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="_static/llvm.css" type="text/css">
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="#RegionPass">The <tt>RegionPass</tt> class</a>
56 <li><a href="#doInitialization_region">The <tt>doInitialization(Region *,
57 RGPassManager &)</tt> method</a></li>
58 <li><a href="#runOnRegion">The <tt>runOnRegion</tt> method</a></li>
59 <li><a href="#doFinalization_region">The <tt>doFinalization()
62 <li><a href="#BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
64 <li><a href="#doInitialization_fn">The <tt>doInitialization(Function
65 &)</tt> method</a></li>
66 <li><a href="#runOnBasicBlock">The <tt>runOnBasicBlock</tt>
68 <li><a href="#doFinalization_fn">The <tt>doFinalization(Function
69 &)</tt> method</a></li>
71 <li><a href="#MachineFunctionPass">The <tt>MachineFunctionPass</tt>
74 <li><a href="#runOnMachineFunction">The
75 <tt>runOnMachineFunction(MachineFunction &)</tt> method</a></li>
78 <li><a href="#registration">Pass Registration</a>
80 <li><a href="#print">The <tt>print</tt> method</a></li>
82 <li><a href="#interaction">Specifying interactions between passes</a>
84 <li><a href="#getAnalysisUsage">The <tt>getAnalysisUsage</tt>
86 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
87 <li><a href="#AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a></li>
88 <li><a href="#AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a></li>
89 <li><a href="#getAnalysis">The <tt>getAnalysis<></tt> and
90 <tt>getAnalysisIfAvailable<></tt> methods</a></li>
92 <li><a href="#analysisgroup">Implementing Analysis Groups</a>
94 <li><a href="#agconcepts">Analysis Group Concepts</a></li>
95 <li><a href="#registerag">Using <tt>RegisterAnalysisGroup</tt></a></li>
97 <li><a href="#passStatistics">Pass Statistics</a>
98 <li><a href="#passmanager">What PassManager does</a>
100 <li><a href="#releaseMemory">The <tt>releaseMemory</tt> method</a></li>
102 <li><a href="#registering">Registering dynamically loaded passes</a>
104 <li><a href="#registering_existing">Using existing registries</a></li>
105 <li><a href="#registering_new">Creating new registries</a></li>
107 <li><a href="#debughints">Using GDB with dynamically loaded passes</a>
109 <li><a href="#breakpoint">Setting a breakpoint in your pass</a></li>
110 <li><a href="#debugmisc">Miscellaneous Problems</a></li>
112 <li><a href="#future">Future extensions planned</a>
114 <li><a href="#SMP">Multithreaded LLVM</a></li>
118 <div class="doc_author">
119 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
120 <a href="mailto:jlaskey@mac.com">Jim Laskey</a></p>
123 <!-- *********************************************************************** -->
125 <a name="introduction">Introduction - What is a pass?</a>
127 <!-- *********************************************************************** -->
131 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
132 passes are where most of the interesting parts of the compiler exist. Passes
133 perform the transformations and optimizations that make up the compiler, they
134 build the analysis results that are used by these transformations, and they are,
135 above all, a structuring technique for compiler code.</p>
137 <p>All LLVM passes are subclasses of the <tt><a
138 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
139 class, which implement functionality by overriding virtual methods inherited
140 from <tt>Pass</tt>. Depending on how your pass works, you should inherit from
141 the <tt><a href="#ModulePass">ModulePass</a></tt>, <tt><a
142 href="#CallGraphSCCPass">CallGraphSCCPass</a></tt>, <tt><a
143 href="#FunctionPass">FunctionPass</a></tt>, or <tt><a
144 href="#LoopPass">LoopPass</a></tt>, or <tt><a
145 href="#RegionPass">RegionPass</a></tt>, or <tt><a
146 href="#BasicBlockPass">BasicBlockPass</a></tt> classes, which gives the system
147 more information about what your pass does, and how it can be combined with
148 other passes. One of the main features of the LLVM Pass Framework is that it
149 schedules passes to run in an efficient way based on the constraints that your
150 pass meets (which are indicated by which class they derive from).</p>
152 <p>We start by showing you how to construct a pass, everything from setting up
153 the code, to compiling, loading, and executing it. After the basics are down,
154 more advanced features are discussed.</p>
158 <!-- *********************************************************************** -->
160 <a name="quickstart">Quick Start - Writing hello world</a>
162 <!-- *********************************************************************** -->
166 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
167 is designed to simply print out the name of non-external functions that exist in
168 the program being compiled. It does not modify the program at all, it just
169 inspects it. The source code and files for this pass are available in the LLVM
170 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
172 <!-- ======================================================================= -->
174 <a name="makefile">Setting up the build environment</a>
179 <p>First, configure and build LLVM. This needs to be done directly inside the
180 LLVM source tree rather than in a separate objects directory.
181 Next, you need to create a new directory somewhere in the LLVM source
182 base. For this example, we'll assume that you made
183 <tt>lib/Transforms/Hello</tt>. Finally, you must set up a build script
184 (Makefile) that will compile the source code for the new pass. To do this,
185 copy the following into <tt>Makefile</tt>:</p>
188 <div class="doc_code"><pre>
189 # Makefile for hello pass
191 # Path to top level of LLVM hierarchy
194 # Name of the library to build
197 # Make the shared library become a loadable module so the tools can
198 # dlopen/dlsym on the resulting library.
201 # Include the makefile implementation stuff
202 include $(LEVEL)/Makefile.common
205 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
206 directory are to be compiled and linked together into a shared object
207 <tt>$(LEVEL)/Debug+Asserts/lib/Hello.so</tt> that can be dynamically loaded by
208 the <tt>opt</tt> or <tt>bugpoint</tt> tools via their <tt>-load</tt> options.
209 If your operating system uses a suffix other than .so (such as windows or
210 Mac OS/X), the appropriate extension will be used.</p>
212 <p>If you are used CMake to build LLVM, see
213 <a href="CMake.html#passdev">Developing an LLVM pass with CMake</a>.</p>
215 <p>Now that we have the build scripts set up, we just need to write the code for
220 <!-- ======================================================================= -->
222 <a name="basiccode">Basic code required</a>
227 <p>Now that we have a way to compile our new pass, we just have to write it.
