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="#BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
48 <li><a href="#doInitialization_fn">The <tt>doInitialization(Function
49 &)</tt> method</a></li>
50 <li><a href="#runOnBasicBlock">The <tt>runOnBasicBlock</tt>
52 <li><a href="#doFinalization_fn">The <tt>doFinalization(Function
53 &)</tt> method</a></li>
55 <li><a href="#MachineFunctionPass">The <tt>MachineFunctionPass</tt>
58 <li><a href="#runOnMachineFunction">The
59 <tt>runOnMachineFunction(MachineFunction &)</tt> method</a></li>
62 <li><a href="#registration">Pass Registration</a>
64 <li><a href="#print">The <tt>print</tt> method</a></li>
66 <li><a href="#interaction">Specifying interactions between passes</a>
68 <li><a href="#getAnalysisUsage">The <tt>getAnalysisUsage</tt>
70 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
71 <li><a href="#AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a></li>
72 <li><a href="#AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a></li>
73 <li><a href="#getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a></li>
75 <li><a href="#analysisgroup">Implementing Analysis Groups</a>
77 <li><a href="#agconcepts">Analysis Group Concepts</a></li>
78 <li><a href="#registerag">Using <tt>RegisterAnalysisGroup</tt></a></li>
80 <li><a href="#passStatistics">Pass Statistics</a>
81 <li><a href="#passmanager">What PassManager does</a>
83 <li><a href="#releaseMemory">The <tt>releaseMemory</tt> method</a></li>
85 <li><a href="#registering">Registering dynamically loaded passes</a>
87 <li><a href="#registering_existing">Using existing registries</a></li>
88 <li><a href="#registering_new">Creating new registries</a></li>
90 <li><a href="#debughints">Using GDB with dynamically loaded passes</a>
92 <li><a href="#breakpoint">Setting a breakpoint in your pass</a></li>
93 <li><a href="#debugmisc">Miscellaneous Problems</a></li>
95 <li><a href="#future">Future extensions planned</a>
97 <li><a href="#SMP">Multithreaded LLVM</a></li>
98 <li><a href="#PassFunctionPass"><tt>ModulePass</tt>es requiring
99 <tt>FunctionPass</tt>es</a></li>
103 <div class="doc_author">
104 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
105 <a href="mailto:jlaskey@apple.com">Jim Laskey</a></p>
108 <!-- *********************************************************************** -->
109 <div class="doc_section">
110 <a name="introduction">Introduction - What is a pass?</a>
112 <!-- *********************************************************************** -->
114 <div class="doc_text">
116 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
117 passes are where most of the interesting parts of the compiler exist. Passes
118 perform the transformations and optimizations that make up the compiler, they
119 build the analysis results that are used by these transformations, and they are,
120 above all, a structuring technique for compiler code.</p>
122 <p>All LLVM passes are subclasses of the <tt><a
123 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
124 class, which implement functionality by overriding virtual methods inherited
125 from <tt>Pass</tt>. Depending on how your pass works, you should inherit from
126 the <tt><a href="#ModulePass">ModulePass</a></tt>, <tt><a
127 href="#CallGraphSCCPass">CallGraphSCCPass</a></tt>, <tt><a
128 href="#FunctionPass">FunctionPass</a></tt>, or <tt><a
129 href="#BasicBlockPass">BasicBlockPass</a></tt> classes, which gives the system
130 more information about what your pass does, and how it can be combined with
131 other passes. One of the main features of the LLVM Pass Framework is that it
132 schedules passes to run in an efficient way based on the constraints that your
133 pass meets (which are indicated by which class they derive from).</p>
135 <p>We start by showing you how to construct a pass, everything from setting up
136 the code, to compiling, loading, and executing it. After the basics are down,
137 more advanced features are discussed.</p>
141 <!-- *********************************************************************** -->
142 <div class="doc_section">
143 <a name="quickstart">Quick Start - Writing hello world</a>
145 <!-- *********************************************************************** -->
147 <div class="doc_text">
149 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
150 is designed to simply print out the name of non-external functions that exist in
151 the program being compiled. It does not modify the program at all, it just
152 inspects it. The source code and files for this pass are available in the LLVM
153 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
157 <!-- ======================================================================= -->
158 <div class="doc_subsection">
159 <a name="makefile">Setting up the build environment</a>
162 <div class="doc_text">
164 <p>First, you need to create a new directory somewhere in the LLVM source
165 base. For this example, we'll assume that you made
166 <tt>lib/Transforms/Hello</tt>. Next, you must set up a build script
167 (Makefile) that will compile the source code for the new pass. To do this,
168 copy the following into <tt>Makefile</tt>:</p>
171 <div class="doc_code"><pre>
172 # Makefile for hello pass
174 # Path to top level of LLVM heirarchy
177 # Name of the library to build
180 # Make the shared library become a loadable module so the tools can
181 # dlopen/dlsym on the resulting library.
