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 <tt>getAnalysisToUpdate<></tt> methods</a></li>
83 <li><a href="#analysisgroup">Implementing Analysis Groups</a>
85 <li><a href="#agconcepts">Analysis Group Concepts</a></li>
86 <li><a href="#registerag">Using <tt>RegisterAnalysisGroup</tt></a></li>
88 <li><a href="#passStatistics">Pass Statistics</a>
89 <li><a href="#passmanager">What PassManager does</a>
91 <li><a href="#releaseMemory">The <tt>releaseMemory</tt> method</a></li>
93 <li><a href="#registering">Registering dynamically loaded passes</a>
95 <li><a href="#registering_existing">Using existing registries</a></li>
96 <li><a href="#registering_new">Creating new registries</a></li>
98 <li><a href="#debughints">Using GDB with dynamically loaded passes</a>
100 <li><a href="#breakpoint">Setting a breakpoint in your pass</a></li>
101 <li><a href="#debugmisc">Miscellaneous Problems</a></li>
103 <li><a href="#future">Future extensions planned</a>
105 <li><a href="#SMP">Multithreaded LLVM</a></li>
109 <div class="doc_author">
110 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
111 <a href="mailto:jlaskey@mac.com">Jim Laskey</a></p>
114 <!-- *********************************************************************** -->
115 <div class="doc_section">
116 <a name="introduction">Introduction - What is a pass?</a>
118 <!-- *********************************************************************** -->
120 <div class="doc_text">
122 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
123 passes are where most of the interesting parts of the compiler exist. Passes
124 perform the transformations and optimizations that make up the compiler, they
125 build the analysis results that are used by these transformations, and they are,
126 above all, a structuring technique for compiler code.</p>
128 <p>All LLVM passes are subclasses of the <tt><a
129 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
130 class, which implement functionality by overriding virtual methods inherited
131 from <tt>Pass</tt>. Depending on how your pass works, you should inherit from
132 the <tt><a href="#ModulePass">ModulePass</a></tt>, <tt><a
133 href="#CallGraphSCCPass">CallGraphSCCPass</a></tt>, <tt><a
134 href="#FunctionPass">FunctionPass</a></tt>, or <tt><a
135 href="#LoopPass">LoopPass</a></tt>, or <tt><a
136 href="#BasicBlockPass">BasicBlockPass</a></tt> classes, which gives the system
137 more information about what your pass does, and how it can be combined with
138 other passes. One of the main features of the LLVM Pass Framework is that it
139 schedules passes to run in an efficient way based on the constraints that your
140 pass meets (which are indicated by which class they derive from).</p>
142 <p>We start by showing you how to construct a pass, everything from setting up
143 the code, to compiling, loading, and executing it. After the basics are down,
144 more advanced features are discussed.</p>
148 <!-- *********************************************************************** -->
149 <div class="doc_section">
150 <a name="quickstart">Quick Start - Writing hello world</a>
152 <!-- *********************************************************************** -->
154 <div class="doc_text">
156 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
157 is designed to simply print out the name of non-external functions that exist in
158 the program being compiled. It does not modify the program at all, it just
159 inspects it. The source code and files for this pass are available in the LLVM
160 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
164 <!-- ======================================================================= -->
165 <div class="doc_subsection">
166 <a name="makefile">Setting up the build environment</a>
169 <div class="doc_text">
171 <p>First, you need to create a new directory somewhere in the LLVM source
172 base. For this example, we'll assume that you made
173 <tt>lib/Transforms/Hello</tt>. Next, you must set up a build script
174 (Makefile) that will compile the source code for the new pass. To do this,
175 copy the following into <tt>Makefile</tt>:</p>
178 <div class="doc_code"><pre>
179 # Makefile for hello pass
181 # Path to top level of LLVM heirarchy
184 # Name of the library to build
187 # Make the shared library become a loadable module so the tools can
188 # dlopen/dlsym on the resulting library.
191 # Tell the build system which LLVM libraries your pass needs. You'll probably
192 # need at least LLVMSystem.a, LLVMSupport.a, LLVMCore.a but possibly several
194 LLVMLIBS = LLVMCore.a LLVMSupport.a LLVMSystem.a
196 # Include the makefile implementation stuff
197 include $(LEVEL)/Makefile.common
200 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
201 directory are to be compiled and linked together into a
202 <tt>Debug/lib/Hello.so</tt> shared object that can be dynamically loaded by
203 the <tt>opt</tt> or <tt>bugpoint</tt> tools via their <tt>-load</tt> options.
204 If your operating system uses a suffix other than .so (such as windows or
205 Mac OS/X), the appropriate extension will be used.</p>
207 <p>Now that we have the build scripts set up, we just need to write the code for
212 <!-- ======================================================================= -->
213 <div class="doc_subsection">
214 <a name="basiccode">Basic code required</a>
217 <div class="doc_text">
219 <p>Now that we have a way to compile our new pass, we just have to write it.
222 <div class="doc_code"><pre>
223 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
224 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
227 <p>Which are needed because we are writing a <tt><a
228 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>, and
229 we are operating on <tt><a
230 href="http://llvm.org/doxygen/classllvm_1_1Function.html">Function</a></tt>'s.</p>
233 <div class="doc_code"><pre>
234 <b>using namespace llvm;</b>
236 <p>... which is required because the functions from the include files
237 live in the llvm namespace.