230 <div class="doc_code">
232 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
233 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
234 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
238 <p>Which are needed because we are writing a <tt><a
239 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>,
240 we are operating on <tt><a
241 href="http://llvm.org/doxygen/classllvm_1_1Function.html">Function</a></tt>'s,
242 and we will be doing some printing.</p>
246 <div class="doc_code">
248 <b>using namespace llvm;</b>
252 <p>... which is required because the functions from the include files
253 live in the llvm namespace.</p>
257 <div class="doc_code">
263 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
264 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
265 things declared inside of the anonymous namespace visible only to the current
266 file. If you're not familiar with them, consult a decent C++ book for more
269 <p>Next, we declare our pass itself:</p>
271 <div class="doc_code">
273 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
277 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
278 href="http://llvm.org/doxygen/classllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
279 The different builtin pass subclasses are described in detail <a
280 href="#passtype">later</a>, but for now, know that <a
281 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate on a function at a
284 <div class="doc_code">
287 Hello() : FunctionPass(ID) {}
291 <p>This declares pass identifier used by LLVM to identify pass. This allows LLVM
292 to avoid using expensive C++ runtime information.</p>
294 <div class="doc_code">
296 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
297 errs() << "<i>Hello: </i>";
298 errs().write_escaped(F.getName()) << "\n";
301 }; <i>// end of struct Hello</i>
302 } <i>// end of anonymous namespace</i>
306 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
307 which overloads an abstract virtual method inherited from <a
308 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
309 to do our thing, so we just print out our message with the name of each
312 <div class="doc_code">
318 <p>We initialize pass ID here. LLVM uses ID's address to identify a pass, so
319 initialization value is not important.</p>
321 <div class="doc_code">
323 static RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>",
324 false /* Only looks at CFG */,
325 false /* Analysis Pass */);
329 <p>Lastly, we <a href="#registration">register our class</a> <tt>Hello</tt>,
330 giving it a command line argument "<tt>hello</tt>", and a name "<tt>Hello World
331 Pass</tt>". The last two arguments describe its behavior: if a pass walks CFG
332 without modifying it then the third argument is set to <tt>true</tt>; if a pass
333 is an analysis pass, for example dominator tree pass, then <tt>true</tt> is
334 supplied as the fourth argument.</p>
336 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
338 <div class="doc_code">
340 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
341 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
342 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
344 <b>using namespace llvm;</b>
347 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
350 Hello() : FunctionPass(ID) {}
352 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
353 errs() << "<i>Hello: </i>";
354 errs().write_escaped(F.getName()) << '\n';
362 static RegisterPass<Hello> X("hello", "Hello World Pass", false, false);
366 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
367 command in the local directory and you should get a new file
368 "<tt>Debug+Asserts/lib/Hello.so</tt>" under the top level directory of the LLVM
369 source tree (not in the local directory). Note that everything in this file is
370 contained in an anonymous namespace — this reflects the fact that passes
371 are self contained units that do not need external interfaces (although they can
372 have them) to be useful.</p>
376 <!-- ======================================================================= -->
378 <a name="running">Running a pass with <tt>opt</tt></a>
383 <p>Now that you have a brand new shiny shared object file, we can use the
384 <tt>opt</tt> command to run an LLVM program through your pass. Because you
385 registered your pass with <tt>RegisterPass</tt>, you will be able to
386 use the <tt>opt</tt> tool to access it, once loaded.</p>
388 <p>To test it, follow the example at the end of the <a
389 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
390 LLVM. We can now run the bitcode file (<tt>hello.bc</tt>) for the program
391 through our transformation like this (or course, any bitcode file will
394 <div class="doc_code"><pre>
395 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello < hello.bc > /dev/null
401 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
402 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
403 argument (which is one reason you need to <a href="#registration">register your
404 pass</a>). Because the hello pass does not modify the program in any
405 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
406 <tt>/dev/null</tt>).</p>
408 <p>To see what happened to the other string you registered, try running
409 <tt>opt</tt> with the <tt>-help</tt> option:</p>
411 <div class="doc_code"><pre>
412 $ opt -load ../../../Debug+Asserts/lib/Hello.so -help
413 OVERVIEW: llvm .bc -> .bc modular optimizer
415 USAGE: opt [options] <input bitcode>
418 Optimizations available:
420 -globalopt - Global Variable Optimizer
421 -globalsmodref-aa - Simple mod/ref analysis for globals
422 -gvn - Global Value Numbering
423 <b>-hello - Hello World Pass</b>
424 -indvars - Induction Variable Simplification
425 -inline - Function Integration/Inlining
426 -insert-edge-profiling - Insert instrumentation for edge profiling
430 <p>The pass name gets added as the information string for your pass, giving some
431 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
432 would go ahead and make it do the cool transformations you want. Once you get
433 it all working and tested, it may become useful to find out how fast your pass
434 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
435 line option (<tt>--time-passes</tt>) that allows you to get information about
436 the execution time of your pass along with the other passes you queue up. For
439 <div class="doc_code"><pre>
440 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
444 ===============================================================================
445 ... Pass execution timing report ...
