 <a name="runOnModule">The <tt>runOnModule</tt> method</a>
513 <div class="doc_text">
515 <div class="doc_code"><pre>
516 <b>virtual bool</b> runOnModule(Module &M) = 0;
519 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
520 It should return true if the module was modified by the transformation and
525 <!-- ======================================================================= -->
526 <div class="doc_subsection">
527 <a name="CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
530 <div class="doc_text">
533 href="http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html">CallGraphSCCPass</a></tt>"
534 is used by passes that need to traverse the program bottom-up on the call graph
535 (callees before callers). Deriving from CallGraphSCCPass provides some
536 mechanics for building and traversing the CallGraph, but also allows the system
537 to optimize execution of CallGraphSCCPass's. If your pass meets the
538 requirements outlined below, and doesn't meet the requirements of a <tt><a
539 href="#FunctionPass">FunctionPass</a></tt> or <tt><a
540 href="#BasicBlockPass">BasicBlockPass</a></tt>, you should derive from
541 <tt>CallGraphSCCPass</tt>.</p>
543 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
545 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
549 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
550 the current SCC.</li>
552 <li>... <em>not allowed</em> to inspect any Function's other than those in the
553 current SCC and the direct callees of the SCC.</li>
555 <li>... <em>required</em> to preserve the current CallGraph object, updating it
556 to reflect any changes made to the program.</li>
558 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
559 though they may change the contents of an SCC.</li>
561 <li>... <em>allowed</em> to add or remove global variables from the current
564 <li>... <em>allowed</em> to maintain state across invocations of
565 <a href="#runOnSCC"><tt>runOnSCC</tt></a> (including global data).</li>
568 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
569 because it has to handle SCCs with more than one node in it. All of the virtual
570 methods described below should return true if they modified the program, or
571 false if they didn't.</p>
575 <!-- _______________________________________________________________________ -->
576 <div class="doc_subsubsection">
577 <a name="doInitialization_scc">The <tt>doInitialization(CallGraph &)</tt>
581 <div class="doc_text">
583 <div class="doc_code"><pre>
584 <b>virtual bool</b> doInitialization(CallGraph &CG);
587 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
588 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
589 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
590 is designed to do simple initialization type of stuff that does not depend on
591 the SCCs being processed. The <tt>doInitialization</tt> method call is not
592 scheduled to overlap with any other pass executions (thus it should be very
597 <!-- _______________________________________________________________________ -->
598 <div class="doc_subsubsection">
599 <a name="runOnSCC">The <tt>runOnSCC</tt> method</a>
602 <div class="doc_text">
604 <div class="doc_code"><pre>
605 <b>virtual bool</b> runOnSCC(CallGraphSCC &SCC) = 0;
608 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
609 should return true if the module was modified by the transformation, false
614 <!-- _______________________________________________________________________ -->
615 <div class="doc_subsubsection">
616 <a name="doFinalization_scc">The <tt>doFinalization(CallGraph
617 &)</tt> method</a>
620 <div class="doc_text">
622 <div class="doc_code"><pre>
623 <b>virtual bool</b> doFinalization(CallGraph &CG);
626 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
627 called when the pass framework has finished calling <a
628 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
629 program being compiled.</p>
633 <!-- ======================================================================= -->
634 <div class="doc_subsection">
635 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
638 <div class="doc_text">
640 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
641 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
642 subclasses do have a predictable, local behavior that can be expected by the
643 system. All <tt>FunctionPass</tt> execute on each function in the program
644 independent of all of the other functions in the program.