242 <div class="doc_code"><pre>
246 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
247 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
248 things declared inside of the anonymous namespace only visible to the current
249 file. If you're not familiar with them, consult a decent C++ book for more
252 <p>Next, we declare our pass itself:</p>
254 <div class="doc_code"><pre>
255 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
258 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
259 href="http://llvm.org/doxygen/classllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
260 The different builtin pass subclasses are described in detail <a
261 href="#passtype">later</a>, but for now, know that <a
262 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate a function at a
265 <div class="doc_code"><pre>
267 Hello() : FunctionPass((intptr_t)&ID) {}
270 <p> This declares pass identifier used by LLVM to identify pass. This allows LLVM to
271 avoid using expensive C++ runtime information.</p>
273 <div class="doc_code"><pre>
274 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
275 llvm::cerr << "<i>Hello: </i>" << F.getName() << "\n";
278 }; <i>// end of struct Hello</i>
281 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
282 which overloads an abstract virtual method inherited from <a
283 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
284 to do our thing, so we just print out our message with the name of each
287 <div class="doc_code"><pre>
291 <p> We initialize pass ID here. LLVM uses ID's address to identify pass so
292 initialization value is not important.</p>
294 <div class="doc_code"><pre>
295 RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>",
296 false /* Only looks at CFG */,
297 false /* Analysis Pass */);
298 } <i>// end of anonymous namespace</i>
301 <p>Lastly, we <a href="#registration">register our class</a> <tt>Hello</tt>,
302 giving it a command line
303 argument "<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>".
304 Last two RegisterPass arguments are optional. Their default value is false.
305 If a pass walks CFG without modifying it then third argument is set to true.
306 If a pass is an analysis pass, for example dominator tree pass, then true
307 is supplied as fourth argument. </p>
309 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
311 <div class="doc_code"><pre>
312 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
313 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
315 <b>using namespace llvm;</b>
318 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
321 Hello() : FunctionPass((intptr_t)&ID) {}
323 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
324 llvm::cerr << "<i>Hello: </i>" << F.getName() << "\n";
330 RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
334 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
335 command in the local directory and you should get a new
336 "<tt>Debug/lib/Hello.so</tt> file. Note that everything in this file is
337 contained in an anonymous namespace: this reflects the fact that passes are self
338 contained units that do not need external interfaces (although they can have
339 them) to be useful.</p>
343 <!-- ======================================================================= -->
344 <div class="doc_subsection">
345 <a name="running">Running a pass with <tt>opt</tt></a>
348 <div class="doc_text">
350 <p>Now that you have a brand new shiny shared object file, we can use the
351 <tt>opt</tt> command to run an LLVM program through your pass. Because you
352 registered your pass with the <tt>RegisterPass</tt> template, you will be able to
353 use the <tt>opt</tt> tool to access it, once loaded.</p>
355 <p>To test it, follow the example at the end of the <a
356 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
357 LLVM. We can now run the bitcode file (<tt>hello.bc</tt>) for the program
358 through our transformation like this (or course, any bitcode file will
361 <div class="doc_code"><pre>
362 $ opt -load ../../../Debug/lib/Hello.so -hello < hello.bc > /dev/null
368 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
369 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
370 argument (which is one reason you need to <a href="#registration">register your
371 pass</a>). Because the hello pass does not modify the program in any
372 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
373 <tt>/dev/null</tt>).</p>
375 <p>To see what happened to the other string you registered, try running
376 <tt>opt</tt> with the <tt>--help</tt> option:</p>
378 <div class="doc_code"><pre>
379 $ opt -load ../../../Debug/lib/Hello.so --help
380 OVERVIEW: llvm .bc -> .bc modular optimizer
382 USAGE: opt [options] <input bitcode>
385 Optimizations available:
387 -funcresolve - Resolve Functions
388 -gcse - Global Common Subexpression Elimination
389 -globaldce - Dead Global Elimination
390 <b>-hello - Hello World Pass</b>
391 -indvars - Canonicalize Induction Variables
392 -inline - Function Integration/Inlining
393 -instcombine - Combine redundant instructions
397 <p>The pass name get added as the information string for your pass, giving some
398 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
399 would go ahead and make it do the cool transformations you want. Once you get
400 it all working and tested, it may become useful to find out how fast your pass
401 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
402 line option (<tt>--time-passes</tt>) that allows you to get information about
403 the execution time of your pass along with the other passes you queue up. For
406 <div class="doc_code"><pre>
407 $ opt -load ../../../Debug/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
411 ===============================================================================
412 ... Pass execution timing report ...
413 ===============================================================================
414 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
416 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
417 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bitcode Writer
418 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
419 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
420 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
421 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
424 <p>As you can see, our implementation above is pretty fast :). The additional
425 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
426 that the LLVM emitted by your pass is still valid and well formed LLVM, which
427 hasn't been broken somehow.</p>
429 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
430 about some more details of how they work and how to use them.</p>
434 <!-- *********************************************************************** -->
435 <div class="doc_section">
436 <a name="passtype">Pass classes and requirements</a>
438 <!-- *********************************************************************** -->
440 <div class="doc_text">
442 <p>One of the first things that you should do when designing a new pass is to
443 decide what class you should subclass for your pass. The <a
444 href="#basiccode">Hello World</a> example uses the <tt><a
445 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
446 did not discuss why or when this should occur. Here we talk about the classes
447 available, from the most general to the most specific.</p>
449 <p>When choosing a superclass for your Pass, you should choose the <b>most
450 specific</b> class possible, while still being able to meet the requirements
451 listed. This gives the LLVM Pass Infrastructure information necessary to
452 optimize how passes are run, so that the resultant compiler isn't unneccesarily
457 <!-- ======================================================================= -->
458 <div class="doc_subsection">
459 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
462 <div class="doc_text">
464 <p>The most plain and boring type of pass is the "<tt><a
465 href="http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
466 class. This pass type is used for passes that do not have to be run, do not
467 change state, and never need to be updated. This is not a normal type of
468 transformation or analysis, but can provide information about the current
469 compiler configuration.</p>
471 <p>Although this pass class is very infrequently used, it is important for
472 providing information about the current target machine being compiled for, and
473 other static information that can affect the various transformations.</p>
475 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