446 ===============================================================================
447 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
449 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
450 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bitcode Writer
451 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
452 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
453 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
454 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
457 <p>As you can see, our implementation above is pretty fast :). The additional
458 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
459 that the LLVM emitted by your pass is still valid and well formed LLVM, which
460 hasn't been broken somehow.</p>
462 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
463 about some more details of how they work and how to use them.</p>
469 <!-- *********************************************************************** -->
471 <a name="passtype">Pass classes and requirements</a>
473 <!-- *********************************************************************** -->
477 <p>One of the first things that you should do when designing a new pass is to
478 decide what class you should subclass for your pass. The <a
479 href="#basiccode">Hello World</a> example uses the <tt><a
480 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
481 did not discuss why or when this should occur. Here we talk about the classes
482 available, from the most general to the most specific.</p>
484 <p>When choosing a superclass for your Pass, you should choose the <b>most
485 specific</b> class possible, while still being able to meet the requirements
486 listed. This gives the LLVM Pass Infrastructure information necessary to
487 optimize how passes are run, so that the resultant compiler isn't unnecessarily
490 <!-- ======================================================================= -->
492 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
497 <p>The most plain and boring type of pass is the "<tt><a
498 href="http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
499 class. This pass type is used for passes that do not have to be run, do not
500 change state, and never need to be updated. This is not a normal type of
501 transformation or analysis, but can provide information about the current
502 compiler configuration.</p>
504 <p>Although this pass class is very infrequently used, it is important for
505 providing information about the current target machine being compiled for, and
506 other static information that can affect the various transformations.</p>
508 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
509 invalidated, and are never "run".</p>
513 <!-- ======================================================================= -->
515 <a name="ModulePass">The <tt>ModulePass</tt> class</a>
521 href="http://llvm.org/doxygen/classllvm_1_1ModulePass.html">ModulePass</a></tt>"
522 class is the most general of all superclasses that you can use. Deriving from
523 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
524 referring to function bodies in no predictable order, or adding and removing
525 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
526 subclasses, no optimization can be done for their execution.</p>
528 <p>A module pass can use function level passes (e.g. dominators) using
529 the getAnalysis interface
530 <tt>getAnalysis<DominatorTree>(llvm::Function *)</tt> to provide the
531 function to retrieve analysis result for, if the function pass does not require
532 any module or immutable passes. Note that this can only be done for functions for which the
533 analysis ran, e.g. in the case of dominators you should only ask for the
534 DominatorTree for function definitions, not declarations.</p>
536 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
537 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
538 following signature:</p>
540 <!-- _______________________________________________________________________ -->
542 <a name="runOnModule">The <tt>runOnModule</tt> method</a>
547 <div class="doc_code"><pre>
548 <b>virtual bool</b> runOnModule(Module &M) = 0;
551 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
552 It should return true if the module was modified by the transformation and
559 <!-- ======================================================================= -->
561 <a name="CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
567 href="http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html">CallGraphSCCPass</a></tt>"
568 is used by passes that need to traverse the program bottom-up on the call graph
569 (callees before callers). Deriving from CallGraphSCCPass provides some
570 mechanics for building and traversing the CallGraph, but also allows the system
571 to optimize execution of CallGraphSCCPass's. If your pass meets the
572 requirements outlined below, and doesn't meet the requirements of a <tt><a
573 href="#FunctionPass">FunctionPass</a></tt> or <tt><a
574 href="#BasicBlockPass">BasicBlockPass</a></tt>, you should derive from
575 <tt>CallGraphSCCPass</tt>.</p>
577 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
579 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
583 <li>... <em>not allowed</em> to inspect or modify any <tt>Function</tt>s other
584 than those in the current SCC and the direct callers and direct callees of the
587 <li>... <em>required</em> to preserve the current CallGraph object, updating it
588 to reflect any changes made to the program.</li>
590 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
591 though they may change the contents of an SCC.</li>
593 <li>... <em>allowed</em> to add or remove global variables from the current
596 <li>... <em>allowed</em> to maintain state across invocations of
597 <a href="#runOnSCC"><tt>runOnSCC</tt></a> (including global data).</li>
600 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
601 because it has to handle SCCs with more than one node in it. All of the virtual
602 methods described below should return true if they modified the program, or
603 false if they didn't.</p>
605 <!-- _______________________________________________________________________ -->
607 <a name="doInitialization_scc">
608 The <tt>doInitialization(CallGraph &)</tt> method
614 <div class="doc_code"><pre>
615 <b>virtual bool</b> doInitialization(CallGraph &CG);
618 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
619 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
620 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
621 is designed to do simple initialization type of stuff that does not depend on
622 the SCCs being processed. The <tt>doInitialization</tt> method call is not
623 scheduled to overlap with any other pass executions (thus it should be very
628 <!-- _______________________________________________________________________ -->
630 <a name="runOnSCC">The <tt>runOnSCC</tt> method</a>
635 <div class="doc_code"><pre>
636 <b>virtual bool</b> runOnSCC(CallGraphSCC &SCC) = 0;
639 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
640 should return true if the module was modified by the transformation, false
645 <!-- _______________________________________________________________________ -->
647 <a name="doFinalization_scc">
648 The <tt>doFinalization(CallGraph &)</tt> method
654 <div class="doc_code"><pre>
655 <b>virtual bool</b> doFinalization(CallGraph &CG);
658 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
659 called when the pass framework has finished calling <a
660 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
661 program being compiled.</p>
667 <!-- ======================================================================= -->
669 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
674 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
675 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
676 subclasses do have a predictable, local behavior that can be expected by the
677 system. All <tt>FunctionPass</tt> execute on each function in the program
678 independent of all of the other functions in the program.
679 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
680 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
682 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
685 <li>Modify a Function other than the one currently being processed.</li>
686 <li>Add or remove Function's from the current Module.</li>
687 <li>Add or remove global variables from the current Module.</li>
688 <li>Maintain state across invocations of
689 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
692 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
693 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
694 may overload three virtual methods to do their work. All of these methods
695 should return true if they modified the program, or false if they didn't.</p>
697 <!-- _______________________________________________________________________ -->
699 <a name="doInitialization_mod">
700 The <tt>doInitialization(Module &)</tt> method
706 <div class="doc_code"><pre>
707 <b>virtual bool</b> doInitialization(Module &M);
710 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
711 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
712 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
713 is designed to do simple initialization type of stuff that does not depend on
