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>
106 <li><a href="#PassFunctionPass"><tt>ModulePass</tt>es requiring
107 <tt>FunctionPass</tt>es</a></li>
111 <div class="doc_author">
112 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
113 <a href="mailto:jlaskey@mac.com">Jim Laskey</a></p>
116 <!-- *********************************************************************** -->
117 <div class="doc_section">
118 <a name="introduction">Introduction - What is a pass?</a>
120 <!-- *********************************************************************** -->
122 <div class="doc_text">
124 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
125 passes are where most of the interesting parts of the compiler exist. Passes
126 perform the transformations and optimizations that make up the compiler, they
127 build the analysis results that are used by these transformations, and they are,
128 above all, a structuring technique for compiler code.</p>
130 <p>All LLVM passes are subclasses of the <tt><a
131 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
132 class, which implement functionality by overriding virtual methods inherited
133 from <tt>Pass</tt>. Depending on how your pass works, you should inherit from
134 the <tt><a href="#ModulePass">ModulePass</a></tt>, <tt><a
135 href="#CallGraphSCCPass">CallGraphSCCPass</a></tt>, <tt><a
136 href="#FunctionPass">FunctionPass</a></tt>, or <tt><a
137 href="#LoopPass">LoopPass</a></tt>, or <tt><a
138 href="#BasicBlockPass">BasicBlockPass</a></tt> classes, which gives the system
139 more information about what your pass does, and how it can be combined with
140 other passes. One of the main features of the LLVM Pass Framework is that it
141 schedules passes to run in an efficient way based on the constraints that your
142 pass meets (which are indicated by which class they derive from).</p>
144 <p>We start by showing you how to construct a pass, everything from setting up
145 the code, to compiling, loading, and executing it. After the basics are down,
146 more advanced features are discussed.</p>
150 <!-- *********************************************************************** -->
151 <div class="doc_section">
152 <a name="quickstart">Quick Start - Writing hello world</a>
154 <!-- *********************************************************************** -->
156 <div class="doc_text">
158 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
159 is designed to simply print out the name of non-external functions that exist in
160 the program being compiled. It does not modify the program at all, it just
161 inspects it. The source code and files for this pass are available in the LLVM
162 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
166 <!-- ======================================================================= -->
167 <div class="doc_subsection">
168 <a name="makefile">Setting up the build environment</a>
171 <div class="doc_text">
173 <p>First, you need to create a new directory somewhere in the LLVM source
174 base. For this example, we'll assume that you made
175 <tt>lib/Transforms/Hello</tt>. Next, you must set up a build script
176 (Makefile) that will compile the source code for the new pass. To do this,
177 copy the following into <tt>Makefile</tt>:</p>
180 <div class="doc_code"><pre>
181 # Makefile for hello pass
183 # Path to top level of LLVM heirarchy
186 # Name of the library to build
189 # Make the shared library become a loadable module so the tools can
190 # dlopen/dlsym on the resulting library.
193 # Tell the build system which LLVM libraries your pass needs. You'll probably
194 # need at least LLVMSystem.a, LLVMSupport.a, LLVMCore.a but possibly several
196 LLVMLIBS = LLVMCore.a LLVMSupport.a LLVMSystem.a
198 # Include the makefile implementation stuff
199 include $(LEVEL)/Makefile.common
202 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
203 directory are to be compiled and linked together into a
204 <tt>Debug/lib/Hello.so</tt> shared object that can be dynamically loaded by
205 the <tt>opt</tt> or <tt>bugpoint</tt> tools via their <tt>-load</tt> options.
206 If your operating system uses a suffix other than .so (such as windows or
207 Mac OS/X), the appropriate extension will be used.</p>
209 <p>Now that we have the build scripts set up, we just need to write the code for
214 <!-- ======================================================================= -->
215 <div class="doc_subsection">
216 <a name="basiccode">Basic code required</a>
219 <div class="doc_text">
221 <p>Now that we have a way to compile our new pass, we just have to write it.
224 <div class="doc_code"><pre>
225 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
226 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
229 <p>Which are needed because we are writing a <tt><a
230 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>, and
231 we are operating on <tt><a
232 href="http://llvm.org/doxygen/classllvm_1_1Function.html">Function</a></tt>'s.</p>
235 <div class="doc_code"><pre>
236 <b>using namespace llvm;</b>
238 <p>... which is required because the functions from the include files
239 live in the llvm namespace.