476 invalidated, and are never "run".</p>
480 <!-- ======================================================================= -->
481 <div class="doc_subsection">
482 <a name="ModulePass">The <tt>ModulePass</tt> class</a>
485 <div class="doc_text">
488 href="http://llvm.org/doxygen/classllvm_1_1ModulePass.html">ModulePass</a></tt>"
489 class is the most general of all superclasses that you can use. Deriving from
490 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
491 refering to function bodies in no predictable order, or adding and removing
492 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
493 subclasses, no optimization can be done for their execution. A module pass
494 can use function level passes (e.g. dominators) using getAnalysis interface
495 <tt> getAnalysis<DominatorTree>(Function)</tt>, if the function pass
496 does not require any module passes. </p>
498 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
499 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
500 following signature:</p>
504 <!-- _______________________________________________________________________ -->
505 <div class="doc_subsubsection">
506 <a name="runOnModule">The <tt>runOnModule</tt> method</a>
509 <div class="doc_text">
511 <div class="doc_code"><pre>
512 <b>virtual bool</b> runOnModule(Module &M) = 0;
515 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
516 It should return true if the module was modified by the transformation and
521 <!-- ======================================================================= -->
522 <div class="doc_subsection">
523 <a name="CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
526 <div class="doc_text">
529 href="http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html">CallGraphSCCPass</a></tt>"
530 is used by passes that need to traverse the program bottom-up on the call graph
531 (callees before callers). Deriving from CallGraphSCCPass provides some
532 mechanics for building and traversing the CallGraph, but also allows the system
533 to optimize execution of CallGraphSCCPass's. If your pass meets the
534 requirements outlined below, and doesn't meet the requirements of a <tt><a
535 href="#FunctionPass">FunctionPass</a></tt> or <tt><a
536 href="#BasicBlockPass">BasicBlockPass</a></tt>, you should derive from
537 <tt>CallGraphSCCPass</tt>.</p>
539 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
541 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
545 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
546 the current SCC.</li>
548 <li>... <em>not allowed</em> to inspect any Function's other than those in the
549 current SCC and the direct callees of the SCC.</li>
551 <li>... <em>required</em> to preserve the current CallGraph object, updating it
552 to reflect any changes made to the program.</li>
554 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
555 though they may change the contents of an SCC.</li>
557 <li>... <em>allowed</em> to add or remove global variables from the current
560 <li>... <em>allowed</em> to maintain state across invocations of
561 <a href="#runOnSCC"><tt>runOnSCC</tt></a> (including global data).</li>
564 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
565 because it has to handle SCCs with more than one node in it. All of the virtual
566 methods described below should return true if they modified the program, or
567 false if they didn't.</p>
571 <!-- _______________________________________________________________________ -->
572 <div class="doc_subsubsection">
573 <a name="doInitialization_scc">The <tt>doInitialization(CallGraph &)</tt>
577 <div class="doc_text">
579 <div class="doc_code"><pre>
580 <b>virtual bool</b> doInitialization(CallGraph &CG);
583 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
584 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
585 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
586 is designed to do simple initialization type of stuff that does not depend on
587 the SCCs being processed. The <tt>doInitialization</tt> method call is not
588 scheduled to overlap with any other pass executions (thus it should be very
593 <!-- _______________________________________________________________________ -->
594 <div class="doc_subsubsection">
595 <a name="runOnSCC">The <tt>runOnSCC</tt> method</a>
598 <div class="doc_text">
600 <div class="doc_code"><pre>
601 <b>virtual bool</b> runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;
604 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
605 should return true if the module was modified by the transformation, false
610 <!-- _______________________________________________________________________ -->
611 <div class="doc_subsubsection">
612 <a name="doFinalization_scc">The <tt>doFinalization(CallGraph
613 &)</tt> method</a>
616 <div class="doc_text">
618 <div class="doc_code"><pre>
619 <b>virtual bool</b> doFinalization(CallGraph &CG);
622 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
623 called when the pass framework has finished calling <a
624 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
625 program being compiled.</p>
629 <!-- ======================================================================= -->
630 <div class="doc_subsection">
631 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
634 <div class="doc_text">
636 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
637 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
638 subclasses do have a predictable, local behavior that can be expected by the
639 system. All <tt>FunctionPass</tt> execute on each function in the program
640 independent of all of the other functions in the program.
641 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
642 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
644 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
647 <li>Modify a Function other than the one currently being processed.</li>
648 <li>Add or remove Function's from the current Module.</li>
649 <li>Add or remove global variables from the current Module.</li>
650 <li>Maintain state across invocations of
651 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
654 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
655 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
656 may overload three virtual methods to do their work. All of these methods
657 should return true if they modified the program, or false if they didn't.</p>
661 <!-- _______________________________________________________________________ -->
662 <div class="doc_subsubsection">
663 <a name="doInitialization_mod">The <tt>doInitialization(Module &)</tt>
667 <div class="doc_text">
669 <div class="doc_code"><pre>
670 <b>virtual bool</b> doInitialization(Module &M);
673 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
674 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
675 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
676 is designed to do simple initialization type of stuff that does not depend on
677 the functions being processed. The <tt>doInitialization</tt> method call is not
678 scheduled to overlap with any other pass executions (thus it should be very
681 <p>A good example of how this method should be used is the <a
682 href="http://llvm.org/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
683 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
684 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
685 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
686 free functions that it needs, adding prototypes to the module if necessary.</p>
690 <!-- _______________________________________________________________________ -->
691 <div class="doc_subsubsection">
692 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
695 <div class="doc_text">
697 <div class="doc_code"><pre>
698 <b>virtual bool</b> runOnFunction(Function &F) = 0;
701 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
702 the transformation or analysis work of your pass. As usual, a true value should
703 be returned if the function is modified.</p>
707 <!-- _______________________________________________________________________ -->
708 <div class="doc_subsubsection">
709 <a name="doFinalization_mod">The <tt>doFinalization(Module
710 &)</tt> method</a>
713 <div class="doc_text">
715 <div class="doc_code"><pre>
716 <b>virtual bool</b> doFinalization(Module &M);
719 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
720 called when the pass framework has finished calling <a
721 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
722 program being compiled.</p>
726 <!-- ======================================================================= -->