184 # Tell the build system which LLVM libraries your pass needs. You'll probably
185 # need at least LLVMSystem.a, LLVMSupport.a, LLVMCore.a but possibly several
187 LLVMLIBS = LLVMCore.a LLVMSupport.a LLVMSystem.a
189 # Include the makefile implementation stuff
190 include $(LEVEL)/Makefile.common
193 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
194 directory are to be compiled and linked together into a
195 <tt>Debug/lib/Hello.so</tt> shared object that can be dynamically loaded by
196 the <tt>opt</tt> or <tt>bugpoint</tt> tools via their <tt>-load</tt> options.
197 If your operating system uses a suffix other than .so (such as windows or
198 Mac OS/X), the appropriate extension will be used.</p>
200 <p>Now that we have the build scripts set up, we just need to write the code for
205 <!-- ======================================================================= -->
206 <div class="doc_subsection">
207 <a name="basiccode">Basic code required</a>
210 <div class="doc_text">
212 <p>Now that we have a way to compile our new pass, we just have to write it.
215 <div class="doc_code"><pre>
216 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
217 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
220 <p>Which are needed because we are writing a <tt><a
221 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">Pass</a></tt>, and
222 we are operating on <tt><a
223 href="http://llvm.org/doxygen/classllvm_1_1Function.html">Function</a></tt>'s.</p>
226 <div class="doc_code"><pre>
227 <b>using namespace llvm;</b>
229 <p>... which is required because the functions from the include files
230 live in the llvm namespace.
235 <div class="doc_code"><pre>
239 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
240 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
241 things declared inside of the anonymous namespace only visible to the current
242 file. If you're not familiar with them, consult a decent C++ book for more
245 <p>Next, we declare our pass itself:</p>
247 <div class="doc_code"><pre>
248 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
251 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
252 href="http://llvm.org/doxygen/classllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
253 The different builtin pass subclasses are described in detail <a
254 href="#passtype">later</a>, but for now, know that <a
255 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate a function at a
258 <div class="doc_code"><pre>
259 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
260 llvm::cerr << "<i>Hello: </i>" << F.getName() << "\n";
263 }; <i>// end of struct Hello</i>
266 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
267 which overloads an abstract virtual method inherited from <a
268 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
269 to do our thing, so we just print out our message with the name of each
272 <div class="doc_code"><pre>
273 RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
274 } <i>// end of anonymous namespace</i>
277 <p>Lastly, we <a href="#registration">register our class</a> <tt>Hello</tt>,
278 giving it a command line
279 argument "<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>".</p>
281 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
283 <div class="doc_code"><pre>
284 <b>#include</b> "<a href="http://llvm.org/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
285 <b>#include</b> "<a href="http://llvm.org/doxygen/Function_8h-source.html">llvm/Function.h</a>"
287 <b>using namespace llvm;</b>
290 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
291 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
292 llvm::cerr << "<i>Hello: </i>" << F.getName() << "\n";
297 RegisterPass<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
301 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
302 command in the local directory and you should get a new
303 "<tt>Debug/lib/Hello.so</tt> file. Note that everything in this file is
304 contained in an anonymous namespace: this reflects the fact that passes are self
305 contained units that do not need external interfaces (although they can have
306 them) to be useful.</p>
310 <!-- ======================================================================= -->
311 <div class="doc_subsection">
312 <a name="running">Running a pass with <tt>opt</tt></a>
315 <div class="doc_text">
317 <p>Now that you have a brand new shiny shared object file, we can use the
318 <tt>opt</tt> command to run an LLVM program through your pass. Because you
319 registered your pass with the <tt>RegisterPass</tt> template, you will be able to
320 use the <tt>opt</tt> tool to access it, once loaded.</p>
322 <p>To test it, follow the example at the end of the <a
323 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
324 LLVM. We can now run the bytecode file (<tt>hello.bc</tt>) for the program
325 through our transformation like this (or course, any bytecode file will
328 <div class="doc_code"><pre>
329 $ opt -load ../../../Debug/lib/Hello.so -hello < hello.bc > /dev/null