244 <div class="doc_code"><pre>
248 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
249 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
250 things declared inside of the anonymous namespace only visible to the current
251 file. If you're not familiar with them, consult a decent C++ book for more
254 <p>Next, we declare our pass itself:</p>
256 <div class="doc_code"><pre>
257 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
260 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
261 href="http://llvm.org/doxygen/classllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
262 The different builtin pass subclasses are described in detail <a
263 href="#passtype">later</a>, but for now, know that <a
264 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate a function at a
267 <div class="doc_code"><pre>
268 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
269 llvm::cerr << "<i>Hello: </i>" << F.getName() << "\n";
272 }; <i>// end of struct Hello</i>
275 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
276 which overloads an abstract virtual method inherited from <a
277 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
278 to do our thing, so we just print out our message with the name of each
281 <div class="doc_code"><pre>
282 RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
283 } <i>// end of anonymous namespace</i>
286 <p>Lastly, we <a href="#registration">register our class</a> <tt>Hello</tt>,
287 giving it a command line
288 argument "<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>".</p>
290 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
292 <div class="doc_code"><pre>
293 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
294 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
296 <b>using namespace llvm;</b>
299 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
300 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
301 llvm::cerr << "<i>Hello: </i>" << F.getName() << "\n";
306 RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
310 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
311 command in the local directory and you should get a new
312 "<tt>Debug/lib/Hello.so</tt> file. Note that everything in this file is
313 contained in an anonymous namespace: this reflects the fact that passes are self
314 contained units that do not need external interfaces (although they can have
315 them) to be useful.</p>
319 <!-- ======================================================================= -->
320 <div class="doc_subsection">
321 <a name="running">Running a pass with <tt>opt</tt></a>
324 <div class="doc_text">
326 <p>Now that you have a brand new shiny shared object file, we can use the
327 <tt>opt</tt> command to run an LLVM program through your pass. Because you
328 registered your pass with the <tt>RegisterPass</tt> template, you will be able to
329 use the <tt>opt</tt> tool to access it, once loaded.</p>
331 <p>To test it, follow the example at the end of the <a
332 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
333 LLVM. We can now run the bytecode file (<tt>hello.bc</tt>) for the program
334 through our transformation like this (or course, any bytecode file will
337 <div class="doc_code"><pre>
338 $ opt -load ../../../Debug/lib/Hello.so -hello < hello.bc > /dev/null
344 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
345 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
346 argument (which is one reason you need to <a href="#registration">register your
347 pass</a>). Because the hello pass does not modify the program in any
348 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
349 <tt>/dev/null</tt>).</p>
351 <p>To see what happened to the other string you registered, try running
352 <tt>opt</tt> with the <tt>--help</tt> option:</p>
354 <div class="doc_code"><pre>
355 $ opt -load ../../../Debug/lib/Hello.so --help
356 OVERVIEW: llvm .bc -> .bc modular optimizer
358 USAGE: opt [options] <input bytecode>
361 Optimizations available:
363 -funcresolve - Resolve Functions
364 -gcse - Global Common Subexpression Elimination
365 -globaldce - Dead Global Elimination
366 <b>-hello - Hello World Pass</b>
367 -indvars - Canonicalize Induction Variables
368 -inline - Function Integration/Inlining
369 -instcombine - Combine redundant instructions
373 <p>The pass name get added as the information string for your pass, giving some
374 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
375 would go ahead and make it do the cool transformations you want. Once you get
376 it all working and tested, it may become useful to find out how fast your pass
377 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
378 line option (<tt>--time-passes</tt>) that allows you to get information about
379 the execution time of your pass along with the other passes you queue up. For
382 <div class="doc_code"><pre>
383 $ opt -load ../../../Debug/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
387 ===============================================================================
388 ... Pass execution timing report ...
389 ===============================================================================
390 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
392 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
393 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bytecode Writer
394 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
395 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
396 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
397 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
400 <p>As you can see, our implementation above is pretty fast :). The additional
401 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
402 that the LLVM emitted by your pass is still valid and well formed LLVM, which
403 hasn't been broken somehow.</p>
405 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
406 about some more details of how they work and how to use them.</p>
410 <!-- *********************************************************************** -->
411 <div class="doc_section">
412 <a name="passtype">Pass classes and requirements</a>
414 <!-- *********************************************************************** -->
416 <div class="doc_text">
418 <p>One of the first things that you should do when designing a new pass is to
419 decide what class you should subclass for your pass. The <a
420 href="#basiccode">Hello World</a> example uses the <tt><a
421 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
422 did not discuss why or when this should occur. Here we talk about the classes
423 available, from the most general to the most specific.</p>
425 <p>When choosing a superclass for your Pass, you should choose the <b>most
426 specific</b> class possible, while still being able to meet the requirements
427 listed. This gives the LLVM Pass Infrastructure information necessary to
428 optimize how passes are run, so that the resultant compiler isn't unneccesarily
433 <!-- ======================================================================= -->
434 <div class="doc_subsection">
435 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
438 <div class="doc_text">
440 <p>The most plain and boring type of pass is the "<tt><a
441 href="http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
442 class. This pass type is used for passes that do not have to be run, do not
443 change state, and never need to be updated. This is not a normal type of
444 transformation or analysis, but can provide information about the current
445 compiler configuration.</p>
447 <p>Although this pass class is very infrequently used, it is important for
448 providing information about the current target machine being compiled for, and
449 other static information that can affect the various transformations.</p>
451 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
452 invalidated, and are never "run".</p>
456 <!-- ======================================================================= -->
457 <div class="doc_subsection">
458 <a name="ModulePass">The <tt>ModulePass</tt> class</a>
461 <div class="doc_text">
464 href="http://llvm.org/doxygen/classllvm_1_1ModulePass.html">ModulePass</a></tt>"
465 class is the most general of all superclasses that you can use. Deriving from
466 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
467 refering to function bodies in no predictable order, or adding and removing
468 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
469 subclasses, no optimization can be done for their execution.</p>
471 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
472 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
473 following signature:</p>
477 <!-- _______________________________________________________________________ -->
478 <div class="doc_subsubsection">
479 <a name="runOnModule">The <tt>runOnModule</tt> method</a>
482 <div class="doc_text">
484 <div class="doc_code"><pre>
485 <b>virtual bool</b> runOnModule(Module &M) = 0;
488 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
489 It should return true if the module was modified by the transformation and
494 <!-- ======================================================================= -->
495 <div class="doc_subsection">
496 <a name="CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
499 <div class="doc_text">
502 href="http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html">CallGraphSCCPass</a></tt>"
503 is used by passes that need to traverse the program bottom-up on the call graph
504 (callees before callers). Deriving from CallGraphSCCPass provides some