727 <div class="doc_subsection">
728 <a name="LoopPass">The <tt>LoopPass</tt> class </a>
731 <div class="doc_text">
733 <p> All <tt>LoopPass</tt> execute on each loop in the function independent of
734 all of the other loops in the function. <tt>LoopPass</tt> processes loops in
735 loop nest order such that outer most loop is processed last. </p>
737 <p> <tt>LoopPass</tt> subclasses are allowed to update loop nest using
738 <tt>LPPassManager</tt> interface. Implementing a loop pass is usually
739 straightforward. <tt>Looppass</tt>'s may overload three virtual methods to
740 do their work. All these methods should return true if they modified the
741 program, or false if they didn't. </p>
744 <!-- _______________________________________________________________________ -->
745 <div class="doc_subsubsection">
746 <a name="doInitialization_loop">The <tt>doInitialization(Loop *,
747 LPPassManager &)</tt>
751 <div class="doc_text">
753 <div class="doc_code"><pre>
754 <b>virtual bool</b> doInitialization(Loop *, LPPassManager &LPM);
757 <p>The <tt>doInitialization</tt> method is designed to do simple initialization
758 type of stuff that does not depend on the functions being processed. The
759 <tt>doInitialization</tt> method call is not scheduled to overlap with any
760 other pass executions (thus it should be very fast). LPPassManager
761 interface should be used to access Function or Module level analysis
767 <!-- _______________________________________________________________________ -->
768 <div class="doc_subsubsection">
769 <a name="runOnLoop">The <tt>runOnLoop</tt> method</a>
772 <div class="doc_text">
774 <div class="doc_code"><pre>
775 <b>virtual bool</b> runOnLoop(Loop *, LPPassManager &LPM) = 0;
778 <p>The <tt>runOnLoop</tt> method must be implemented by your subclass to do
779 the transformation or analysis work of your pass. As usual, a true value should
780 be returned if the function is modified. <tt>LPPassManager</tt> interface
781 should be used to update loop nest.</p>
785 <!-- _______________________________________________________________________ -->
786 <div class="doc_subsubsection">
787 <a name="doFinalization_loop">The <tt>doFinalization()</tt> method</a>
790 <div class="doc_text">
792 <div class="doc_code"><pre>
793 <b>virtual bool</b> doFinalization();
796 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
797 called when the pass framework has finished calling <a
798 href="#runOnLoop"><tt>runOnLoop</tt></a> for every loop in the
799 program being compiled. </p>
805 <!-- ======================================================================= -->
806 <div class="doc_subsection">
807 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
810 <div class="doc_text">
812 <p><tt>BasicBlockPass</tt>'s are just like <a
813 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
814 their scope of inspection and modification to a single basic block at a time.
815 As such, they are <b>not</b> allowed to do any of the following:</p>
818 <li>Modify or inspect any basic blocks outside of the current one</li>
819 <li>Maintain state across invocations of
820 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
821 <li>Modify the control flow graph (by altering terminator instructions)</li>
822 <li>Any of the things forbidden for
823 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
826 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
827 optimizations. They may override the same <a
828 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
829 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
830 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
834 <!-- _______________________________________________________________________ -->
835 <div class="doc_subsubsection">
836 <a name="doInitialization_fn">The <tt>doInitialization(Function
837 &)</tt> method</a>
840 <div class="doc_text">
842 <div class="doc_code"><pre>
843 <b>virtual bool</b> doInitialization(Function &F);
846 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
847 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
848 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
849 to do simple initialization that does not depend on the
850 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
851 scheduled to overlap with any other pass executions (thus it should be very
856 <!-- _______________________________________________________________________ -->
857 <div class="doc_subsubsection">
858 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
861 <div class="doc_text">
863 <div class="doc_code"><pre>
864 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
867 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
868 function is not allowed to inspect or modify basic blocks other than the
869 parameter, and are not allowed to modify the CFG. A true value must be returned
870 if the basic block is modified.</p>
874 <!-- _______________________________________________________________________ -->
875 <div class="doc_subsubsection">
876 <a name="doFinalization_fn">The <tt>doFinalization(Function &)</tt>
880 <div class="doc_text">
882 <div class="doc_code"><pre>
883 <b>virtual bool</b> doFinalization(Function &F);
886 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
887 called when the pass framework has finished calling <a
888 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
889 program being compiled. This can be used to perform per-function
894 <!-- ======================================================================= -->
895 <div class="doc_subsection">
896 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
899 <div class="doc_text">
901 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
902 executes on the machine-dependent representation of each LLVM function in the
903 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
904 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
905 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
906 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
909 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
910 <li>Modify a MachineFunction other than the one currently being processed.</li>
911 <li>Add or remove MachineFunctions from the current Module.</li>
912 <li>Add or remove global variables from the current Module.</li>
913 <li>Maintain state across invocations of <a
914 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
920 <!-- _______________________________________________________________________ -->
921 <div class="doc_subsubsection">
922 <a name="runOnMachineFunction">The <tt>runOnMachineFunction(MachineFunction
923 &MF)</tt> method</a>
926 <div class="doc_text">
928 <div class="doc_code"><pre>
929 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
932 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
933 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
934 work of your <tt>MachineFunctionPass</tt>.</p>
936 <p>The <tt>runOnMachineFunction</tt> method is called on every
937 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
938 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
939 representation of the function. If you want to get at the LLVM <tt>Function</tt>
940 for the <tt>MachineFunction</tt> you're working on, use
941 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
942 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
943 <tt>MachineFunctionPass</tt>.</p>
947 <!-- *********************************************************************** -->
948 <div class="doc_section">
949 <a name="registration">Pass registration</a>
951 <!-- *********************************************************************** -->
953 <div class="doc_text">
955 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
956 pass registration works, and discussed some of the reasons that it is used and
957 what it does. Here we discuss how and why passes are registered.</p>
959 <p>As we saw above, passes are registered with the <b><tt>RegisterPass</tt></b>
960 template, which requires you to pass at least two
961 parameters. The first parameter is the name of the pass that is to be used on
962 the command line to specify that the pass should be added to a program (for
963 example, with <tt>opt</tt> or <tt>bugpoint</tt>). The second argument is the
964 name of the pass, which is to be used for the <tt>--help</tt> output of
966 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
968 <p>If you want your pass to be easily dumpable, you should
969 implement the virtual <tt>print</tt> method:</p>
973 <!-- _______________________________________________________________________ -->
974 <div class="doc_subsubsection">
975 <a name="print">The <tt>print</tt> method</a>