645 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
646 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
648 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
651 <li>Modify a Function other than the one currently being processed.</li>
652 <li>Add or remove Function's from the current Module.</li>
653 <li>Add or remove global variables from the current Module.</li>
654 <li>Maintain state across invocations of
655 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
658 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
659 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
660 may overload three virtual methods to do their work. All of these methods
661 should return true if they modified the program, or false if they didn't.</p>
665 <!-- _______________________________________________________________________ -->
666 <div class="doc_subsubsection">
667 <a name="doInitialization_mod">The <tt>doInitialization(Module &)</tt>
671 <div class="doc_text">
673 <div class="doc_code"><pre>
674 <b>virtual bool</b> doInitialization(Module &M);
677 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
678 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
679 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
680 is designed to do simple initialization type of stuff that does not depend on
681 the functions being processed. The <tt>doInitialization</tt> method call is not
682 scheduled to overlap with any other pass executions (thus it should be very
685 <p>A good example of how this method should be used is the <a
686 href="http://llvm.org/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
687 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
688 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
689 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
690 free functions that it needs, adding prototypes to the module if necessary.</p>
694 <!-- _______________________________________________________________________ -->
695 <div class="doc_subsubsection">
696 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
699 <div class="doc_text">
701 <div class="doc_code"><pre>
702 <b>virtual bool</b> runOnFunction(Function &F) = 0;
705 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
706 the transformation or analysis work of your pass. As usual, a true value should
707 be returned if the function is modified.</p>
711 <!-- _______________________________________________________________________ -->
712 <div class="doc_subsubsection">
713 <a name="doFinalization_mod">The <tt>doFinalization(Module
714 &)</tt> method</a>
717 <div class="doc_text">
719 <div class="doc_code"><pre>
720 <b>virtual bool</b> doFinalization(Module &M);
723 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
724 called when the pass framework has finished calling <a
725 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
726 program being compiled.</p>
730 <!-- ======================================================================= -->
731 <div class="doc_subsection">
732 <a name="LoopPass">The <tt>LoopPass</tt> class </a>
735 <div class="doc_text">
737 <p> All <tt>LoopPass</tt> execute on each loop in the function independent of
738 all of the other loops in the function. <tt>LoopPass</tt> processes loops in
739 loop nest order such that outer most loop is processed last. </p>
741 <p> <tt>LoopPass</tt> subclasses are allowed to update loop nest using
742 <tt>LPPassManager</tt> interface. Implementing a loop pass is usually
743 straightforward. <tt>Looppass</tt>'s may overload three virtual methods to
744 do their work. All these methods should return true if they modified the
745 program, or false if they didn't. </p>
748 <!-- _______________________________________________________________________ -->
749 <div class="doc_subsubsection">
750 <a name="doInitialization_loop">The <tt>doInitialization(Loop *,
751 LPPassManager &)</tt>
755 <div class="doc_text">
757 <div class="doc_code"><pre>
758 <b>virtual bool</b> doInitialization(Loop *, LPPassManager &LPM);
761 <p>The <tt>doInitialization</tt> method is designed to do simple initialization
762 type of stuff that does not depend on the functions being processed. The
763 <tt>doInitialization</tt> method call is not scheduled to overlap with any
764 other pass executions (thus it should be very fast). LPPassManager
765 interface should be used to access Function or Module level analysis
771 <!-- _______________________________________________________________________ -->
772 <div class="doc_subsubsection">
773 <a name="runOnLoop">The <tt>runOnLoop</tt> method</a>
776 <div class="doc_text">
778 <div class="doc_code"><pre>
779 <b>virtual bool</b> runOnLoop(Loop *, LPPassManager &LPM) = 0;
782 <p>The <tt>runOnLoop</tt> method must be implemented by your subclass to do
783 the transformation or analysis work of your pass. As usual, a true value should
784 be returned if the function is modified. <tt>LPPassManager</tt> interface
785 should be used to update loop nest.</p>
789 <!-- _______________________________________________________________________ -->
790 <div class="doc_subsubsection">
791 <a name="doFinalization_loop">The <tt>doFinalization()</tt> method</a>
794 <div class="doc_text">
796 <div class="doc_code"><pre>
797 <b>virtual bool</b> doFinalization();
800 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
801 called when the pass framework has finished calling <a
802 href="#runOnLoop"><tt>runOnLoop</tt></a> for every loop in the
803 program being compiled. </p>
809 <!-- ======================================================================= -->
810 <div class="doc_subsection">
811 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
814 <div class="doc_text">
816 <p><tt>BasicBlockPass</tt>'s are just like <a
817 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
818 their scope of inspection and modification to a single basic block at a time.