505 mechanics for building and traversing the CallGraph, but also allows the system
506 to optimize execution of CallGraphSCCPass's. If your pass meets the
507 requirements outlined below, and doesn't meet the requirements of a <tt><a
508 href="#FunctionPass">FunctionPass</a></tt> or <tt><a
509 href="#BasicBlockPass">BasicBlockPass</a></tt>, you should derive from
510 <tt>CallGraphSCCPass</tt>.</p>
512 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
514 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
518 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
519 the current SCC.</li>
521 <li>... <em>allowed</em> to inspect any Function's other than those in the
522 current SCC and the direct callees of the SCC.</li>
524 <li>... <em>required</em> to preserve the current CallGraph object, updating it
525 to reflect any changes made to the program.</li>
527 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
528 though they may change the contents of an SCC.</li>
530 <li>... <em>allowed</em> to add or remove global variables from the current
533 <li>... <em>allowed</em> to maintain state across invocations of
534 <a href="#runOnSCC"><tt>runOnSCC</tt></a> (including global data).</li>
537 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
538 because it has to handle SCCs with more than one node in it. All of the virtual
539 methods described below should return true if they modified the program, or
540 false if they didn't.</p>
544 <!-- _______________________________________________________________________ -->
545 <div class="doc_subsubsection">
546 <a name="doInitialization_scc">The <tt>doInitialization(CallGraph &)</tt>
550 <div class="doc_text">
552 <div class="doc_code"><pre>
553 <b>virtual bool</b> doInitialization(CallGraph &CG);
556 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
557 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
558 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
559 is designed to do simple initialization type of stuff that does not depend on
560 the SCCs being processed. The <tt>doInitialization</tt> method call is not
561 scheduled to overlap with any other pass executions (thus it should be very
566 <!-- _______________________________________________________________________ -->
567 <div class="doc_subsubsection">
568 <a name="runOnSCC">The <tt>runOnSCC</tt> method</a>
571 <div class="doc_text">
573 <div class="doc_code"><pre>
574 <b>virtual bool</b> runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;
577 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
578 should return true if the module was modified by the transformation, false
583 <!-- _______________________________________________________________________ -->
584 <div class="doc_subsubsection">
585 <a name="doFinalization_scc">The <tt>doFinalization(CallGraph
586 &)</tt> method</a>
589 <div class="doc_text">
591 <div class="doc_code"><pre>
592 <b>virtual bool</b> doFinalization(CallGraph &CG);
595 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
596 called when the pass framework has finished calling <a
597 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
598 program being compiled.</p>
602 <!-- ======================================================================= -->
603 <div class="doc_subsection">
604 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
607 <div class="doc_text">
609 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
610 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
611 subclasses do have a predictable, local behavior that can be expected by the
612 system. All <tt>FunctionPass</tt> execute on each function in the program
613 independent of all of the other functions in the program.
614 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
615 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
617 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
620 <li>Modify a Function other than the one currently being processed.</li>
621 <li>Add or remove Function's from the current Module.</li>
622 <li>Add or remove global variables from the current Module.</li>
623 <li>Maintain state across invocations of
624 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
627 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
628 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
629 may overload three virtual methods to do their work. All of these methods
630 should return true if they modified the program, or false if they didn't.</p>
634 <!-- _______________________________________________________________________ -->
635 <div class="doc_subsubsection">
636 <a name="doInitialization_mod">The <tt>doInitialization(Module &)</tt>
640 <div class="doc_text">
642 <div class="doc_code"><pre>
643 <b>virtual bool</b> doInitialization(Module &M);
646 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
647 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
648 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
649 is designed to do simple initialization type of stuff that does not depend on
650 the functions being processed. The <tt>doInitialization</tt> method call is not
651 scheduled to overlap with any other pass executions (thus it should be very
654 <p>A good example of how this method should be used is the <a
655 href="http://llvm.org/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
656 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
657 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
658 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
659 free functions that it needs, adding prototypes to the module if necessary.</p>
663 <!-- _______________________________________________________________________ -->
664 <div class="doc_subsubsection">
665 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
668 <div class="doc_text">
670 <div class="doc_code"><pre>
671 <b>virtual bool</b> runOnFunction(Function &F) = 0;
674 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
675 the transformation or analysis work of your pass. As usual, a true value should
676 be returned if the function is modified.</p>
680 <!-- _______________________________________________________________________ -->
681 <div class="doc_subsubsection">
682 <a name="doFinalization_mod">The <tt>doFinalization(Module
683 &)</tt> method</a>
686 <div class="doc_text">
688 <div class="doc_code"><pre>
689 <b>virtual bool</b> doFinalization(Module &M);
692 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
693 called when the pass framework has finished calling <a
694 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
695 program being compiled.</p>
699 <!-- ======================================================================= -->
700 <div class="doc_subsection">
701 <a name="LoopPass">The <tt>LoopPass</tt> class </a>
704 <div class="doc_text">
706 <p> All <tt>LoopPass</tt> execute on each loop in the function independent of
707 all of the other loops in the function. <tt>LoopPass</tt> processes loops in
708 loop nest order such that outer most loop is processed last. </p>
710 <p> <tt>LoopPass</tt> subclasses are allowed to update loop nest using
711 <tt>LPPassManager</tt> interface. Implementing a loop pass is usually
712 straightforward. <tt>Looppass</tt>'s may overload three virtual methods to
713 do their work. All these methods should return true if they modified the
714 program, or false if they didn't. </p>
717 <!-- _______________________________________________________________________ -->
718 <div class="doc_subsubsection">
719 <a name="doInitialization_loop">The <tt>doInitialization(Loop *,
720 LPPassManager &)</tt>
724 <div class="doc_text">
726 <div class="doc_code"><pre>
727 <b>virtual bool</b> doInitialization(Loop *, LPPassManager &LPM);
730 The <tt>doInitialization</tt> method is designed to do simple initialization
731 type of stuff that does not depend on the functions being processed. The
732 <tt>doInitialization</tt> method call is not scheduled to overlap with any
733 other pass executions (thus it should be very fast). LPPassManager
734 interface should be used to access Function or Module level analysis
740 <!-- _______________________________________________________________________ -->
741 <div class="doc_subsubsection">
742 <a name="runOnLoop">The <tt>runOnLoop</tt> method</a>
745 <div class="doc_text">
747 <div class="doc_code"><pre>
748 <b>virtual bool</b> runOnLoop(Loop *, LPPassManager &LPM) = 0;
751 <p>The <tt>runOnLoop</tt> method must be implemented by your subclass to do
752 the transformation or analysis work of your pass. As usual, a true value should
753 be returned if the function is modified. <tt>LPPassManager</tt> interface
754 should be used to update loop nest.</p>
758 <!-- _______________________________________________________________________ -->
759 <div class="doc_subsubsection">
760 <a name="doFinalization_loop">The <tt>doFinalization()</tt> method</a>
763 <div class="doc_text">
765 <div class="doc_code"><pre>
766 <b>virtual bool</b> doFinalization();
769 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
770 called when the pass framework has finished calling <a
771 href="#runOnLoop"><tt>runOnLoop</tt></a> for every loop in the
772 program being compiled. </p>
778 <!-- ======================================================================= -->
779 <div class="doc_subsection">
780 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
783 <div class="doc_text">
785 <p><tt>BasicBlockPass</tt>'s are just like <a
786 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
787 their scope of inspection and modification to a single basic block at a time.