978 <div class="doc_text">
980 <div class="doc_code"><pre>
981 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
984 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
985 a human readable version of the analysis results. This is useful for debugging
986 an analysis itself, as well as for other people to figure out how an analysis
987 works. Use the <tt>opt -analyze</tt> argument to invoke this method.</p>
989 <p>The <tt>llvm::OStream</tt> parameter specifies the stream to write the results on,
990 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
991 program that has been analyzed. Note however that this pointer may be null in
992 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
993 debugger), so it should only be used to enhance debug output, it should not be
998 <!-- *********************************************************************** -->
999 <div class="doc_section">
1000 <a name="interaction">Specifying interactions between passes</a>
1002 <!-- *********************************************************************** -->
1004 <div class="doc_text">
1006 <p>One of the main responsibilities of the <tt>PassManager</tt> is to make sure
1007 that passes interact with each other correctly. Because <tt>PassManager</tt>
1008 tries to <a href="#passmanager">optimize the execution of passes</a> it must
1009 know how the passes interact with each other and what dependencies exist between
1010 the various passes. To track this, each pass can declare the set of passes that
1011 are required to be executed before the current pass, and the passes which are
1012 invalidated by the current pass.</p>
1014 <p>Typically this functionality is used to require that analysis results are
1015 computed before your pass is run. Running arbitrary transformation passes can
1016 invalidate the computed analysis results, which is what the invalidation set
1017 specifies. If a pass does not implement the <tt><a
1018 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
1019 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
1023 <!-- _______________________________________________________________________ -->
1024 <div class="doc_subsubsection">
1025 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
1028 <div class="doc_text">
1030 <div class="doc_code"><pre>
1031 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
1034 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
1035 invalidated sets may be specified for your transformation. The implementation
1036 should fill in the <tt><a
1037 href="http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
1038 object with information about which passes are required and not invalidated. To
1039 do this, a pass may call any of the following methods on the AnalysisUsage
1043 <!-- _______________________________________________________________________ -->
1044 <div class="doc_subsubsection">
1045 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
1048 <div class="doc_text">
1050 If your pass requires a previous pass to be executed (an analysis for example),
1051 it can use one of these methods to arrange for it to be run before your pass.
1052 LLVM has many different types of analyses and passes that can be required,
1053 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
1054 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
1055 be no critical edges in the CFG when your pass has been run.
1059 Some analyses chain to other analyses to do their job. For example, an <a
1060 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
1061 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
1062 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
1063 used instead of the <tt>addRequired</tt> method. This informs the PassManager
1064 that the transitively required pass should be alive as long as the requiring
1069 <!-- _______________________________________________________________________ -->
1070 <div class="doc_subsubsection">
1071 <a name="AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a>
1074 <div class="doc_text">
1076 One of the jobs of the PassManager is to optimize how and when analyses are run.
1077 In particular, it attempts to avoid recomputing data unless it needs to. For
1078 this reason, passes are allowed to declare that they preserve (i.e., they don't
1079 invalidate) an existing analysis if it's available. For example, a simple
1080 constant folding pass would not modify the CFG, so it can't possibly affect the
1081 results of dominator analysis. By default, all passes are assumed to invalidate
1086 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
1087 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
1088 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
1089 not modify the LLVM program at all (which is true for analyses), and the
1090 <tt>setPreservesCFG</tt> method can be used by transformations that change
1091 instructions in the program but do not modify the CFG or terminator instructions
1092 (note that this property is implicitly set for <a
1093 href="#BasicBlockPass">BasicBlockPass</a>'s).
1097 <tt>addPreserved</tt> is particularly useful for transformations like
1098 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
1099 and dominator related analyses if they exist, so it can preserve them, despite
1100 the fact that it hacks on the CFG.
1104 <!-- _______________________________________________________________________ -->
1105 <div class="doc_subsubsection">
1106 <a name="AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a>
1109 <div class="doc_text">
1111 <div class="doc_code"><pre>
1112 <i>// This is an example implementation from an analysis, which does not modify
1113 // the program at all, yet has a prerequisite.</i>
1114 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1115 AU.setPreservesAll();
1116 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1122 <div class="doc_code"><pre>
1123 <i>// This example modifies the program, but does not modify the CFG</i>
1124 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1125 AU.setPreservesCFG();
1126 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1132 <!-- _______________________________________________________________________ -->
1133 <div class="doc_subsubsection">
1134 <a name="getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a>
1137 <div class="doc_text">
1139 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1140 your class, providing you with access to the passes that you declared that you
1141 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
1142 method. It takes a single template argument that specifies which pass class you
1143 want, and returns a reference to that pass. For example:</p>
1145 <div class="doc_code"><pre>
1146 bool LICM::runOnFunction(Function &F) {
1147 LoopInfo &LI = getAnalysis<LoopInfo>();
1152 <p>This method call returns a reference to the pass desired. You may get a
1153 runtime assertion failure if you attempt to get an analysis that you did not
1154 declare as required in your <a
1155 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
1156 method can be called by your <tt>run*</tt> method implementation, or by any
1157 other local method invoked by your <tt>run*</tt> method.
1159 A module level pass can use function level analysis info using this interface.
1162 <div class="doc_code"><pre>
1163 bool ModuleLevelPass::runOnModule(Module &M) {
1165 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
1170 <p>In above example, runOnFunction for DominatorTree is called by pass manager
1171 before returning a reference to the desired pass.</p>
1174 If your pass is capable of updating analyses if they exist (e.g.,
1175 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1176 <tt>getAnalysisToUpdate</tt> method, which returns a pointer to the analysis if
1177 it is active. For example:</p>
1179 <div class="doc_code"><pre>
1181 if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
1182 <i>// A DominatorSet is active. This code will update it.</i>
1189 <!-- *********************************************************************** -->
1190 <div class="doc_section">
1191 <a name="analysisgroup">Implementing Analysis Groups</a>
1193 <!-- *********************************************************************** -->
1195 <div class="doc_text">
1197 <p>Now that we understand the basics of how passes are defined, how they are
1198 used, and how they are required from other passes, it's time to get a little bit
1199 fancier. All of the pass relationships that we have seen so far are very
1200 simple: one pass depends on one other specific pass to be run before it can run.