819 As such, they are <b>not</b> allowed to do any of the following:</p>
822 <li>Modify or inspect any basic blocks outside of the current one</li>
823 <li>Maintain state across invocations of
824 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
825 <li>Modify the control flow graph (by altering terminator instructions)</li>
826 <li>Any of the things forbidden for
827 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
830 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
831 optimizations. They may override the same <a
832 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
833 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
834 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
838 <!-- _______________________________________________________________________ -->
839 <div class="doc_subsubsection">
840 <a name="doInitialization_fn">The <tt>doInitialization(Function
841 &)</tt> method</a>
844 <div class="doc_text">
846 <div class="doc_code"><pre>
847 <b>virtual bool</b> doInitialization(Function &F);
850 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
851 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
852 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
853 to do simple initialization that does not depend on the
854 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
855 scheduled to overlap with any other pass executions (thus it should be very
860 <!-- _______________________________________________________________________ -->
861 <div class="doc_subsubsection">
862 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
865 <div class="doc_text">
867 <div class="doc_code"><pre>
868 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
871 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
872 function is not allowed to inspect or modify basic blocks other than the
873 parameter, and are not allowed to modify the CFG. A true value must be returned
874 if the basic block is modified.</p>
878 <!-- _______________________________________________________________________ -->
879 <div class="doc_subsubsection">
880 <a name="doFinalization_fn">The <tt>doFinalization(Function &)</tt>
884 <div class="doc_text">
886 <div class="doc_code"><pre>
887 <b>virtual bool</b> doFinalization(Function &F);
890 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
891 called when the pass framework has finished calling <a
892 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
893 program being compiled. This can be used to perform per-function
898 <!-- ======================================================================= -->
899 <div class="doc_subsection">
900 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
903 <div class="doc_text">
905 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
906 executes on the machine-dependent representation of each LLVM function in the
909 <p>Code generator passes are registered and initialized specially by
910 <tt>TargetMachine::addPassesToEmitFile</tt> and similar routines, so they
911 cannot generally be run from the <tt>opt</tt> or <tt>bugpoint</tt>
914 <p>A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
915 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
916 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
917 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
920 <li>Modify or create any LLVM IR Instructions, BasicBlocks, Arguments,
921 Functions, GlobalVariables, GlobalAliases, or Modules.</li>
922 <li>Modify a MachineFunction other than the one currently being processed.</li>
923 <li>Maintain state across invocations of <a
924 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
930 <!-- _______________________________________________________________________ -->
931 <div class="doc_subsubsection">
932 <a name="runOnMachineFunction">The <tt>runOnMachineFunction(MachineFunction
933 &MF)</tt> method</a>
936 <div class="doc_text">
938 <div class="doc_code"><pre>
939 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
942 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
943 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
944 work of your <tt>MachineFunctionPass</tt>.</p>
946 <p>The <tt>runOnMachineFunction</tt> method is called on every
947 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
948 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
949 representation of the function. If you want to get at the LLVM <tt>Function</tt>
950 for the <tt>MachineFunction</tt> you're working on, use
951 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
952 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
953 <tt>MachineFunctionPass</tt>.</p>
957 <!-- *********************************************************************** -->
958 <div class="doc_section">
959 <a name="registration">Pass registration</a>
961 <!-- *********************************************************************** -->
963 <div class="doc_text">
965 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
966 pass registration works, and discussed some of the reasons that it is used and
967 what it does. Here we discuss how and why passes are registered.</p>
969 <p>As we saw above, passes are registered with the <b><tt>RegisterPass</tt></b>
970 template, which requires you to pass at least two
971 parameters. The first parameter is the name of the pass that is to be used on
972 the command line to specify that the pass should be added to a program (for
973 example, with <tt>opt</tt> or <tt>bugpoint</tt>). The second argument is the
974 name of the pass, which is to be used for the <tt>-help</tt> output of
976 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
978 <p>If you want your pass to be easily dumpable, you should
979 implement the virtual <tt>print</tt> method:</p>
983 <!-- _______________________________________________________________________ -->
984 <div class="doc_subsubsection">
985 <a name="print">The <tt>print</tt> method</a>
988 <div class="doc_text">
990 <div class="doc_code"><pre>
991 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
994 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
995 a human readable version of the analysis results. This is useful for debugging
996 an analysis itself, as well as for other people to figure out how an analysis
997 works. Use the <tt>opt -analyze</tt> argument to invoke this method.</p>
999 <p>The <tt>llvm::OStream</tt> parameter specifies the stream to write the results on,
1000 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
1001 program that has been analyzed. Note however that this pointer may be null in
1002 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
1003 debugger), so it should only be used to enhance debug output, it should not be
1008 <!-- *********************************************************************** -->
1009 <div class="doc_section">
1010 <a name="interaction">Specifying interactions between passes</a>
1012 <!-- *********************************************************************** -->
1014 <div class="doc_text">
1016 <p>One of the main responsibilities of the <tt>PassManager</tt> is to make sure
1017 that passes interact with each other correctly. Because <tt>PassManager</tt>
1018 tries to <a href="#passmanager">optimize the execution of passes</a> it must
1019 know how the passes interact with each other and what dependencies exist between
1020 the various passes. To track this, each pass can declare the set of passes that
1021 are required to be executed before the current pass, and the passes which are
1022 invalidated by the current pass.</p>
1024 <p>Typically this functionality is used to require that analysis results are
1025 computed before your pass is run. Running arbitrary transformation passes can
1026 invalidate the computed analysis results, which is what the invalidation set
1027 specifies. If a pass does not implement the <tt><a
1028 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
1029 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
1033 <!-- _______________________________________________________________________ -->
1034 <div class="doc_subsubsection">
1035 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
1038 <div class="doc_text">
1040 <div class="doc_code"><pre>
1041 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
1044 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
1045 invalidated sets may be specified for your transformation. The implementation
1046 should fill in the <tt><a
1047 href="http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
1048 object with information about which passes are required and not invalidated. To
1049 do this, a pass may call any of the following methods on the AnalysisUsage
1053 <!-- _______________________________________________________________________ -->
1054 <div class="doc_subsubsection">
1055 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
1058 <div class="doc_text">
1060 If your pass requires a previous pass to be executed (an analysis for example),
1061 it can use one of these methods to arrange for it to be run before your pass.