788 As such, they are <b>not</b> allowed to do any of the following:</p>
791 <li>Modify or inspect any basic blocks outside of the current one</li>
792 <li>Maintain state across invocations of
793 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
794 <li>Modify the control flow graph (by altering terminator instructions)</li>
795 <li>Any of the things forbidden for
796 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
799 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
800 optimizations. They may override the same <a
801 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
802 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
803 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
807 <!-- _______________________________________________________________________ -->
808 <div class="doc_subsubsection">
809 <a name="doInitialization_fn">The <tt>doInitialization(Function
810 &)</tt> method</a>
813 <div class="doc_text">
815 <div class="doc_code"><pre>
816 <b>virtual bool</b> doInitialization(Function &F);
819 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
820 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
821 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
822 to do simple initialization that does not depend on the
823 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
824 scheduled to overlap with any other pass executions (thus it should be very
829 <!-- _______________________________________________________________________ -->
830 <div class="doc_subsubsection">
831 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
834 <div class="doc_text">
836 <div class="doc_code"><pre>
837 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
840 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
841 function is not allowed to inspect or modify basic blocks other than the
842 parameter, and are not allowed to modify the CFG. A true value must be returned
843 if the basic block is modified.</p>
847 <!-- _______________________________________________________________________ -->
848 <div class="doc_subsubsection">
849 <a name="doFinalization_fn">The <tt>doFinalization(Function &)</tt>
853 <div class="doc_text">
855 <div class="doc_code"><pre>
856 <b>virtual bool</b> doFinalization(Function &F);
859 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
860 called when the pass framework has finished calling <a
861 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
862 program being compiled. This can be used to perform per-function
867 <!-- ======================================================================= -->
868 <div class="doc_subsection">
869 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
872 <div class="doc_text">
874 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
875 executes on the machine-dependent representation of each LLVM function in the
876 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
877 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
878 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
879 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
882 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
883 <li>Modify a MachineFunction other than the one currently being processed.</li>
884 <li>Add or remove MachineFunctions from the current Module.</li>
885 <li>Add or remove global variables from the current Module.</li>
886 <li>Maintain state across invocations of <a
887 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
893 <!-- _______________________________________________________________________ -->
894 <div class="doc_subsubsection">
895 <a name="runOnMachineFunction">The <tt>runOnMachineFunction(MachineFunction
896 &MF)</tt> method</a>
899 <div class="doc_text">
901 <div class="doc_code"><pre>
902 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
905 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
906 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
907 work of your <tt>MachineFunctionPass</tt>.</p>
909 <p>The <tt>runOnMachineFunction</tt> method is called on every
910 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
911 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
912 representation of the function. If you want to get at the LLVM <tt>Function</tt>
913 for the <tt>MachineFunction</tt> you're working on, use
914 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
915 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
916 <tt>MachineFunctionPass</tt>.</p>
920 <!-- *********************************************************************** -->
921 <div class="doc_section">
922 <a name="registration">Pass registration</a>
924 <!-- *********************************************************************** -->
926 <div class="doc_text">
928 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
929 pass registration works, and discussed some of the reasons that it is used and
930 what it does. Here we discuss how and why passes are registered.</p>
932 <p>As we saw above, passes are registered with the <b><tt>RegisterPass</tt></b>
933 template, which requires you to pass at least two
934 parameters. The first parameter is the name of the pass that is to be used on
935 the command line to specify that the pass should be added to a program (for
936 example, with <tt>opt</tt> or <tt>bugpoint</tt>). The second argument is the
937 name of the pass, which is to be used for the <tt>--help</tt> output of
939 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
941 <p>If you want your pass to be easily dumpable, you should
942 implement the virtual <tt>print</tt> method:</p>
946 <!-- _______________________________________________________________________ -->
947 <div class="doc_subsubsection">
948 <a name="print">The <tt>print</tt> method</a>
951 <div class="doc_text">
953 <div class="doc_code"><pre>
954 <b>virtual void</b> print(llvm::OStream &O, <b>const</b> Module *M) <b>const</b>;
957 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
958 a human readable version of the analysis results. This is useful for debugging
959 an analysis itself, as well as for other people to figure out how an analysis
960 works. Use the <tt>opt -analyze</tt> argument to invoke this method.</p>
962 <p>The <tt>llvm::OStream</tt> parameter specifies the stream to write the results on,
963 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
964 program that has been analyzed. Note however that this pointer may be null in
965 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
966 debugger), so it should only be used to enhance debug output, it should not be
971 <!-- *********************************************************************** -->
972 <div class="doc_section">
973 <a name="interaction">Specifying interactions between passes</a>
975 <!-- *********************************************************************** -->
977 <div class="doc_text">
979 <p>One of the main responsibilities of the <tt>PassManager</tt> is the make sure
980 that passes interact with each other correctly. Because <tt>PassManager</tt>
981 tries to <a href="#passmanager">optimize the execution of passes</a> it must
982 know how the passes interact with each other and what dependencies exist between
983 the various passes. To track this, each pass can declare the set of passes that
984 are required to be executed before the current pass, and the passes which are
985 invalidated by the current pass.</p>
987 <p>Typically this functionality is used to require that analysis results are
988 computed before your pass is run. Running arbitrary transformation passes can
989 invalidate the computed analysis results, which is what the invalidation set
990 specifies. If a pass does not implement the <tt><a
991 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
992 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
996 <!-- _______________________________________________________________________ -->
997 <div class="doc_subsubsection">
998 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
1001 <div class="doc_text">
1003 <div class="doc_code"><pre>
1004 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
1007 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
1008 invalidated sets may be specified for your transformation. The implementation
1009 should fill in the <tt><a
1010 href="http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
1011 object with information about which passes are required and not invalidated. To
1012 do this, a pass may call any of the following methods on the AnalysisUsage
1016 <!-- _______________________________________________________________________ -->
1017 <div class="doc_subsubsection">
1018 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
1021 <div class="doc_text">
1023 If your pass requires a previous pass to be executed (an analysis for example),
1024 it can use one of these methods to arrange for it to be run before your pass.