1201 For many applications, this is great, for others, more flexibility is
1204 <p>In particular, some analyses are defined such that there is a single simple
1205 interface to the analysis results, but multiple ways of calculating them.
1206 Consider alias analysis for example. The most trivial alias analysis returns
1207 "may alias" for any alias query. The most sophisticated analysis a
1208 flow-sensitive, context-sensitive interprocedural analysis that can take a
1209 significant amount of time to execute (and obviously, there is a lot of room
1210 between these two extremes for other implementations). To cleanly support
1211 situations like this, the LLVM Pass Infrastructure supports the notion of
1212 Analysis Groups.</p>
1216 <!-- _______________________________________________________________________ -->
1217 <div class="doc_subsubsection">
1218 <a name="agconcepts">Analysis Group Concepts</a>
1221 <div class="doc_text">
1223 <p>An Analysis Group is a single simple interface that may be implemented by
1224 multiple different passes. Analysis Groups can be given human readable names
1225 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1226 class. An analysis group may have one or more implementations, one of which is
1227 the "default" implementation.</p>
1229 <p>Analysis groups are used by client passes just like other passes are: the
1230 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1231 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
1232 scans the available passes to see if any implementations of the analysis group
1233 are available. If none is available, the default implementation is created for
1234 the pass to use. All standard rules for <A href="#interaction">interaction
1235 between passes</a> still apply.</p>
1237 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
1238 passes, all analysis group implementations must be registered, and must use the
1239 <A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
1240 implementation pool. Also, a default implementation of the interface
1241 <b>must</b> be registered with <A
1242 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
1244 <p>As a concrete example of an Analysis Group in action, consider the <a
1245 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1246 analysis group. The default implementation of the alias analysis interface (the
1248 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1249 pass) just does a few simple checks that don't require significant analysis to
1250 compute (such as: two different globals can never alias each other, etc).
1251 Passes that use the <tt><a
1252 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1253 interface (for example the <tt><a
1254 href="http://llvm.org/doxygen/structGCSE.html">gcse</a></tt> pass), do
1255 not care which implementation of alias analysis is actually provided, they just
1256 use the designated interface.</p>
1258 <p>From the user's perspective, commands work just like normal. Issuing the
1259 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1260 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1261 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1262 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1263 hypothetical example) instead.</p>
1267 <!-- _______________________________________________________________________ -->
1268 <div class="doc_subsubsection">
1269 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1272 <div class="doc_text">
1274 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1275 group itself as well as add pass implementations to the analysis group. First,
1276 an analysis should be registered, with a human readable name provided for it.
1277 Unlike registration of passes, there is no command line argument to be specified
1278 for the Analysis Group Interface itself, because it is "abstract":</p>
1280 <div class="doc_code"><pre>
1281 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1284 <p>Once the analysis is registered, passes can declare that they are valid
1285 implementations of the interface by using the following code:</p>
1287 <div class="doc_code"><pre>
1289 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1290 RegisterPass<FancyAA>
1291 B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
1293 //<i> Declare that we implement the AliasAnalysis interface</i>
1294 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> C(B);
1298 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1299 uses the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
1300 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1301 analysis group. Every implementation of an analysis group should join using
1302 this template. A single pass may join multiple different analysis groups with
1305 <div class="doc_code"><pre>
1307 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1308 RegisterPass<<a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1309 D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
1311 //<i> Declare that we implement the AliasAnalysis interface</i>
1312 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <b>true</b>> E(D);
1316 <p>Here we show how the default implementation is specified (using the extra
1317 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1318 one default implementation available at all times for an Analysis Group to be
1319 used. Only default implementation can derive from <tt>ImmutablePass</tt>.
1320 Here we declare that the
1321 <tt><a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1322 pass is the default implementation for the interface.</p>
1326 <!-- *********************************************************************** -->
1327 <div class="doc_section">
1328 <a name="passStatistics">Pass Statistics</a>
1330 <!-- *********************************************************************** -->
1332 <div class="doc_text">
1334 href="http://llvm.org/doxygen/Statistic_8h-source.html"><tt>Statistic</tt></a>
1335 class is designed to be an easy way to expose various success
1336 metrics from passes. These statistics are printed at the end of a
1337 run, when the -stats command line option is enabled on the command
1338 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1343 <!-- *********************************************************************** -->
1344 <div class="doc_section">
1345 <a name="passmanager">What PassManager does</a>
1347 <!-- *********************************************************************** -->
1349 <div class="doc_text">
1352 href="http://llvm.org/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1354 href="http://llvm.org/doxygen/classllvm_1_1PassManager.html">class</a>
1355 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1356 are set up correctly, and then schedules passes to run efficiently. All of the
1357 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1360 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1361 time of a series of passes:</p>
1364 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1365 recomputing analysis results as much as possible. This means keeping track of
1366 which analyses are available already, which analyses get invalidated, and which
1367 analyses are needed to be run for a pass. An important part of work is that the
1368 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1369 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1370 results as soon as they are no longer needed.</li>
1372 <li><b>Pipeline the execution of passes on the program</b> - The
1373 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1374 of a series of passes by pipelining the passes together. This means that, given
1375 a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1376 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1377 the first function, then all of the <a
1378 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1379 etc... until the entire program has been run through the passes.