714 the functions being processed. The <tt>doInitialization</tt> method call is not
715 scheduled to overlap with any other pass executions (thus it should be very
718 <p>A good example of how this method should be used is the <a
719 href="http://llvm.org/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
720 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
721 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
722 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
723 free functions that it needs, adding prototypes to the module if necessary.</p>
727 <!-- _______________________________________________________________________ -->
729 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
734 <div class="doc_code"><pre>
735 <b>virtual bool</b> runOnFunction(Function &F) = 0;
738 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
739 the transformation or analysis work of your pass. As usual, a true value should
740 be returned if the function is modified.</p>
744 <!-- _______________________________________________________________________ -->
746 <a name="doFinalization_mod">
747 The <tt>doFinalization(Module &)</tt> method
753 <div class="doc_code"><pre>
754 <b>virtual bool</b> doFinalization(Module &M);
757 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
758 called when the pass framework has finished calling <a
759 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
760 program being compiled.</p>
766 <!-- ======================================================================= -->
768 <a name="LoopPass">The <tt>LoopPass</tt> class </a>
773 <p> All <tt>LoopPass</tt> execute on each loop in the function independent of
774 all of the other loops in the function. <tt>LoopPass</tt> processes loops in
775 loop nest order such that outer most loop is processed last. </p>
777 <p> <tt>LoopPass</tt> subclasses are allowed to update loop nest using
778 <tt>LPPassManager</tt> interface. Implementing a loop pass is usually
779 straightforward. <tt>LoopPass</tt>'s may overload three virtual methods to
780 do their work. All these methods should return true if they modified the
781 program, or false if they didn't. </p>
783 <!-- _______________________________________________________________________ -->
785 <a name="doInitialization_loop">
786 The <tt>doInitialization(Loop *,LPPassManager &)</tt> method
792 <div class="doc_code"><pre>
793 <b>virtual bool</b> doInitialization(Loop *, LPPassManager &LPM);
796 <p>The <tt>doInitialization</tt> method is designed to do simple initialization
797 type of stuff that does not depend on the functions being processed. The
798 <tt>doInitialization</tt> method call is not scheduled to overlap with any
799 other pass executions (thus it should be very fast). LPPassManager
800 interface should be used to access Function or Module level analysis
806 <!-- _______________________________________________________________________ -->
808 <a name="runOnLoop">The <tt>runOnLoop</tt> method</a>
813 <div class="doc_code"><pre>
814 <b>virtual bool</b> runOnLoop(Loop *, LPPassManager &LPM) = 0;
817 <p>The <tt>runOnLoop</tt> method must be implemented by your subclass to do
818 the transformation or analysis work of your pass. As usual, a true value should
819 be returned if the function is modified. <tt>LPPassManager</tt> interface
820 should be used to update loop nest.</p>
824 <!-- _______________________________________________________________________ -->
826 <a name="doFinalization_loop">The <tt>doFinalization()</tt> method</a>
831 <div class="doc_code"><pre>
832 <b>virtual bool</b> doFinalization();
835 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
836 called when the pass framework has finished calling <a
837 href="#runOnLoop"><tt>runOnLoop</tt></a> for every loop in the
838 program being compiled. </p>
844 <!-- ======================================================================= -->
846 <a name="RegionPass">The <tt>RegionPass</tt> class </a>
851 <p> <tt>RegionPass</tt> is similar to <a href="#LoopPass"><tt>LoopPass</tt></a>,
852 but executes on each single entry single exit region in the function.
853 <tt>RegionPass</tt> processes regions in nested order such that the outer most
854 region is processed last. </p>
856 <p> <tt>RegionPass</tt> subclasses are allowed to update the region tree by using
857 the <tt>RGPassManager</tt> interface. You may overload three virtual methods of
858 <tt>RegionPass</tt> to implement your own region pass. All these
859 methods should return true if they modified the program, or false if they didn not.
862 <!-- _______________________________________________________________________ -->
864 <a name="doInitialization_region">
865 The <tt>doInitialization(Region *, RGPassManager &)</tt> method
871 <div class="doc_code"><pre>
872 <b>virtual bool</b> doInitialization(Region *, RGPassManager &RGM);
875 <p>The <tt>doInitialization</tt> method is designed to do simple initialization
876 type of stuff that does not depend on the functions being processed. The
877 <tt>doInitialization</tt> method call is not scheduled to overlap with any
878 other pass executions (thus it should be very fast). RPPassManager
879 interface should be used to access Function or Module level analysis
885 <!-- _______________________________________________________________________ -->
887 <a name="runOnRegion">The <tt>runOnRegion</tt> method</a>
892 <div class="doc_code"><pre>
893 <b>virtual bool</b> runOnRegion(Region *, RGPassManager &RGM) = 0;
896 <p>The <tt>runOnRegion</tt> method must be implemented by your subclass to do
897 the transformation or analysis work of your pass. As usual, a true value should
898 be returned if the region is modified. <tt>RGPassManager</tt> interface
899 should be used to update region tree.</p>
903 <!-- _______________________________________________________________________ -->
905 <a name="doFinalization_region">The <tt>doFinalization()</tt> method</a>
910 <div class="doc_code"><pre>
911 <b>virtual bool</b> doFinalization();
914 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
915 called when the pass framework has finished calling <a
916 href="#runOnRegion"><tt>runOnRegion</tt></a> for every region in the
917 program being compiled. </p>
923 <!-- ======================================================================= -->
925 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
930 <p><tt>BasicBlockPass</tt>'s are just like <a
931 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
932 their scope of inspection and modification to a single basic block at a time.
933 As such, they are <b>not</b> allowed to do any of the following:</p>
936 <li>Modify or inspect any basic blocks outside of the current one</li>
937 <li>Maintain state across invocations of
938 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
939 <li>Modify the control flow graph (by altering terminator instructions)</li>
940 <li>Any of the things forbidden for
941 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
944 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
945 optimizations. They may override the same <a
946 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
947 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
948 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
950 <!-- _______________________________________________________________________ -->
952 <a name="doInitialization_fn">
953 The <tt>doInitialization(Function &)</tt> method
959 <div class="doc_code"><pre>
960 <b>virtual bool</b> doInitialization(Function &F);
963 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
964 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
965 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
966 to do simple initialization that does not depend on the
967 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
968 scheduled to overlap with any other pass executions (thus it should be very
973 <!-- _______________________________________________________________________ -->
975 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
980 <div class="doc_code"><pre>
981 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
984 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
985 function is not allowed to inspect or modify basic blocks other than the
986 parameter, and are not allowed to modify the CFG. A true value must be returned
987 if the basic block is modified.</p>
991 <!-- _______________________________________________________________________ -->
993 <a name="doFinalization_fn">
994 The <tt>doFinalization(Function &)</tt> method
1000 <div class="doc_code"><pre>
1001 <b>virtual bool</b> doFinalization(Function &F);
1004 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
1005 called when the pass framework has finished calling <a
1006 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
1007 program being compiled. This can be used to perform per-function
1014 <!-- ======================================================================= -->
1016 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
1021 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
1022 executes on the machine-dependent representation of each LLVM function in the
1025 <p>Code generator passes are registered and initialized specially by
1026 <tt>TargetMachine::addPassesToEmitFile</tt> and similar routines, so they