1025 LLVM has many different types of analyses and passes that can be required,
1026 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
1027 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
1028 be no critical edges in the CFG when your pass has been run.
1032 Some analyses chain to other analyses to do their job. For example, an <a
1033 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
1034 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
1035 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
1036 used instead of the <tt>addRequired</tt> method. This informs the PassManager
1037 that the transitively required pass should be alive as long as the requiring
1042 <!-- _______________________________________________________________________ -->
1043 <div class="doc_subsubsection">
1044 <a name="AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a>
1047 <div class="doc_text">
1049 One of the jobs of the PassManager is to optimize how and when analyses are run.
1050 In particular, it attempts to avoid recomputing data unless it needs to. For
1051 this reason, passes are allowed to declare that they preserve (i.e., they don't
1052 invalidate) an existing analysis if it's available. For example, a simple
1053 constant folding pass would not modify the CFG, so it can't possibly affect the
1054 results of dominator analysis. By default, all passes are assumed to invalidate
1059 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
1060 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
1061 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
1062 not modify the LLVM program at all (which is true for analyses), and the
1063 <tt>setPreservesCFG</tt> method can be used by transformations that change
1064 instructions in the program but do not modify the CFG or terminator instructions
1065 (note that this property is implicitly set for <a
1066 href="#BasicBlockPass">BasicBlockPass</a>'s).
1070 <tt>addPreserved</tt> is particularly useful for transformations like
1071 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
1072 and dominator related analyses if they exist, so it can preserve them, despite
1073 the fact that it hacks on the CFG.
1077 <!-- _______________________________________________________________________ -->
1078 <div class="doc_subsubsection">
1079 <a name="AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a>
1082 <div class="doc_text">
1084 <div class="doc_code"><pre>
1085 <i>// This is an example implementation from an analysis, which does not modify
1086 // the program at all, yet has a prerequisite.</i>
1087 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1088 AU.setPreservesAll();
1089 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1095 <div class="doc_code"><pre>
1096 <i>// This example modifies the program, but does not modify the CFG</i>
1097 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1098 AU.setPreservesCFG();
1099 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1105 <!-- _______________________________________________________________________ -->
1106 <div class="doc_subsubsection">
1107 <a name="getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a>
1110 <div class="doc_text">
1112 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1113 your class, providing you with access to the passes that you declared that you
1114 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
1115 method. It takes a single template argument that specifies which pass class you
1116 want, and returns a reference to that pass. For example:</p>
1118 <div class="doc_code"><pre>
1119 bool LICM::runOnFunction(Function &F) {
1120 LoopInfo &LI = getAnalysis<LoopInfo>();
1125 <p>This method call returns a reference to the pass desired. You may get a
1126 runtime assertion failure if you attempt to get an analysis that you did not
1127 declare as required in your <a
1128 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
1129 method can be called by your <tt>run*</tt> method implementation, or by any
1130 other local method invoked by your <tt>run*</tt> method.</p>
1133 If your pass is capable of updating analyses if they exist (e.g.,
1134 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1135 <tt>getAnalysisToUpdate</tt> method, which returns a pointer to the analysis if
1136 it is active. For example:</p>
1138 <div class="doc_code"><pre>
1140 if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
1141 <i>// A DominatorSet is active. This code will update it.</i>
1148 <!-- *********************************************************************** -->
1149 <div class="doc_section">
1150 <a name="analysisgroup">Implementing Analysis Groups</a>
1152 <!-- *********************************************************************** -->
1154 <div class="doc_text">
1156 <p>Now that we understand the basics of how passes are defined, how the are
1157 used, and how they are required from other passes, it's time to get a little bit
1158 fancier. All of the pass relationships that we have seen so far are very
1159 simple: one pass depends on one other specific pass to be run before it can run.
1160 For many applications, this is great, for others, more flexibility is
1163 <p>In particular, some analyses are defined such that there is a single simple
1164 interface to the analysis results, but multiple ways of calculating them.
1165 Consider alias analysis for example. The most trivial alias analysis returns
1166 "may alias" for any alias query. The most sophisticated analysis a
1167 flow-sensitive, context-sensitive interprocedural analysis that can take a
1168 significant amount of time to execute (and obviously, there is a lot of room
1169 between these two extremes for other implementations). To cleanly support
1170 situations like this, the LLVM Pass Infrastructure supports the notion of
1171 Analysis Groups.</p>
1175 <!-- _______________________________________________________________________ -->
1176 <div class="doc_subsubsection">
1177 <a name="agconcepts">Analysis Group Concepts</a>
1180 <div class="doc_text">
1182 <p>An Analysis Group is a single simple interface that may be implemented by
1183 multiple different passes. Analysis Groups can be given human readable names
1184 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1185 class. An analysis group may have one or more implementations, one of which is
1186 the "default" implementation.</p>
1188 <p>Analysis groups are used by client passes just like other passes are: the
1189 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1190 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
1191 scans the available passes to see if any implementations of the analysis group
1192 are available. If none is available, the default implementation is created for
1193 the pass to use. All standard rules for <A href="#interaction">interaction
1194 between passes</a> still apply.</p>
1196 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
1197 passes, all analysis group implementations must be registered, and must use the
1198 <A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
1199 implementation pool. Also, a default implementation of the interface
1200 <b>must</b> be registered with <A
1201 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
1203 <p>As a concrete example of an Analysis Group in action, consider the <a
1204 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1205 analysis group. The default implementation of the alias analysis interface (the
1207 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1208 pass) just does a few simple checks that don't require significant analysis to
1209 compute (such as: two different globals can never alias each other, etc).