1381 <p>This improves the cache behavior of the compiler, because it is only touching
1382 the LLVM program representation for a single function at a time, instead of
1383 traversing the entire program. It reduces the memory consumption of compiler,
1384 because, for example, only one <a
1385 href="http://llvm.org/doxygen/classllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1386 needs to be calculated at a time. This also makes it possible to implement
1388 href="#SMP">interesting enhancements</a> in the future.</p></li>
1392 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1393 much information it has about the behaviors of the passes it is scheduling. For
1394 example, the "preserved" set is intentionally conservative in the face of an
1395 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1396 Not implementing when it should be implemented will have the effect of not
1397 allowing any analysis results to live across the execution of your pass.</p>
1399 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1400 options that is useful for debugging pass execution, seeing how things work, and
1401 diagnosing when you should be preserving more analyses than you currently are
1402 (To get information about all of the variants of the <tt>--debug-pass</tt>
1403 option, just type '<tt>opt --help-hidden</tt>').</p>
1405 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1406 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1407 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1409 <div class="doc_code"><pre>
1410 $ opt -load ../../../Debug/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1412 Function Pass Manager
1413 Dominator Set Construction
1414 Immediate Dominators Construction
1415 Global Common Subexpression Elimination
1416 -- Immediate Dominators Construction
1417 -- Global Common Subexpression Elimination
1418 Natural Loop Construction
1419 Loop Invariant Code Motion
1420 -- Natural Loop Construction
1421 -- Loop Invariant Code Motion
1423 -- Dominator Set Construction
1429 <p>This output shows us when passes are constructed and when the analysis
1430 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1431 GCSE uses dominator and immediate dominator information to do its job. The LICM
1432 pass uses natural loop information, which uses dominator sets, but not immediate
1433 dominators. Because immediate dominators are no longer useful after the GCSE
1434 pass, it is immediately destroyed. The dominator sets are then reused to
1435 compute natural loop information, which is then used by the LICM pass.</p>
1437 <p>After the LICM pass, the module verifier runs (which is automatically added
1438 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1439 resultant LLVM code is well formed. After it finishes, the dominator set
1440 information is destroyed, after being computed once, and shared by three
1443 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1444 World</a> pass in between the two passes:</p>
1446 <div class="doc_code"><pre>
1447 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1449 Function Pass Manager
1450 Dominator Set Construction
1451 Immediate Dominators Construction
1452 Global Common Subexpression Elimination
1453 <b>-- Dominator Set Construction</b>
1454 -- Immediate Dominators Construction
1455 -- Global Common Subexpression Elimination
1456 <b> Hello World Pass
1458 Dominator Set Construction</b>
1459 Natural Loop Construction
1460 Loop Invariant Code Motion
1461 -- Natural Loop Construction
1462 -- Loop Invariant Code Motion
1464 -- Dominator Set Construction
1473 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1474 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1475 we need to add the following <a
1476 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1478 <div class="doc_code"><pre>
1479 <i>// We don't modify the program, so we preserve all analyses</i>
1480 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1481 AU.setPreservesAll();
1485 <p>Now when we run our pass, we get this output:</p>
1487 <div class="doc_code"><pre>
1488 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1489 Pass Arguments: -gcse -hello -licm
1491 Function Pass Manager
1492 Dominator Set Construction
1493 Immediate Dominators Construction
1494 Global Common Subexpression Elimination
1495 -- Immediate Dominators Construction
1496 -- Global Common Subexpression Elimination
1499 Natural Loop Construction
1500 Loop Invariant Code Motion
1501 -- Loop Invariant Code Motion
1502 -- Natural Loop Construction
1504 -- Dominator Set Construction
1513 <p>Which shows that we don't accidentally invalidate dominator information
1514 anymore, and therefore do not have to compute it twice.</p>
1518 <!-- _______________________________________________________________________ -->
1519 <div class="doc_subsubsection">
1520 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1523 <div class="doc_text">
1525 <div class="doc_code"><pre>
1526 <b>virtual void</b> releaseMemory();
1529 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1530 results, and how long to keep them around for. Because the lifetime of the pass
1531 object itself is effectively the entire duration of the compilation process, we
1532 need some way to free analysis results when they are no longer useful. The
1533 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1535 <p>If you are writing an analysis or any other pass that retains a significant
1536 amount of state (for use by another pass which "requires" your pass and uses the
1537 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1538 <tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
1539 internal state. This method is called after the <tt>run*</tt> method for the
1540 class, before the next call of <tt>run*</tt> in your pass.</p>
1544 <!-- *********************************************************************** -->
1545 <div class="doc_section">
1546 <a name="registering">Registering dynamically loaded passes</a>
1548 <!-- *********************************************************************** -->
1550 <div class="doc_text">
1552 <p><i>Size matters</i> when constructing production quality tools using llvm,
1553 both for the purposes of distribution, and for regulating the resident code size
1554 when running on the target system. Therefore, it becomes desirable to
1555 selectively use some passes, while omitting others and maintain the flexibility
1556 to change configurations later on. You want to be able to do all this, and,
1557 provide feedback to the user. This is where pass registration comes into
1560 <p>The fundamental mechanisms for pass registration are the
1561 <tt>MachinePassRegistry</tt> class and subclasses of
1562 <tt>MachinePassRegistryNode</tt>.</p>
1564 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1565 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1566 communicates additions and deletions to the command line interface.</p>
1568 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1569 information provided about a particular pass. This information includes the
1570 command line name, the command help string and the address of the function used
1571 to create an instance of the pass. A global static constructor of one of these
1572 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1573 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1574 in the tool will be registered at start up. A dynamically loaded pass will
1575 register on load and unregister at unload.</p>
1579 <!-- _______________________________________________________________________ -->
1580 <div class="doc_subsection">
1581 <a name="registering_existing">Using existing registries</a>
1584 <div class="doc_text">
1586 <p>There are predefined registries to track instruction scheduling
1587 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1588 machine passes. Here we will describe how to <i>register</i> a register
1589 allocator machine pass.</p>
1591 <p>Implement your register allocator machine pass. In your register allocator
1592 .cpp file add the following include;</p>
1594 <div class="doc_code"><pre>
1595 #include "llvm/CodeGen/RegAllocRegistry.h"
1598 <p>Also in your register allocator .cpp file, define a creator function in the