335 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
336 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
337 argument (which is one reason you need to <a href="#registration">register your
338 pass</a>). Because the hello pass does not modify the program in any
339 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
340 <tt>/dev/null</tt>).</p>
342 <p>To see what happened to the other string you registered, try running
343 <tt>opt</tt> with the <tt>--help</tt> option:</p>
345 <div class="doc_code"><pre>
346 $ opt -load ../../../Debug/lib/Hello.so --help
347 OVERVIEW: llvm .bc -> .bc modular optimizer
349 USAGE: opt [options] <input bytecode>
352 Optimizations available:
354 -funcresolve - Resolve Functions
355 -gcse - Global Common Subexpression Elimination
356 -globaldce - Dead Global Elimination
357 <b>-hello - Hello World Pass</b>
358 -indvars - Canonicalize Induction Variables
359 -inline - Function Integration/Inlining
360 -instcombine - Combine redundant instructions
364 <p>The pass name get added as the information string for your pass, giving some
365 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
366 would go ahead and make it do the cool transformations you want. Once you get
367 it all working and tested, it may become useful to find out how fast your pass
368 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
369 line option (<tt>--time-passes</tt>) that allows you to get information about
370 the execution time of your pass along with the other passes you queue up. For
373 <div class="doc_code"><pre>
374 $ opt -load ../../../Debug/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
378 ===============================================================================
379 ... Pass execution timing report ...
380 ===============================================================================
381 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
383 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
384 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bytecode Writer
385 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
386 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
387 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
388 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
391 <p>As you can see, our implementation above is pretty fast :). The additional
392 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
393 that the LLVM emitted by your pass is still valid and well formed LLVM, which
394 hasn't been broken somehow.</p>
396 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
397 about some more details of how they work and how to use them.</p>
401 <!-- *********************************************************************** -->
402 <div class="doc_section">
403 <a name="passtype">Pass classes and requirements</a>
405 <!-- *********************************************************************** -->
407 <div class="doc_text">
409 <p>One of the first things that you should do when designing a new pass is to
410 decide what class you should subclass for your pass. The <a
411 href="#basiccode">Hello World</a> example uses the <tt><a
412 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
413 did not discuss why or when this should occur. Here we talk about the classes
414 available, from the most general to the most specific.</p>
416 <p>When choosing a superclass for your Pass, you should choose the <b>most
417 specific</b> class possible, while still being able to meet the requirements
418 listed. This gives the LLVM Pass Infrastructure information necessary to
419 optimize how passes are run, so that the resultant compiler isn't unneccesarily
424 <!-- ======================================================================= -->
425 <div class="doc_subsection">
426 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
429 <div class="doc_text">
431 <p>The most plain and boring type of pass is the "<tt><a
432 href="http://llvm.org/doxygen/classllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
433 class. This pass type is used for passes that do not have to be run, do not
434 change state, and never need to be updated. This is not a normal type of
435 transformation or analysis, but can provide information about the current
436 compiler configuration.</p>
438 <p>Although this pass class is very infrequently used, it is important for
439 providing information about the current target machine being compiled for, and
440 other static information that can affect the various transformations.</p>
442 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
443 invalidated, and are never "run".</p>
447 <!-- ======================================================================= -->
448 <div class="doc_subsection">
449 <a name="ModulePass">The <tt>ModulePass</tt> class</a>
452 <div class="doc_text">
455 href="http://llvm.org/doxygen/classllvm_1_1ModulePass.html">ModulePass</a></tt>"
456 class is the most general of all superclasses that you can use. Deriving from