1210 Passes that use the <tt><a
1211 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1212 interface (for example the <tt><a
1213 href="http://llvm.org/doxygen/structGCSE.html">gcse</a></tt> pass), do
1214 not care which implementation of alias analysis is actually provided, they just
1215 use the designated interface.</p>
1217 <p>From the user's perspective, commands work just like normal. Issuing the
1218 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1219 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1220 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1221 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1222 hypothetical example) instead.</p>
1226 <!-- _______________________________________________________________________ -->
1227 <div class="doc_subsubsection">
1228 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1231 <div class="doc_text">
1233 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1234 group itself as well as add pass implementations to the analysis group. First,
1235 an analysis should be registered, with a human readable name provided for it.
1236 Unlike registration of passes, there is no command line argument to be specified
1237 for the Analysis Group Interface itself, because it is "abstract":</p>
1239 <div class="doc_code"><pre>
1240 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1243 <p>Once the analysis is registered, passes can declare that they are valid
1244 implementations of the interface by using the following code:</p>
1246 <div class="doc_code"><pre>
1248 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1249 RegisterPass<FancyAA>
1250 B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
1252 //<i> Declare that we implement the AliasAnalysis interface</i>
1253 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> C(B);
1257 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1258 uses the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
1259 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1260 analysis group. Every implementation of an analysis group should join using
1261 this template. A single pass may join multiple different analysis groups with
1264 <div class="doc_code"><pre>
1266 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1267 RegisterPass<<a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1268 D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
1270 //<i> Declare that we implement the AliasAnalysis interface</i>
1271 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <b>true</b>> E(D);
1275 <p>Here we show how the default implementation is specified (using the extra
1276 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1277 one default implementation available at all times for an Analysis Group to be
1278 used. Here we declare that the <tt><a
1279 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1280 pass is the default implementation for the interface.</p>
1284 <!-- *********************************************************************** -->
1285 <div class="doc_section">
1286 <a name="passStatistics">Pass Statistics</a>
1288 <!-- *********************************************************************** -->
1290 <div class="doc_text">
1292 href="http://llvm.org/doxygen/Statistic_8h-source.html"><tt>Statistic</tt></a>
1293 class is designed to be an easy way to expose various success
1294 metrics from passes. These statistics are printed at the end of a
1295 run, when the -stats command line option is enabled on the command
1296 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1301 <!-- *********************************************************************** -->
1302 <div class="doc_section">
1303 <a name="passmanager">What PassManager does</a>
1305 <!-- *********************************************************************** -->
1307 <div class="doc_text">
1310 href="http://llvm.org/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1312 href="http://llvm.org/doxygen/classllvm_1_1PassManager.html">class</a>
1313 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1314 are set up correctly, and then schedules passes to run efficiently. All of the
1315 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1318 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1319 time of a series of passes:</p>
1322 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1323 recomputing analysis results as much as possible. This means keeping track of
1324 which analyses are available already, which analyses get invalidated, and which
1325 analyses are needed to be run for a pass. An important part of work is that the
1326 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1327 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1328 results as soon as they are no longer needed.</li>
1330 <li><b>Pipeline the execution of passes on the program</b> - The
1331 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1332 of a series of passes by pipelining the passes together. This means that, given
1333 a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1334 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1335 the first function, then all of the <a
1336 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1337 etc... until the entire program has been run through the passes.
1339 <p>This improves the cache behavior of the compiler, because it is only touching
1340 the LLVM program representation for a single function at a time, instead of
1341 traversing the entire program. It reduces the memory consumption of compiler,
1342 because, for example, only one <a
1343 href="http://llvm.org/doxygen/classllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1344 needs to be calculated at a time. This also makes it possible some <a
1345 href="#SMP">interesting enhancements</a> in the future.</p></li>
1349 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1350 much information it has about the behaviors of the passes it is scheduling. For
1351 example, the "preserved" set is intentionally conservative in the face of an
1352 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1353 Not implementing when it should be implemented will have the effect of not
1354 allowing any analysis results to live across the execution of your pass.</p>
1356 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1357 options that is useful for debugging pass execution, seeing how things work, and
1358 diagnosing when you should be preserving more analyses than you currently are
1359 (To get information about all of the variants of the <tt>--debug-pass</tt>
1360 option, just type '<tt>opt --help-hidden</tt>').</p>
1362 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1363 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1364 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1366 <div class="doc_code"><pre>
1367 $ opt -load ../../../Debug/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1369 Function Pass Manager
1370 Dominator Set Construction
1371 Immediate Dominators Construction
1372 Global Common Subexpression Elimination
1373 -- Immediate Dominators Construction
1374 -- Global Common Subexpression Elimination
1375 Natural Loop Construction
1376 Loop Invariant Code Motion
1377 -- Natural Loop Construction
1378 -- Loop Invariant Code Motion
1380 -- Dominator Set Construction
1386 <p>This output shows us when passes are constructed and when the analysis
1387 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1388 GCSE uses dominator and immediate dominator information to do its job. The LICM
1389 pass uses natural loop information, which uses dominator sets, but not immediate
1390 dominators. Because immediate dominators are no longer useful after the GCSE
1391 pass, it is immediately destroyed. The dominator sets are then reused to
1392 compute natural loop information, which is then used by the LICM pass.</p>
1394 <p>After the LICM pass, the module verifier runs (which is automatically added
1395 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1396 resultant LLVM code is well formed. After it finishes, the dominator set
1397 information is destroyed, after being computed once, and shared by three
1400 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1401 World</a> pass in between the two passes:</p>
1403 <div class="doc_code"><pre>
1404 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1406 Function Pass Manager
1407 Dominator Set Construction
1408 Immediate Dominators Construction
1409 Global Common Subexpression Elimination