1062 LLVM has many different types of analyses and passes that can be required,
1063 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
1064 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
1065 be no critical edges in the CFG when your pass has been run.
1069 Some analyses chain to other analyses to do their job. For example, an <a
1070 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
1071 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
1072 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
1073 used instead of the <tt>addRequired</tt> method. This informs the PassManager
1074 that the transitively required pass should be alive as long as the requiring
1079 <!-- _______________________________________________________________________ -->
1080 <div class="doc_subsubsection">
1081 <a name="AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a>
1084 <div class="doc_text">
1086 One of the jobs of the PassManager is to optimize how and when analyses are run.
1087 In particular, it attempts to avoid recomputing data unless it needs to. For
1088 this reason, passes are allowed to declare that they preserve (i.e., they don't
1089 invalidate) an existing analysis if it's available. For example, a simple
1090 constant folding pass would not modify the CFG, so it can't possibly affect the
1091 results of dominator analysis. By default, all passes are assumed to invalidate
1096 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
1097 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
1098 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
1099 not modify the LLVM program at all (which is true for analyses), and the
1100 <tt>setPreservesCFG</tt> method can be used by transformations that change
1101 instructions in the program but do not modify the CFG or terminator instructions
1102 (note that this property is implicitly set for <a
1103 href="#BasicBlockPass">BasicBlockPass</a>'s).
1107 <tt>addPreserved</tt> is particularly useful for transformations like
1108 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
1109 and dominator related analyses if they exist, so it can preserve them, despite
1110 the fact that it hacks on the CFG.
1114 <!-- _______________________________________________________________________ -->
1115 <div class="doc_subsubsection">
1116 <a name="AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a>
1119 <div class="doc_text">
1121 <div class="doc_code"><pre>
1122 <i>// This is an example implementation from an analysis, which does not modify
1123 // the program at all, yet has a prerequisite.</i>
1124 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1125 AU.setPreservesAll();
1126 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1132 <div class="doc_code"><pre>
1133 <i>// This example modifies the program, but does not modify the CFG</i>
1134 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1135 AU.setPreservesCFG();
1136 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1142 <!-- _______________________________________________________________________ -->
1143 <div class="doc_subsubsection">
1144 <a name="getAnalysis">The <tt>getAnalysis<></tt> and
1145 <tt>getAnalysisIfAvailable<></tt> methods</a>
1148 <div class="doc_text">
1150 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1151 your class, providing you with access to the passes that you declared that you
1152 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
1153 method. It takes a single template argument that specifies which pass class you
1154 want, and returns a reference to that pass. For example:</p>
1156 <div class="doc_code"><pre>
1157 bool LICM::runOnFunction(Function &F) {
1158 LoopInfo &LI = getAnalysis<LoopInfo>();
1163 <p>This method call returns a reference to the pass desired. You may get a
1164 runtime assertion failure if you attempt to get an analysis that you did not
1165 declare as required in your <a
1166 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
1167 method can be called by your <tt>run*</tt> method implementation, or by any
1168 other local method invoked by your <tt>run*</tt> method.
1170 A module level pass can use function level analysis info using this interface.
1173 <div class="doc_code"><pre>
1174 bool ModuleLevelPass::runOnModule(Module &M) {
1176 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
1181 <p>In above example, runOnFunction for DominatorTree is called by pass manager
1182 before returning a reference to the desired pass.</p>
1185 If your pass is capable of updating analyses if they exist (e.g.,
1186 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1187 <tt>getAnalysisIfAvailable</tt> method, which returns a pointer to the analysis
1188 if it is active. For example:</p>
1190 <div class="doc_code"><pre>
1192 if (DominatorSet *DS = getAnalysisIfAvailable<DominatorSet>()) {
1193 <i>// A DominatorSet is active. This code will update it.</i>
1200 <!-- *********************************************************************** -->
1201 <div class="doc_section">
1202 <a name="analysisgroup">Implementing Analysis Groups</a>
1204 <!-- *********************************************************************** -->
1206 <div class="doc_text">
1208 <p>Now that we understand the basics of how passes are defined, how they are
1209 used, and how they are required from other passes, it's time to get a little bit
1210 fancier. All of the pass relationships that we have seen so far are very
1211 simple: one pass depends on one other specific pass to be run before it can run.