457 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
458 refering to function bodies in no predictable order, or adding and removing
459 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
460 subclasses, no optimization can be done for their execution.</p>
462 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
463 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
464 following signature:</p>
468 <!-- _______________________________________________________________________ -->
469 <div class="doc_subsubsection">
470 <a name="runOnModule">The <tt>runOnModule</tt> method</a>
473 <div class="doc_text">
475 <div class="doc_code"><pre>
476 <b>virtual bool</b> runOnModule(Module &M) = 0;
479 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
480 It should return true if the module was modified by the transformation and
485 <!-- ======================================================================= -->
486 <div class="doc_subsection">
487 <a name="CallGraphSCCPass">The <tt>CallGraphSCCPass</tt> class</a>
490 <div class="doc_text">
493 href="http://llvm.org/doxygen/classllvm_1_1CallGraphSCCPass.html">CallGraphSCCPass</a></tt>"
494 is used by passes that need to traverse the program bottom-up on the call graph
495 (callees before callers). Deriving from CallGraphSCCPass provides some
496 mechanics for building and traversing the CallGraph, but also allows the system
497 to optimize execution of CallGraphSCCPass's. If your pass meets the
498 requirements outlined below, and doesn't meet the requirements of a <tt><a
499 href="#FunctionPass">FunctionPass</a></tt> or <tt><a
500 href="#BasicBlockPass">BasicBlockPass</a></tt>, you should derive from
501 <tt>CallGraphSCCPass</tt>.</p>
503 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
505 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
509 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
510 the current SCC.</li>
512 <li>... <em>allowed</em> to inspect any Function's other than those in the
513 current SCC and the direct callees of the SCC.</li>
515 <li>... <em>required</em> to preserve the current CallGraph object, updating it
516 to reflect any changes made to the program.</li>
518 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
519 though they may change the contents of an SCC.</li>
521 <li>... <em>allowed</em> to add or remove global variables from the current
524 <li>... <em>allowed</em> to maintain state across invocations of
525 <a href="#runOnSCC"><tt>runOnSCC</tt></a> (including global data).</li>
528 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
529 because it has to handle SCCs with more than one node in it. All of the virtual
530 methods described below should return true if they modified the program, or
531 false if they didn't.</p>
535 <!-- _______________________________________________________________________ -->
536 <div class="doc_subsubsection">
537 <a name="doInitialization_scc">The <tt>doInitialization(CallGraph &)</tt>
541 <div class="doc_text">
543 <div class="doc_code"><pre>
544 <b>virtual bool</b> doInitialization(CallGraph &CG);
547 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
548 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
549 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
550 is designed to do simple initialization type of stuff that does not depend on
551 the SCCs being processed. The <tt>doInitialization</tt> method call is not
552 scheduled to overlap with any other pass executions (thus it should be very
557 <!-- _______________________________________________________________________ -->
558 <div class="doc_subsubsection">
559 <a name="runOnSCC">The <tt>runOnSCC</tt> method</a>
562 <div class="doc_text">
564 <div class="doc_code"><pre>
565 <b>virtual bool</b> runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;
568 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
569 should return true if the module was modified by the transformation, false
574 <!-- _______________________________________________________________________ -->
575 <div class="doc_subsubsection">
576 <a name="doFinalization_scc">The <tt>doFinalization(CallGraph
577 &)</tt> method</a>
580 <div class="doc_text">
582 <div class="doc_code"><pre>
583 <b>virtual bool</b> doFinalization(CallGraph &CG);
586 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
587 called when the pass framework has finished calling <a
588 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
589 program being compiled.</p>
593 <!-- ======================================================================= -->
594 <div class="doc_subsection">
595 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
598 <div class="doc_text">
600 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
601 href="http://llvm.org/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
602 subclasses do have a predictable, local behavior that can be expected by the
603 system. All <tt>FunctionPass</tt> execute on each function in the program
604 independent of all of the other functions in the program.