1027 cannot generally be run from the <tt>opt</tt> or <tt>bugpoint</tt>
1030 <p>A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
1031 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
1032 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
1033 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
1036 <li>Modify or create any LLVM IR Instructions, BasicBlocks, Arguments,
1037 Functions, GlobalVariables, GlobalAliases, or Modules.</li>
1038 <li>Modify a MachineFunction other than the one currently being processed.</li>
1039 <li>Maintain state across invocations of <a
1040 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
1044 <!-- _______________________________________________________________________ -->
1046 <a name="runOnMachineFunction">
1047 The <tt>runOnMachineFunction(MachineFunction &MF)</tt> method
1053 <div class="doc_code"><pre>
1054 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
1057 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
1058 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
1059 work of your <tt>MachineFunctionPass</tt>.</p>
1061 <p>The <tt>runOnMachineFunction</tt> method is called on every
1062 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
1063 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
1064 representation of the function. If you want to get at the LLVM <tt>Function</tt>
1065 for the <tt>MachineFunction</tt> you're working on, use
1066 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
1067 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
1068 <tt>MachineFunctionPass</tt>.</p>
1076 <!-- *********************************************************************** -->
1078 <a name="registration">Pass registration</a>
1080 <!-- *********************************************************************** -->
1084 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
1085 pass registration works, and discussed some of the reasons that it is used and
1086 what it does. Here we discuss how and why passes are registered.</p>
1088 <p>As we saw above, passes are registered with the <b><tt>RegisterPass</tt></b>
1089 template. The template parameter is the name of the pass that is to be used on
1090 the command line to specify that the pass should be added to a program (for
1091 example, with <tt>opt</tt> or <tt>bugpoint</tt>). The first argument is the
1092 name of the pass, which is to be used for the <tt>-help</tt> output of
1094 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
1096 <p>If you want your pass to be easily dumpable, you should
1097 implement the virtual <tt>print</tt> method:</p>
1099 <!-- _______________________________________________________________________ -->
1101 <a name="print">The <tt>print</tt> method</a>
1106 <div class="doc_code"><pre>
1107 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
1110 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
1111 a human readable version of the analysis results. This is useful for debugging
1112 an analysis itself, as well as for other people to figure out how an analysis
1113 works. Use the <tt>opt -analyze</tt> argument to invoke this method.</p>
1115 <p>The <tt>llvm::OStream</tt> parameter specifies the stream to write the results on,
1116 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
1117 program that has been analyzed. Note however that this pointer may be null in
1118 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
1119 debugger), so it should only be used to enhance debug output, it should not be
1126 <!-- *********************************************************************** -->
1128 <a name="interaction">Specifying interactions between passes</a>
1130 <!-- *********************************************************************** -->
1134 <p>One of the main responsibilities of the <tt>PassManager</tt> is to make sure
1135 that passes interact with each other correctly. Because <tt>PassManager</tt>
1136 tries to <a href="#passmanager">optimize the execution of passes</a> it must
1137 know how the passes interact with each other and what dependencies exist between
1138 the various passes. To track this, each pass can declare the set of passes that
1139 are required to be executed before the current pass, and the passes which are
1140 invalidated by the current pass.</p>
1142 <p>Typically this functionality is used to require that analysis results are
1143 computed before your pass is run. Running arbitrary transformation passes can
1144 invalidate the computed analysis results, which is what the invalidation set
1145 specifies. If a pass does not implement the <tt><a
1146 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
1147 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
1149 <!-- _______________________________________________________________________ -->
1151 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
1156 <div class="doc_code"><pre>
1157 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
1160 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
1161 invalidated sets may be specified for your transformation. The implementation
1162 should fill in the <tt><a
1163 href="http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
1164 object with information about which passes are required and not invalidated. To
1165 do this, a pass may call any of the following methods on the AnalysisUsage
1169 <!-- _______________________________________________________________________ -->
1171 <a name="AU::addRequired">
1172 The <tt>AnalysisUsage::addRequired<></tt>
1173 and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods
1179 If your pass requires a previous pass to be executed (an analysis for example),
1180 it can use one of these methods to arrange for it to be run before your pass.
1181 LLVM has many different types of analyses and passes that can be required,
1182 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
1183 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
1184 be no critical edges in the CFG when your pass has been run.
1188 Some analyses chain to other analyses to do their job. For example, an <a
1189 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
1190 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
1191 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
1192 used instead of the <tt>addRequired</tt> method. This informs the PassManager
1193 that the transitively required pass should be alive as long as the requiring
1198 <!-- _______________________________________________________________________ -->
1200 <a name="AU::addPreserved">
1201 The <tt>AnalysisUsage::addPreserved<></tt> method
1207 One of the jobs of the PassManager is to optimize how and when analyses are run.
1208 In particular, it attempts to avoid recomputing data unless it needs to. For
1209 this reason, passes are allowed to declare that they preserve (i.e., they don't
1210 invalidate) an existing analysis if it's available. For example, a simple
1211 constant folding pass would not modify the CFG, so it can't possibly affect the
1212 results of dominator analysis. By default, all passes are assumed to invalidate
1217 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
1218 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
1219 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
1220 not modify the LLVM program at all (which is true for analyses), and the
1221 <tt>setPreservesCFG</tt> method can be used by transformations that change
1222 instructions in the program but do not modify the CFG or terminator instructions
1223 (note that this property is implicitly set for <a
1224 href="#BasicBlockPass">BasicBlockPass</a>'s).
1228 <tt>addPreserved</tt> is particularly useful for transformations like
1229 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
1230 and dominator related analyses if they exist, so it can preserve them, despite
1231 the fact that it hacks on the CFG.
1235 <!-- _______________________________________________________________________ -->
1237 <a name="AU::examples">
1238 Example implementations of <tt>getAnalysisUsage</tt>
1244 <div class="doc_code"><pre>
1245 <i>// This example modifies the program, but does not modify the CFG</i>
1246 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1247 AU.setPreservesCFG();
1248 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1254 <!-- _______________________________________________________________________ -->
1256 <a name="getAnalysis">
1257 The <tt>getAnalysis<></tt> and
1258 <tt>getAnalysisIfAvailable<></tt> methods
1264 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1265 your class, providing you with access to the passes that you declared that you
1266 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
1267 method. It takes a single template argument that specifies which pass class you
1268 want, and returns a reference to that pass. For example:</p>
1270 <div class="doc_code"><pre>
1271 bool LICM::runOnFunction(Function &F) {
1272 LoopInfo &LI = getAnalysis<LoopInfo>();
1277 <p>This method call returns a reference to the pass desired. You may get a
1278 runtime assertion failure if you attempt to get an analysis that you did not
1279 declare as required in your <a
1280 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
1281 method can be called by your <tt>run*</tt> method implementation, or by any
1282 other local method invoked by your <tt>run*</tt> method.