1212 For many applications, this is great, for others, more flexibility is
1215 <p>In particular, some analyses are defined such that there is a single simple
1216 interface to the analysis results, but multiple ways of calculating them.
1217 Consider alias analysis for example. The most trivial alias analysis returns
1218 "may alias" for any alias query. The most sophisticated analysis a
1219 flow-sensitive, context-sensitive interprocedural analysis that can take a
1220 significant amount of time to execute (and obviously, there is a lot of room
1221 between these two extremes for other implementations). To cleanly support
1222 situations like this, the LLVM Pass Infrastructure supports the notion of
1223 Analysis Groups.</p>
1227 <!-- _______________________________________________________________________ -->
1228 <div class="doc_subsubsection">
1229 <a name="agconcepts">Analysis Group Concepts</a>
1232 <div class="doc_text">
1234 <p>An Analysis Group is a single simple interface that may be implemented by
1235 multiple different passes. Analysis Groups can be given human readable names
1236 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1237 class. An analysis group may have one or more implementations, one of which is
1238 the "default" implementation.</p>
1240 <p>Analysis groups are used by client passes just like other passes are: the
1241 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1242 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
1243 scans the available passes to see if any implementations of the analysis group
1244 are available. If none is available, the default implementation is created for
1245 the pass to use. All standard rules for <A href="#interaction">interaction
1246 between passes</a> still apply.</p>
1248 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
1249 passes, all analysis group implementations must be registered, and must use the
1250 <A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
1251 implementation pool. Also, a default implementation of the interface
1252 <b>must</b> be registered with <A
1253 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
1255 <p>As a concrete example of an Analysis Group in action, consider the <a
1256 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1257 analysis group. The default implementation of the alias analysis interface (the
1259 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1260 pass) just does a few simple checks that don't require significant analysis to
1261 compute (such as: two different globals can never alias each other, etc).
1262 Passes that use the <tt><a
1263 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1264 interface (for example the <tt><a
1265 href="http://llvm.org/doxygen/structGCSE.html">gcse</a></tt> pass), do
1266 not care which implementation of alias analysis is actually provided, they just
1267 use the designated interface.</p>
1269 <p>From the user's perspective, commands work just like normal. Issuing the
1270 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1271 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1272 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1273 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1274 hypothetical example) instead.</p>
1278 <!-- _______________________________________________________________________ -->
1279 <div class="doc_subsubsection">
1280 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1283 <div class="doc_text">
1285 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1286 group itself as well as add pass implementations to the analysis group. First,
1287 an analysis should be registered, with a human readable name provided for it.
1288 Unlike registration of passes, there is no command line argument to be specified
1289 for the Analysis Group Interface itself, because it is "abstract":</p>
1291 <div class="doc_code"><pre>
1292 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1295 <p>Once the analysis is registered, passes can declare that they are valid
1296 implementations of the interface by using the following code:</p>
1298 <div class="doc_code"><pre>
1300 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1301 RegisterPass<FancyAA>
1302 B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
1304 //<i> Declare that we implement the AliasAnalysis interface</i>
1305 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> C(B);
1309 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1310 uses the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
1311 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1312 analysis group. Every implementation of an analysis group should join using
1313 this template. A single pass may join multiple different analysis groups with
1316 <div class="doc_code"><pre>
1318 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1319 RegisterPass<<a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1320 D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
1322 //<i> Declare that we implement the AliasAnalysis interface</i>
1323 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <b>true</b>> E(D);
1327 <p>Here we show how the default implementation is specified (using the extra
1328 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1329 one default implementation available at all times for an Analysis Group to be
1330 used. Only default implementation can derive from <tt>ImmutablePass</tt>.