1601 <div class="doc_code"><pre>
1602 FunctionPass *createMyRegisterAllocator() {
1603 return new MyRegisterAllocator();
1607 <p>Note that the signature of this function should match the type of
1608 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1609 "installing" declaration, in the form;</p>
1611 <div class="doc_code"><pre>
1612 static RegisterRegAlloc myRegAlloc("myregalloc",
1613 " my register allocator help string",
1614 createMyRegisterAllocator);
1617 <p>Note the two spaces prior to the help string produces a tidy result on the
1620 <div class="doc_code"><pre>
1623 -regalloc - Register allocator to use: (default = linearscan)
1624 =linearscan - linear scan register allocator
1625 =local - local register allocator
1626 =simple - simple register allocator
1627 =myregalloc - my register allocator help string
1631 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1632 an option. Registering instruction schedulers is similar except use the
1633 <tt>RegisterScheduler</tt> class. Note that the
1634 <tt>RegisterScheduler::FunctionPassCtor</tt> is significantly different from
1635 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1637 <p>To force the load/linking of your register allocator into the llc/lli tools,
1638 add your creator function's global declaration to "Passes.h" and add a "pseudo"
1639 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1644 <!-- _______________________________________________________________________ -->
1645 <div class="doc_subsection">
1646 <a name="registering_new">Creating new registries</a>
1649 <div class="doc_text">
1651 <p>The easiest way to get started is to clone one of the existing registries; we
1652 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1653 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1655 <p>Then you need to declare the registry. Example: if your pass registry is
1656 <tt>RegisterMyPasses</tt> then define;</p>
1658 <div class="doc_code"><pre>
1659 MachinePassRegistry RegisterMyPasses::Registry;
1662 <p>And finally, declare the command line option for your passes. Example:</p>
1664 <div class="doc_code"><pre>
1665 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1666 RegisterPassParser<RegisterMyPasses> >
1668 cl::init(&createDefaultMyPass),
1669 cl::desc("my pass option help"));
1672 <p>Here the command option is "mypass", with createDefaultMyPass as the default
1677 <!-- *********************************************************************** -->
1678 <div class="doc_section">
1679 <a name="debughints">Using GDB with dynamically loaded passes</a>
1681 <!-- *********************************************************************** -->
1683 <div class="doc_text">
1685 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1686 should be. First of all, you can't set a breakpoint in a shared object that has
1687 not been loaded yet, and second of all there are problems with inlined functions
1688 in shared objects. Here are some suggestions to debugging your pass with
1691 <p>For sake of discussion, I'm going to assume that you are debugging a
1692 transformation invoked by <tt>opt</tt>, although nothing described here depends
1697 <!-- _______________________________________________________________________ -->
1698 <div class="doc_subsubsection">
1699 <a name="breakpoint">Setting a breakpoint in your pass</a>
1702 <div class="doc_text">
1704 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1706 <div class="doc_code"><pre>
1709 Copyright 2000 Free Software Foundation, Inc.
1710 GDB is free software, covered by the GNU General Public License, and you are
1711 welcome to change it and/or distribute copies of it under certain conditions.
1712 Type "show copying" to see the conditions.
1713 There is absolutely no warranty for GDB. Type "show warranty" for details.
1714 This GDB was configured as "sparc-sun-solaris2.6"...
1718 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1719 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1720 (the shared object isn't loaded until runtime), we must execute the process, and
1721 have it stop before it invokes our pass, but after it has loaded the shared
1722 object. The most foolproof way of doing this is to set a breakpoint in
1723 <tt>PassManager::run</tt> and then run the process with the arguments you
1726 <div class="doc_code"><pre>
1727 (gdb) <b>break llvm::PassManager::run</b>
1728 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1729 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]</b>
1730 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
1731 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1732 70 bool PassManager::run(Module &M) { return PM->run(M); }
1736 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1737 now free to set breakpoints in your pass so that you can trace through execution
1738 or do other standard debugging stuff.</p>
1742 <!-- _______________________________________________________________________ -->
1743 <div class="doc_subsubsection">
1744 <a name="debugmisc">Miscellaneous Problems</a>
1747 <div class="doc_text">
1749 <p>Once you have the basics down, there are a couple of problems that GDB has,
1750 some with solutions, some without.</p>
1753 <li>Inline functions have bogus stack information. In general, GDB does a
1754 pretty good job getting stack traces and stepping through inline functions.
1755 When a pass is dynamically loaded however, it somehow completely loses this
1756 capability. The only solution I know of is to de-inline a function (move it
1757 from the body of a class to a .cpp file).</li>
1759 <li>Restarting the program breaks breakpoints. After following the information
1760 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1761 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1762 and you start getting errors about breakpoints being unsettable. The only way I
1763 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1764 already set in your pass, run the program, and re-set the breakpoints once
1765 execution stops in <tt>PassManager::run</tt>.</li>
1769 <p>Hopefully these tips will help with common case debugging situations. If
1770 you'd like to contribute some tips of your own, just contact <a
1771 href="mailto:sabre@nondot.org">Chris</a>.</p>
1775 <!-- *********************************************************************** -->
1776 <div class="doc_section">
1777 <a name="future">Future extensions planned</a>
1779 <!-- *********************************************************************** -->
1781 <div class="doc_text">
1783 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1784 some nifty stuff, there are things we'd like to add in the future. Here is
1785 where we are going:</p>
1789 <!-- _______________________________________________________________________ -->
1790 <div class="doc_subsubsection">
1791 <a name="SMP">Multithreaded LLVM</a>
1794 <div class="doc_text">
1796 <p>Multiple CPU machines are becoming more common and compilation can never be
1797 fast enough: obviously we should allow for a multithreaded compiler. Because of
1798 the semantics defined for passes above (specifically they cannot maintain state
1799 across invocations of their <tt>run*</tt> methods), a nice clean way to
1800 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1801 to create multiple instances of each pass object, and allow the separate
1802 instances to be hacking on different parts of the program at the same time.</p>
1804 <p>This implementation would prevent each of the passes from having to implement
1805 multithreaded constructs, requiring only the LLVM core to have locking in a few
1806 places (for global resources). Although this is a simple extension, we simply
1807 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1808 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1812 <!-- *********************************************************************** -->
1815 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1816 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
1817 <a href="http://validator.w3.org/check/referer"><img
1818 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
1820 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1821 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1822 Last modified: $Date$