1410 <b>-- Dominator Set Construction</b>
1411 -- Immediate Dominators Construction
1412 -- Global Common Subexpression Elimination
1413 <b> Hello World Pass
1415 Dominator Set Construction</b>
1416 Natural Loop Construction
1417 Loop Invariant Code Motion
1418 -- Natural Loop Construction
1419 -- Loop Invariant Code Motion
1421 -- Dominator Set Construction
1430 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1431 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1432 we need to add the following <a
1433 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1435 <div class="doc_code"><pre>
1436 <i>// We don't modify the program, so we preserve all analyses</i>
1437 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1438 AU.setPreservesAll();
1442 <p>Now when we run our pass, we get this output:</p>
1444 <div class="doc_code"><pre>
1445 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1446 Pass Arguments: -gcse -hello -licm
1448 Function Pass Manager
1449 Dominator Set Construction
1450 Immediate Dominators Construction
1451 Global Common Subexpression Elimination
1452 -- Immediate Dominators Construction
1453 -- Global Common Subexpression Elimination
1456 Natural Loop Construction
1457 Loop Invariant Code Motion
1458 -- Loop Invariant Code Motion
1459 -- Natural Loop Construction
1461 -- Dominator Set Construction
1470 <p>Which shows that we don't accidentally invalidate dominator information
1471 anymore, and therefore do not have to compute it twice.</p>
1475 <!-- _______________________________________________________________________ -->
1476 <div class="doc_subsubsection">
1477 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1480 <div class="doc_text">
1482 <div class="doc_code"><pre>
1483 <b>virtual void</b> releaseMemory();
1486 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1487 results, and how long to keep them around for. Because the lifetime of the pass
1488 object itself is effectively the entire duration of the compilation process, we
1489 need some way to free analysis results when they are no longer useful. The
1490 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1492 <p>If you are writing an analysis or any other pass that retains a significant
1493 amount of state (for use by another pass which "requires" your pass and uses the
1494 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1495 <tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
1496 internal state. This method is called after the <tt>run*</tt> method for the
1497 class, before the next call of <tt>run*</tt> in your pass.</p>
1501 <!-- *********************************************************************** -->
1502 <div class="doc_section">
1503 <a name="registering">Registering dynamically loaded passes</a>
1505 <!-- *********************************************************************** -->
1507 <div class="doc_text">
1509 <p><i>Size matters</i> when constructing production quality tools using llvm,
1510 both for the purposes of distribution, and for regulating the resident code size
1511 when running on the target system. Therefore, it becomes desirable to
1512 selectively use some passes, while omitting others and maintain the flexibility
1513 to change configurations later on. You want to be able to do all this, and,
1514 provide feedback to the user. This is where pass registration comes into
1517 <p>The fundamental mechanisms for pass registration are the
1518 <tt>MachinePassRegistry</tt> class and subclasses of
1519 <tt>MachinePassRegistryNode</tt>.</p>
1521 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1522 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1523 communicates additions and deletions to the command line interface.</p>
1525 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1526 information provided about a particular pass. This information includes the
1527 command line name, the command help string and the address of the function used
1528 to create an instance of the pass. A global static constructor of one of these
1529 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1530 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1531 in the tool will be registered at start up. A dynamically loaded pass will
1532 register on load and unregister at unload.</p>
1536 <!-- _______________________________________________________________________ -->
1537 <div class="doc_subsection">
1538 <a name="registering_existing">Using existing registries</a>
1541 <div class="doc_text">
1543 <p>There are predefined registries to track instruction scheduling
1544 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1545 machine passes. Here we will describe how to <i>register</i> a register
1546 allocator machine pass.</p>
1548 <p>Implement your register allocator machine pass. In your register allocator
1549 .cpp file add the following include;</p>
1551 <div class="doc_code"><pre>
1552 #include "llvm/CodeGen/RegAllocRegistry.h"
1555 <p>Also in your register allocator .cpp file, define a creator function in the
1558 <div class="doc_code"><pre>
1559 FunctionPass *createMyRegisterAllocator() {
1560 return new MyRegisterAllocator();
1564 <p>Note that the signature of this function should match the type of
1565 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1566 "installing" declaration, in the form;</p>
1568 <div class="doc_code"><pre>
1569 static RegisterRegAlloc myRegAlloc("myregalloc",
1570 " my register allocator help string",
1571 createMyRegisterAllocator);
1574 <p>Note the two spaces prior to the help string produces a tidy result on the
1577 <div class="doc_code"><pre>
1580 -regalloc - Register allocator to use: (default = linearscan)
1581 =linearscan - linear scan register allocator
1582 =local - local register allocator
1583 =simple - simple register allocator
1584 =myregalloc - my register allocator help string
1588 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1589 an option. Registering instruction schedulers is similar except use the
1590 <tt>RegisterScheduler</tt> class. Note that the
1591 <tt>RegisterScheduler::FunctionPassCtor</tt> is significantly different from
1592 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1594 <p>To force the load/linking of your register allocator into the llc/lli tools,
1595 add your creator function's global declaration to "Passes.h" and add a "pseudo"
1596 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1601 <!-- _______________________________________________________________________ -->
1602 <div class="doc_subsection">
1603 <a name="registering_new">Creating new registries</a>
1606 <div class="doc_text">
1608 <p>The easiest way to get started is to clone one of the existing registries; we
1609 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1610 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1612 <p>Then you need to declare the registry. Example: if your pass registry is
1613 <tt>RegisterMyPasses</tt> then define;</p>
1615 <div class="doc_code"><pre>
1616 MachinePassRegistry RegisterMyPasses::Registry;
1619 <p>And finally, declare the command line option for your passes. Example:</p>
1621 <div class="doc_code"><pre>
1622 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1623 RegisterPassParser<RegisterMyPasses> >
1625 cl::init(&createDefaultMyPass),
1626 cl::desc("my pass option help"));
1629 <p>Here the command option is "mypass", with createDefaultMyPass as the default
1634 <!-- *********************************************************************** -->
1635 <div class="doc_section">
1636 <a name="debughints">Using GDB with dynamically loaded passes</a>
1638 <!-- *********************************************************************** -->
1640 <div class="doc_text">
1642 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1643 should be. First of all, you can't set a breakpoint in a shared object that has
1644 not been loaded yet, and second of all there are problems with inlined functions
1645 in shared objects. Here are some suggestions to debugging your pass with
1648 <p>For sake of discussion, I'm going to assume that you are debugging a
1649 transformation invoked by <tt>opt</tt>, although nothing described here depends
1654 <!-- _______________________________________________________________________ -->
1655 <div class="doc_subsubsection">
1656 <a name="breakpoint">Setting a breakpoint in your pass</a>
1659 <div class="doc_text">
1661 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1663 <div class="doc_code"><pre>