1284 A module level pass can use function level analysis info using this interface.
1287 <div class="doc_code"><pre>
1288 bool ModuleLevelPass::runOnModule(Module &M) {
1290 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
1295 <p>In above example, runOnFunction for DominatorTree is called by pass manager
1296 before returning a reference to the desired pass.</p>
1299 If your pass is capable of updating analyses if they exist (e.g.,
1300 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1301 <tt>getAnalysisIfAvailable</tt> method, which returns a pointer to the analysis
1302 if it is active. For example:</p>
1304 <div class="doc_code"><pre>
1306 if (DominatorSet *DS = getAnalysisIfAvailable<DominatorSet>()) {
1307 <i>// A DominatorSet is active. This code will update it.</i>
1316 <!-- *********************************************************************** -->
1318 <a name="analysisgroup">Implementing Analysis Groups</a>
1320 <!-- *********************************************************************** -->
1324 <p>Now that we understand the basics of how passes are defined, how they are
1325 used, and how they are required from other passes, it's time to get a little bit
1326 fancier. All of the pass relationships that we have seen so far are very
1327 simple: one pass depends on one other specific pass to be run before it can run.
1328 For many applications, this is great, for others, more flexibility is
1331 <p>In particular, some analyses are defined such that there is a single simple
1332 interface to the analysis results, but multiple ways of calculating them.
1333 Consider alias analysis for example. The most trivial alias analysis returns
1334 "may alias" for any alias query. The most sophisticated analysis a
1335 flow-sensitive, context-sensitive interprocedural analysis that can take a
1336 significant amount of time to execute (and obviously, there is a lot of room
1337 between these two extremes for other implementations). To cleanly support
1338 situations like this, the LLVM Pass Infrastructure supports the notion of
1339 Analysis Groups.</p>
1341 <!-- _______________________________________________________________________ -->
1343 <a name="agconcepts">Analysis Group Concepts</a>
1348 <p>An Analysis Group is a single simple interface that may be implemented by
1349 multiple different passes. Analysis Groups can be given human readable names
1350 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1351 class. An analysis group may have one or more implementations, one of which is
1352 the "default" implementation.</p>
1354 <p>Analysis groups are used by client passes just like other passes are: the
1355 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1356 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
1357 scans the available passes to see if any implementations of the analysis group
1358 are available. If none is available, the default implementation is created for
1359 the pass to use. All standard rules for <A href="#interaction">interaction
1360 between passes</a> still apply.</p>
1362 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
1363 passes, all analysis group implementations must be registered, and must use the
1364 <A href="#registerag"><tt>INITIALIZE_AG_PASS</tt></a> template to join the
1365 implementation pool. Also, a default implementation of the interface
1366 <b>must</b> be registered with <A
1367 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
1369 <p>As a concrete example of an Analysis Group in action, consider the <a
1370 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1371 analysis group. The default implementation of the alias analysis interface (the
1373 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1374 pass) just does a few simple checks that don't require significant analysis to
1375 compute (such as: two different globals can never alias each other, etc).
1376 Passes that use the <tt><a
1377 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1378 interface (for example the <tt><a
1379 href="http://llvm.org/doxygen/structGCSE.html">gcse</a></tt> pass), do
1380 not care which implementation of alias analysis is actually provided, they just
1381 use the designated interface.</p>
1383 <p>From the user's perspective, commands work just like normal. Issuing the
1384 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1385 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1386 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1387 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1388 hypothetical example) instead.</p>
1392 <!-- _______________________________________________________________________ -->
1394 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1399 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1400 group itself, while the <tt>INITIALIZE_AG_PASS</tt> is used to add pass
1401 implementations to the analysis group. First,
1402 an analysis group should be registered, with a human readable name
1404 Unlike registration of passes, there is no command line argument to be specified
1405 for the Analysis Group Interface itself, because it is "abstract":</p>
1407 <div class="doc_code"><pre>
1408 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1411 <p>Once the analysis is registered, passes can declare that they are valid
1412 implementations of the interface by using the following code:</p>
1414 <div class="doc_code"><pre>
1416 //<i> Declare that we implement the AliasAnalysis interface</i>
1417 INITIALIZE_AG_PASS(FancyAA, <a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, "<i>somefancyaa</i>",
1418 "<i>A more complex alias analysis implementation</i>",
1419 false, // <i>Is CFG Only?</i>
1420 true, // <i>Is Analysis?</i>
1421 false); // <i>Is default Analysis Group implementation?</i>
1425 <p>This just shows a class <tt>FancyAA</tt> that
1426 uses the <tt>INITIALIZE_AG_PASS</tt> macro both to register and
1427 to "join" the <tt><a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1428 analysis group. Every implementation of an analysis group should join using
1431 <div class="doc_code"><pre>
1433 //<i> Declare that we implement the AliasAnalysis interface</i>
1434 INITIALIZE_AG_PASS(BasicAA, <a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, "<i>basicaa</i>",
1435 "<i>Basic Alias Analysis (default AA impl)</i>",
1436 false, // <i>Is CFG Only?</i>
1437 true, // <i>Is Analysis?</i>
1438 true); // <i>Is default Analysis Group implementation?</i>
1442 <p>Here we show how the default implementation is specified (using the final
1443 argument to the <tt>INITIALIZE_AG_PASS</tt> template). There must be exactly
1444 one default implementation available at all times for an Analysis Group to be
1445 used. Only default implementation can derive from <tt>ImmutablePass</tt>.
1446 Here we declare that the
1447 <tt><a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1448 pass is the default implementation for the interface.</p>
1454 <!-- *********************************************************************** -->
1456 <a name="passStatistics">Pass Statistics</a>
1458 <!-- *********************************************************************** -->
1462 href="http://llvm.org/doxygen/Statistic_8h-source.html"><tt>Statistic</tt></a>
1463 class is designed to be an easy way to expose various success
1464 metrics from passes. These statistics are printed at the end of a
1465 run, when the -stats command line option is enabled on the command
1466 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1471 <!-- *********************************************************************** -->
1473 <a name="passmanager">What PassManager does</a>
1475 <!-- *********************************************************************** -->
1480 href="http://llvm.org/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1482 href="http://llvm.org/doxygen/classllvm_1_1PassManager.html">class</a>
1483 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1484 are set up correctly, and then schedules passes to run efficiently. All of the
1485 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1488 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1489 time of a series of passes:</p>
1492 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1493 recomputing analysis results as much as possible. This means keeping track of
1494 which analyses are available already, which analyses get invalidated, and which
1495 analyses are needed to be run for a pass. An important part of work is that the
1496 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1497 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1498 results as soon as they are no longer needed.</li>
1500 <li><b>Pipeline the execution of passes on the program</b> - The
1501 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1502 of a series of passes by pipelining the passes together. This means that, given
1503 a series of consecutive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1504 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1505 the first function, then all of the <a
1506 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1507 etc... until the entire program has been run through the passes.