1331 Here we declare that the
1332 <tt><a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1333 pass is the default implementation for the interface.</p>
1337 <!-- *********************************************************************** -->
1338 <div class="doc_section">
1339 <a name="passStatistics">Pass Statistics</a>
1341 <!-- *********************************************************************** -->
1343 <div class="doc_text">
1345 href="http://llvm.org/doxygen/Statistic_8h-source.html"><tt>Statistic</tt></a>
1346 class is designed to be an easy way to expose various success
1347 metrics from passes. These statistics are printed at the end of a
1348 run, when the -stats command line option is enabled on the command
1349 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1354 <!-- *********************************************************************** -->
1355 <div class="doc_section">
1356 <a name="passmanager">What PassManager does</a>
1358 <!-- *********************************************************************** -->
1360 <div class="doc_text">
1363 href="http://llvm.org/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1365 href="http://llvm.org/doxygen/classllvm_1_1PassManager.html">class</a>
1366 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1367 are set up correctly, and then schedules passes to run efficiently. All of the
1368 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1371 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1372 time of a series of passes:</p>
1375 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1376 recomputing analysis results as much as possible. This means keeping track of
1377 which analyses are available already, which analyses get invalidated, and which
1378 analyses are needed to be run for a pass. An important part of work is that the
1379 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1380 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1381 results as soon as they are no longer needed.</li>
1383 <li><b>Pipeline the execution of passes on the program</b> - The
1384 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1385 of a series of passes by pipelining the passes together. This means that, given
1386 a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1387 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1388 the first function, then all of the <a
1389 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1390 etc... until the entire program has been run through the passes.
1392 <p>This improves the cache behavior of the compiler, because it is only touching
1393 the LLVM program representation for a single function at a time, instead of
1394 traversing the entire program. It reduces the memory consumption of compiler,
1395 because, for example, only one <a
1396 href="http://llvm.org/doxygen/classllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1397 needs to be calculated at a time. This also makes it possible to implement
1399 href="#SMP">interesting enhancements</a> in the future.</p></li>
1403 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1404 much information it has about the behaviors of the passes it is scheduling. For
1405 example, the "preserved" set is intentionally conservative in the face of an
1406 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1407 Not implementing when it should be implemented will have the effect of not
1408 allowing any analysis results to live across the execution of your pass.</p>
1410 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1411 options that is useful for debugging pass execution, seeing how things work, and
1412 diagnosing when you should be preserving more analyses than you currently are
1413 (To get information about all of the variants of the <tt>--debug-pass</tt>
1414 option, just type '<tt>opt -help-hidden</tt>').</p>
1416 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1417 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1418 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1420 <div class="doc_code"><pre>
1421 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1423 Function Pass Manager
1424 Dominator Set Construction
1425 Immediate Dominators Construction
1426 Global Common Subexpression Elimination
1427 -- Immediate Dominators Construction
1428 -- Global Common Subexpression Elimination
1429 Natural Loop Construction
1430 Loop Invariant Code Motion
1431 -- Natural Loop Construction
1432 -- Loop Invariant Code Motion
1434 -- Dominator Set Construction
1440 <p>This output shows us when passes are constructed and when the analysis
1441 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1442 GCSE uses dominator and immediate dominator information to do its job. The LICM
1443 pass uses natural loop information, which uses dominator sets, but not immediate
1444 dominators. Because immediate dominators are no longer useful after the GCSE
1445 pass, it is immediately destroyed. The dominator sets are then reused to
1446 compute natural loop information, which is then used by the LICM pass.</p>
1448 <p>After the LICM pass, the module verifier runs (which is automatically added
1449 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1450 resultant LLVM code is well formed. After it finishes, the dominator set
1451 information is destroyed, after being computed once, and shared by three
1454 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1455 World</a> pass in between the two passes:</p>
1457 <div class="doc_code"><pre>
1458 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1460 Function Pass Manager
1461 Dominator Set Construction
1462 Immediate Dominators Construction
1463 Global Common Subexpression Elimination
1464 <b>-- Dominator Set Construction</b>
1465 -- Immediate Dominators Construction
1466 -- Global Common Subexpression Elimination
1467 <b> Hello World Pass
1469 Dominator Set Construction</b>
1470 Natural Loop Construction
1471 Loop Invariant Code Motion
1472 -- Natural Loop Construction
1473 -- Loop Invariant Code Motion
1475 -- Dominator Set Construction
1484 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1485 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1486 we need to add the following <a
1487 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1489 <div class="doc_code"><pre>
1490 <i>// We don't modify the program, so we preserve all analyses</i>
1491 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1492 AU.setPreservesAll();
1496 <p>Now when we run our pass, we get this output:</p>
1498 <div class="doc_code"><pre>
1499 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1500 Pass Arguments: -gcse -hello -licm
1502 Function Pass Manager
1503 Dominator Set Construction
1504 Immediate Dominators Construction
1505 Global Common Subexpression Elimination
1506 -- Immediate Dominators Construction
1507 -- Global Common Subexpression Elimination
1510 Natural Loop Construction
1511 Loop Invariant Code Motion
1512 -- Loop Invariant Code Motion
1513 -- Natural Loop Construction
1515 -- Dominator Set Construction
1524 <p>Which shows that we don't accidentally invalidate dominator information