1666 Copyright 2000 Free Software Foundation, Inc.
1667 GDB is free software, covered by the GNU General Public License, and you are
1668 welcome to change it and/or distribute copies of it under certain conditions.
1669 Type "show copying" to see the conditions.
1670 There is absolutely no warranty for GDB. Type "show warranty" for details.
1671 This GDB was configured as "sparc-sun-solaris2.6"...
1675 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1676 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1677 (the shared object isn't loaded until runtime), we must execute the process, and
1678 have it stop before it invokes our pass, but after it has loaded the shared
1679 object. The most foolproof way of doing this is to set a breakpoint in
1680 <tt>PassManager::run</tt> and then run the process with the arguments you
1683 <div class="doc_code"><pre>
1684 (gdb) <b>break llvm::PassManager::run</b>
1685 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1686 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]</b>
1687 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
1688 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1689 70 bool PassManager::run(Module &M) { return PM->run(M); }
1693 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1694 now free to set breakpoints in your pass so that you can trace through execution
1695 or do other standard debugging stuff.</p>
1699 <!-- _______________________________________________________________________ -->
1700 <div class="doc_subsubsection">
1701 <a name="debugmisc">Miscellaneous Problems</a>
1704 <div class="doc_text">
1706 <p>Once you have the basics down, there are a couple of problems that GDB has,
1707 some with solutions, some without.</p>
1710 <li>Inline functions have bogus stack information. In general, GDB does a
1711 pretty good job getting stack traces and stepping through inline functions.
1712 When a pass is dynamically loaded however, it somehow completely loses this
1713 capability. The only solution I know of is to de-inline a function (move it
1714 from the body of a class to a .cpp file).</li>
1716 <li>Restarting the program breaks breakpoints. After following the information
1717 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1718 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1719 and you start getting errors about breakpoints being unsettable. The only way I
1720 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1721 already set in your pass, run the program, and re-set the breakpoints once
1722 execution stops in <tt>PassManager::run</tt>.</li>
1726 <p>Hopefully these tips will help with common case debugging situations. If
1727 you'd like to contribute some tips of your own, just contact <a
1728 href="mailto:sabre@nondot.org">Chris</a>.</p>
1732 <!-- *********************************************************************** -->
1733 <div class="doc_section">
1734 <a name="future">Future extensions planned</a>
1736 <!-- *********************************************************************** -->
1738 <div class="doc_text">
1740 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1741 some nifty stuff, there are things we'd like to add in the future. Here is
1742 where we are going:</p>
1746 <!-- _______________________________________________________________________ -->
1747 <div class="doc_subsubsection">
1748 <a name="SMP">Multithreaded LLVM</a>
1751 <div class="doc_text">
1753 <p>Multiple CPU machines are becoming more common and compilation can never be
1754 fast enough: obviously we should allow for a multithreaded compiler. Because of
1755 the semantics defined for passes above (specifically they cannot maintain state
1756 across invocations of their <tt>run*</tt> methods), a nice clean way to
1757 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1758 to create multiple instances of each pass object, and allow the separate
1759 instances to be hacking on different parts of the program at the same time.</p>
1761 <p>This implementation would prevent each of the passes from having to implement
1762 multithreaded constructs, requiring only the LLVM core to have locking in a few
1763 places (for global resources). Although this is a simple extension, we simply
1764 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1765 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1769 <!-- _______________________________________________________________________ -->
1770 <div class="doc_subsubsection">
1771 <a name="PassFunctionPass"><tt>ModulePass</tt>es requiring <tt>FunctionPass</tt>es</a>
1774 <div class="doc_text">
1776 <p>Currently it is illegal for a <a href="#ModulePass"><tt>ModulePass</tt></a>
1777 to require a <a href="#FunctionPass"><tt>FunctionPass</tt></a>. This is because
1778 there is only one instance of the <a
1779 href="#FunctionPass"><tt>FunctionPass</tt></a> object ever created, thus nowhere
1780 to store information for all of the functions in the program at the same time.
1781 Although this has come up a couple of times before, this has always been worked
1782 around by factoring one big complicated pass into a global and an
1783 interprocedural part, both of which are distinct. In the future, it would be
1784 nice to have this though.</p>
1786 <p>Note that it is no problem for a <a
1787 href="#FunctionPass"><tt>FunctionPass</tt></a> to require the results of a <a
1788 href="#ModulePass"><tt>ModulePass</tt></a>, only the other way around.</p>
1792 <!-- *********************************************************************** -->
1795 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1796 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1797 <a href="http://validator.w3.org/check/referer"><img
1798 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!" /></a>
1800 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1801 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1802 Last modified: $Date$