1509 <p>This improves the cache behavior of the compiler, because it is only touching
1510 the LLVM program representation for a single function at a time, instead of
1511 traversing the entire program. It reduces the memory consumption of compiler,
1512 because, for example, only one <a
1513 href="http://llvm.org/doxygen/classllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1514 needs to be calculated at a time. This also makes it possible to implement
1516 href="#SMP">interesting enhancements</a> in the future.</p></li>
1520 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1521 much information it has about the behaviors of the passes it is scheduling. For
1522 example, the "preserved" set is intentionally conservative in the face of an
1523 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1524 Not implementing when it should be implemented will have the effect of not
1525 allowing any analysis results to live across the execution of your pass.</p>
1527 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1528 options that is useful for debugging pass execution, seeing how things work, and
1529 diagnosing when you should be preserving more analyses than you currently are
1530 (To get information about all of the variants of the <tt>--debug-pass</tt>
1531 option, just type '<tt>opt -help-hidden</tt>').</p>
1533 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1534 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1535 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1537 <div class="doc_code"><pre>
1538 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1540 Function Pass Manager
1541 Dominator Set Construction
1542 Immediate Dominators Construction
1543 Global Common Subexpression Elimination
1544 -- Immediate Dominators Construction
1545 -- Global Common Subexpression Elimination
1546 Natural Loop Construction
1547 Loop Invariant Code Motion
1548 -- Natural Loop Construction
1549 -- Loop Invariant Code Motion
1551 -- Dominator Set Construction
1557 <p>This output shows us when passes are constructed and when the analysis
1558 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1559 GCSE uses dominator and immediate dominator information to do its job. The LICM
1560 pass uses natural loop information, which uses dominator sets, but not immediate
1561 dominators. Because immediate dominators are no longer useful after the GCSE
1562 pass, it is immediately destroyed. The dominator sets are then reused to
1563 compute natural loop information, which is then used by the LICM pass.</p>
1565 <p>After the LICM pass, the module verifier runs (which is automatically added
1566 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1567 resultant LLVM code is well formed. After it finishes, the dominator set
1568 information is destroyed, after being computed once, and shared by three
1571 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1572 World</a> pass in between the two passes:</p>
1574 <div class="doc_code"><pre>
1575 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1577 Function Pass Manager
1578 Dominator Set Construction
1579 Immediate Dominators Construction
1580 Global Common Subexpression Elimination
1581 <b>-- Dominator Set Construction</b>
1582 -- Immediate Dominators Construction
1583 -- Global Common Subexpression Elimination
1584 <b> Hello World Pass
1586 Dominator Set Construction</b>
1587 Natural Loop Construction
1588 Loop Invariant Code Motion
1589 -- Natural Loop Construction
1590 -- Loop Invariant Code Motion
1592 -- Dominator Set Construction
1601 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1602 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1603 we need to add the following <a
1604 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1606 <div class="doc_code"><pre>
1607 <i>// We don't modify the program, so we preserve all analyses</i>
1608 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1609 AU.setPreservesAll();
1613 <p>Now when we run our pass, we get this output:</p>
1615 <div class="doc_code"><pre>
1616 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1617 Pass Arguments: -gcse -hello -licm
1619 Function Pass Manager
1620 Dominator Set Construction
1621 Immediate Dominators Construction
1622 Global Common Subexpression Elimination
1623 -- Immediate Dominators Construction
1624 -- Global Common Subexpression Elimination
1627 Natural Loop Construction
1628 Loop Invariant Code Motion
1629 -- Loop Invariant Code Motion
1630 -- Natural Loop Construction
1632 -- Dominator Set Construction
1641 <p>Which shows that we don't accidentally invalidate dominator information
1642 anymore, and therefore do not have to compute it twice.</p>
1644 <!-- _______________________________________________________________________ -->
1646 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1651 <div class="doc_code"><pre>
1652 <b>virtual void</b> releaseMemory();
1655 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1656 results, and how long to keep them around for. Because the lifetime of the pass
1657 object itself is effectively the entire duration of the compilation process, we
1658 need some way to free analysis results when they are no longer useful. The
1659 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1661 <p>If you are writing an analysis or any other pass that retains a significant
1662 amount of state (for use by another pass which "requires" your pass and uses the
1663 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1664 <tt>releaseMemory</tt> to, well, release the memory allocated to maintain this
1665 internal state. This method is called after the <tt>run*</tt> method for the
1666 class, before the next call of <tt>run*</tt> in your pass.</p>
1672 <!-- *********************************************************************** -->
1674 <a name="registering">Registering dynamically loaded passes</a>
1676 <!-- *********************************************************************** -->
1680 <p><i>Size matters</i> when constructing production quality tools using llvm,
1681 both for the purposes of distribution, and for regulating the resident code size
1682 when running on the target system. Therefore, it becomes desirable to
1683 selectively use some passes, while omitting others and maintain the flexibility
1684 to change configurations later on. You want to be able to do all this, and,
1685 provide feedback to the user. This is where pass registration comes into
1688 <p>The fundamental mechanisms for pass registration are the
1689 <tt>MachinePassRegistry</tt> class and subclasses of
1690 <tt>MachinePassRegistryNode</tt>.</p>
1692 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1693 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1694 communicates additions and deletions to the command line interface.</p>
1696 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1697 information provided about a particular pass. This information includes the
1698 command line name, the command help string and the address of the function used
1699 to create an instance of the pass. A global static constructor of one of these
1700 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1701 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1702 in the tool will be registered at start up. A dynamically loaded pass will
1703 register on load and unregister at unload.</p>
1705 <!-- _______________________________________________________________________ -->
1707 <a name="registering_existing">Using existing registries</a>
1712 <p>There are predefined registries to track instruction scheduling
1713 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1714 machine passes. Here we will describe how to <i>register</i> a register
1715 allocator machine pass.</p>
1717 <p>Implement your register allocator machine pass. In your register allocator
1718 <tt>.cpp</tt> file add the following include;</p>
1720 <div class="doc_code"><pre>
1721 #include "llvm/CodeGen/RegAllocRegistry.h"