1525 anymore, and therefore do not have to compute it twice.</p>
1529 <!-- _______________________________________________________________________ -->
1530 <div class="doc_subsubsection">
1531 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1534 <div class="doc_text">
1536 <div class="doc_code"><pre>
1537 <b>virtual void</b> releaseMemory();
1540 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1541 results, and how long to keep them around for. Because the lifetime of the pass
1542 object itself is effectively the entire duration of the compilation process, we
1543 need some way to free analysis results when they are no longer useful. The
1544 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1546 <p>If you are writing an analysis or any other pass that retains a significant
1547 amount of state (for use by another pass which "requires" your pass and uses the
1548 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1549 <tt>releaseMemory</tt> to, well, release the memory allocated to maintain this
1550 internal state. This method is called after the <tt>run*</tt> method for the
1551 class, before the next call of <tt>run*</tt> in your pass.</p>
1555 <!-- *********************************************************************** -->
1556 <div class="doc_section">
1557 <a name="registering">Registering dynamically loaded passes</a>
1559 <!-- *********************************************************************** -->
1561 <div class="doc_text">
1563 <p><i>Size matters</i> when constructing production quality tools using llvm,
1564 both for the purposes of distribution, and for regulating the resident code size
1565 when running on the target system. Therefore, it becomes desirable to
1566 selectively use some passes, while omitting others and maintain the flexibility
1567 to change configurations later on. You want to be able to do all this, and,
1568 provide feedback to the user. This is where pass registration comes into
1571 <p>The fundamental mechanisms for pass registration are the
1572 <tt>MachinePassRegistry</tt> class and subclasses of
1573 <tt>MachinePassRegistryNode</tt>.</p>
1575 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1576 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1577 communicates additions and deletions to the command line interface.</p>
1579 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1580 information provided about a particular pass. This information includes the
1581 command line name, the command help string and the address of the function used
1582 to create an instance of the pass. A global static constructor of one of these
1583 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1584 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1585 in the tool will be registered at start up. A dynamically loaded pass will
1586 register on load and unregister at unload.</p>
1590 <!-- _______________________________________________________________________ -->
1591 <div class="doc_subsection">
1592 <a name="registering_existing">Using existing registries</a>
1595 <div class="doc_text">
1597 <p>There are predefined registries to track instruction scheduling
1598 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1599 machine passes. Here we will describe how to <i>register</i> a register
1600 allocator machine pass.</p>
1602 <p>Implement your register allocator machine pass. In your register allocator
1603 .cpp file add the following include;</p>
1605 <div class="doc_code"><pre>
1606 #include "llvm/CodeGen/RegAllocRegistry.h"
1609 <p>Also in your register allocator .cpp file, define a creator function in the
1612 <div class="doc_code"><pre>
1613 FunctionPass *createMyRegisterAllocator() {
1614 return new MyRegisterAllocator();
1618 <p>Note that the signature of this function should match the type of
1619 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1620 "installing" declaration, in the form;</p>
1622 <div class="doc_code"><pre>
1623 static RegisterRegAlloc myRegAlloc("myregalloc",
1624 " my register allocator help string",
1625 createMyRegisterAllocator);
1628 <p>Note the two spaces prior to the help string produces a tidy result on the
1631 <div class="doc_code"><pre>
1634 -regalloc - Register allocator to use (default=linearscan)
1635 =linearscan - linear scan register allocator
1636 =local - local register allocator
1637 =simple - simple register allocator
1638 =myregalloc - my register allocator help string
1642 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1643 an option. Registering instruction schedulers is similar except use the
1644 <tt>RegisterScheduler</tt> class. Note that the
1645 <tt>RegisterScheduler::FunctionPassCtor</tt> is significantly different from
1646 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1648 <p>To force the load/linking of your register allocator into the llc/lli tools,
1649 add your creator function's global declaration to "Passes.h" and add a "pseudo"
1650 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1655 <!-- _______________________________________________________________________ -->
1656 <div class="doc_subsection">
1657 <a name="registering_new">Creating new registries</a>
1660 <div class="doc_text">
1662 <p>The easiest way to get started is to clone one of the existing registries; we
1663 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1664 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1666 <p>Then you need to declare the registry. Example: if your pass registry is
1667 <tt>RegisterMyPasses</tt> then define;</p>
1669 <div class="doc_code"><pre>
1670 MachinePassRegistry RegisterMyPasses::Registry;
1673 <p>And finally, declare the command line option for your passes. Example:</p>
1675 <div class="doc_code"><pre>
1676 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1677 RegisterPassParser<RegisterMyPasses> >
1679 cl::init(&createDefaultMyPass),
1680 cl::desc("my pass option help"));
1683 <p>Here the command option is "mypass", with createDefaultMyPass as the default
1688 <!-- *********************************************************************** -->
1689 <div class="doc_section">
1690 <a name="debughints">Using GDB with dynamically loaded passes</a>
1692 <!-- *********************************************************************** -->
1694 <div class="doc_text">
1696 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1697 should be. First of all, you can't set a breakpoint in a shared object that has
1698 not been loaded yet, and second of all there are problems with inlined functions
1699 in shared objects. Here are some suggestions to debugging your pass with
1702 <p>For sake of discussion, I'm going to assume that you are debugging a
1703 transformation invoked by <tt>opt</tt>, although nothing described here depends
1708 <!-- _______________________________________________________________________ -->
1709 <div class="doc_subsubsection">
1710 <a name="breakpoint">Setting a breakpoint in your pass</a>
1713 <div class="doc_text">
1715 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1717 <div class="doc_code"><pre>
1720 Copyright 2000 Free Software Foundation, Inc.
1721 GDB is free software, covered by the GNU General Public License, and you are
1722 welcome to change it and/or distribute copies of it under certain conditions.
1723 Type "show copying" to see the conditions.
1724 There is absolutely no warranty for GDB. Type "show warranty" for details.
1725 This GDB was configured as "sparc-sun-solaris2.6"...
1729 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1730 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1731 (the shared object isn't loaded until runtime), we must execute the process, and
1732 have it stop before it invokes our pass, but after it has loaded the shared
1733 object. The most foolproof way of doing this is to set a breakpoint in
1734 <tt>PassManager::run</tt> and then run the process with the arguments you
1737 <div class="doc_code"><pre>
1738 (gdb) <b>break llvm::PassManager::run</b>
1739 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1740 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]</b>
1741 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1742 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1743 70 bool PassManager::run(Module &M) { return PM->run(M); }
1747 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1748 now free to set breakpoints in your pass so that you can trace through execution
1749 or do other standard debugging stuff.</p>
1753 <!-- _______________________________________________________________________ -->
1754 <div class="doc_subsubsection">
1755 <a name="debugmisc">Miscellaneous Problems</a>
1758 <div class="doc_text">
1760 <p>Once you have the basics down, there are a couple of problems that GDB has,
1761 some with solutions, some without.</p>
1764 <li>Inline functions have bogus stack information. In general, GDB does a
1765 pretty good job getting stack traces and stepping through inline functions.
1766 When a pass is dynamically loaded however, it somehow completely loses this
1767 capability. The only solution I know of is to de-inline a function (move it
1768 from the body of a class to a .cpp file).</li>
1770 <li>Restarting the program breaks breakpoints. After following the information
1771 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1772 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1773 and you start getting errors about breakpoints being unsettable. The only way I
1774 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1775 already set in your pass, run the program, and re-set the breakpoints once
1776 execution stops in <tt>PassManager::run</tt>.</li>
1780 <p>Hopefully these tips will help with common case debugging situations. If
1781 you'd like to contribute some tips of your own, just contact <a
1782 href="mailto:sabre@nondot.org">Chris</a>.</p>
1786 <!-- *********************************************************************** -->
1787 <div class="doc_section">
1788 <a name="future">Future extensions planned</a>
1790 <!-- *********************************************************************** -->
1792 <div class="doc_text">
1794 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1795 some nifty stuff, there are things we'd like to add in the future. Here is
1796 where we are going:</p>
1800 <!-- _______________________________________________________________________ -->
1801 <div class="doc_subsubsection">
1802 <a name="SMP">Multithreaded LLVM</a>
1805 <div class="doc_text">
1807 <p>Multiple CPU machines are becoming more common and compilation can never be
1808 fast enough: obviously we should allow for a multithreaded compiler. Because of
1809 the semantics defined for passes above (specifically they cannot maintain state
1810 across invocations of their <tt>run*</tt> methods), a nice clean way to
1811 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1812 to create multiple instances of each pass object, and allow the separate
1813 instances to be hacking on different parts of the program at the same time.</p>
1815 <p>This implementation would prevent each of the passes from having to implement
1816 multithreaded constructs, requiring only the LLVM core to have locking in a few
1817 places (for global resources). Although this is a simple extension, we simply
1818 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1819 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1823 <!-- *********************************************************************** -->
1826 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1827 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
1828 <a href="http://validator.w3.org/check/referer"><img
1829 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
1831 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1832 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1833 Last modified: $Date$