1724 <p>Also in your register allocator .cpp file, define a creator function in the
1727 <div class="doc_code"><pre>
1728 FunctionPass *createMyRegisterAllocator() {
1729 return new MyRegisterAllocator();
1733 <p>Note that the signature of this function should match the type of
1734 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1735 "installing" declaration, in the form;</p>
1737 <div class="doc_code"><pre>
1738 static RegisterRegAlloc myRegAlloc("myregalloc",
1739 "my register allocator help string",
1740 createMyRegisterAllocator);
1743 <p>Note the two spaces prior to the help string produces a tidy result on the
1746 <div class="doc_code"><pre>
1749 -regalloc - Register allocator to use (default=linearscan)
1750 =linearscan - linear scan register allocator
1751 =local - local register allocator
1752 =simple - simple register allocator
1753 =myregalloc - my register allocator help string
1757 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1758 an option. Registering instruction schedulers is similar except use the
1759 <tt>RegisterScheduler</tt> class. Note that the
1760 <tt>RegisterScheduler::FunctionPassCtor</tt> is significantly different from
1761 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1763 <p>To force the load/linking of your register allocator into the llc/lli tools,
1764 add your creator function's global declaration to "Passes.h" and add a "pseudo"
1765 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1770 <!-- _______________________________________________________________________ -->
1772 <a name="registering_new">Creating new registries</a>
1777 <p>The easiest way to get started is to clone one of the existing registries; we
1778 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1779 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1781 <p>Then you need to declare the registry. Example: if your pass registry is
1782 <tt>RegisterMyPasses</tt> then define;</p>
1784 <div class="doc_code"><pre>
1785 MachinePassRegistry RegisterMyPasses::Registry;
1788 <p>And finally, declare the command line option for your passes. Example:</p>
1790 <div class="doc_code"><pre>
1791 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1792 RegisterPassParser<RegisterMyPasses> >
1794 cl::init(&createDefaultMyPass),
1795 cl::desc("my pass option help"));
1798 <p>Here the command option is "mypass", with createDefaultMyPass as the default
1805 <!-- *********************************************************************** -->
1807 <a name="debughints">Using GDB with dynamically loaded passes</a>
1809 <!-- *********************************************************************** -->
1813 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1814 should be. First of all, you can't set a breakpoint in a shared object that has
1815 not been loaded yet, and second of all there are problems with inlined functions
1816 in shared objects. Here are some suggestions to debugging your pass with
1819 <p>For sake of discussion, I'm going to assume that you are debugging a
1820 transformation invoked by <tt>opt</tt>, although nothing described here depends
1823 <!-- _______________________________________________________________________ -->
1825 <a name="breakpoint">Setting a breakpoint in your pass</a>
1830 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1832 <div class="doc_code"><pre>
1835 Copyright 2000 Free Software Foundation, Inc.
1836 GDB is free software, covered by the GNU General Public License, and you are
1837 welcome to change it and/or distribute copies of it under certain conditions.
1838 Type "show copying" to see the conditions.
1839 There is absolutely no warranty for GDB. Type "show warranty" for details.
1840 This GDB was configured as "sparc-sun-solaris2.6"...
1844 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1845 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1846 (the shared object isn't loaded until runtime), we must execute the process, and
1847 have it stop before it invokes our pass, but after it has loaded the shared
1848 object. The most foolproof way of doing this is to set a breakpoint in
1849 <tt>PassManager::run</tt> and then run the process with the arguments you
1852 <div class="doc_code"><pre>
1853 (gdb) <b>break llvm::PassManager::run</b>
1854 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1855 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]</b>
1856 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1857 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1858 70 bool PassManager::run(Module &M) { return PM->run(M); }
1862 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1863 now free to set breakpoints in your pass so that you can trace through execution
1864 or do other standard debugging stuff.</p>
1868 <!-- _______________________________________________________________________ -->
1870 <a name="debugmisc">Miscellaneous Problems</a>
1875 <p>Once you have the basics down, there are a couple of problems that GDB has,
1876 some with solutions, some without.</p>
1879 <li>Inline functions have bogus stack information. In general, GDB does a
1880 pretty good job getting stack traces and stepping through inline functions.
1881 When a pass is dynamically loaded however, it somehow completely loses this
1882 capability. The only solution I know of is to de-inline a function (move it
1883 from the body of a class to a .cpp file).</li>
1885 <li>Restarting the program breaks breakpoints. After following the information
1886 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1887 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1888 and you start getting errors about breakpoints being unsettable. The only way I
1889 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1890 already set in your pass, run the program, and re-set the breakpoints once
1891 execution stops in <tt>PassManager::run</tt>.</li>
1895 <p>Hopefully these tips will help with common case debugging situations. If
1896 you'd like to contribute some tips of your own, just contact <a
1897 href="mailto:sabre@nondot.org">Chris</a>.</p>
1903 <!-- *********************************************************************** -->
1905 <a name="future">Future extensions planned</a>
1907 <!-- *********************************************************************** -->
1911 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1912 some nifty stuff, there are things we'd like to add in the future. Here is
1913 where we are going:</p>
1915 <!-- _______________________________________________________________________ -->
1917 <a name="SMP">Multithreaded LLVM</a>
1922 <p>Multiple CPU machines are becoming more common and compilation can never be
1923 fast enough: obviously we should allow for a multithreaded compiler. Because of
1924 the semantics defined for passes above (specifically they cannot maintain state
1925 across invocations of their <tt>run*</tt> methods), a nice clean way to
1926 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1927 to create multiple instances of each pass object, and allow the separate
1928 instances to be hacking on different parts of the program at the same time.</p>
1930 <p>This implementation would prevent each of the passes from having to implement
1931 multithreaded constructs, requiring only the LLVM core to have locking in a few
1932 places (for global resources). Although this is a simple extension, we simply
1933 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1934 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1940 <!-- *********************************************************************** -->
1943 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1944 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
1945 <a href="http://validator.w3.org/check/referer"><img
1946 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
1948 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1949 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a><br>
1950 Last modified: $Date$