605 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
606 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
608 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
611 <li>Modify a Function other than the one currently being processed.</li>
612 <li>Add or remove Function's from the current Module.</li>
613 <li>Add or remove global variables from the current Module.</li>
614 <li>Maintain state across invocations of
615 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
618 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
619 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
620 may overload three virtual methods to do their work. All of these methods
621 should return true if they modified the program, or false if they didn't.</p>
625 <!-- _______________________________________________________________________ -->
626 <div class="doc_subsubsection">
627 <a name="doInitialization_mod">The <tt>doInitialization(Module &)</tt>
631 <div class="doc_text">
633 <div class="doc_code"><pre>
634 <b>virtual bool</b> doInitialization(Module &M);
637 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
638 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
639 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
640 is designed to do simple initialization type of stuff that does not depend on
641 the functions being processed. The <tt>doInitialization</tt> method call is not
642 scheduled to overlap with any other pass executions (thus it should be very
645 <p>A good example of how this method should be used is the <a
646 href="http://llvm.org/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
647 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
648 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
649 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
650 free functions that it needs, adding prototypes to the module if necessary.</p>
654 <!-- _______________________________________________________________________ -->
655 <div class="doc_subsubsection">
656 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
659 <div class="doc_text">
661 <div class="doc_code"><pre>
662 <b>virtual bool</b> runOnFunction(Function &F) = 0;
665 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
666 the transformation or analysis work of your pass. As usual, a true value should
667 be returned if the function is modified.</p>
671 <!-- _______________________________________________________________________ -->
672 <div class="doc_subsubsection">
673 <a name="doFinalization_mod">The <tt>doFinalization(Module
674 &)</tt> method</a>
677 <div class="doc_text">
679 <div class="doc_code"><pre>
680 <b>virtual bool</b> doFinalization(Module &M);
683 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
684 called when the pass framework has finished calling <a
685 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
686 program being compiled.</p>
690 <!-- ======================================================================= -->
691 <div class="doc_subsection">
692 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
695 <div class="doc_text">
697 <p><tt>BasicBlockPass</tt>'s are just like <a
698 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
699 their scope of inspection and modification to a single basic block at a time.
700 As such, they are <b>not</b> allowed to do any of the following:</p>
703 <li>Modify or inspect any basic blocks outside of the current one</li>
704 <li>Maintain state across invocations of
705 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
706 <li>Modify the control flow graph (by altering terminator instructions)</li>
707 <li>Any of the things forbidden for
708 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
711 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
712 optimizations. They may override the same <a
713 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
714 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
715 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
719 <!-- _______________________________________________________________________ -->
720 <div class="doc_subsubsection">
721 <a name="doInitialization_fn">The <tt>doInitialization(Function
722 &)</tt> method</a>
725 <div class="doc_text">
727 <div class="doc_code"><pre>
728 <b>virtual bool</b> doInitialization(Function &F);
731 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
732 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
733 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
734 to do simple initialization that does not depend on the
735 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
736 scheduled to overlap with any other pass executions (thus it should be very
741 <!-- _______________________________________________________________________ -->
742 <div class="doc_subsubsection">
743 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
746 <div class="doc_text">
748 <div class="doc_code"><pre>
749 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
752 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
753 function is not allowed to inspect or modify basic blocks other than the
754 parameter, and are not allowed to modify the CFG. A true value must be returned
755 if the basic block is modified.</p>
759 <!-- _______________________________________________________________________ -->
760 <div class="doc_subsubsection">
761 <a name="doFinalization_fn">The <tt>doFinalization(Function &)</tt>
765 <div class="doc_text">
767 <div class="doc_code"><pre>
768 <b>virtual bool</b> doFinalization(Function &F);
771 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
772 called when the pass framework has finished calling <a
773 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
774 program being compiled. This can be used to perform per-function
779 <!-- ======================================================================= -->
780 <div class="doc_subsection">
781 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
784 <div class="doc_text">
786 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
787 executes on the machine-dependent representation of each LLVM function in the
788 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
789 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
790 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
791 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
794 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
795 <li>Modify a MachineFunction other than the one currently being processed.</li>
796 <li>Add or remove MachineFunctions from the current Module.</li>
797 <li>Add or remove global variables from the current Module.</li>
798 <li>Maintain state across invocations of <a
799 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
805 <!-- _______________________________________________________________________ -->
806 <div class="doc_subsubsection">
807 <a name="runOnMachineFunction">The <tt>runOnMachineFunction(MachineFunction
808 &MF)</tt> method</a>
811 <div class="doc_text">
813 <div class="doc_code"><pre>
814 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
817 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
818 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
819 work of your <tt>MachineFunctionPass</tt>.</p>
821 <p>The <tt>runOnMachineFunction</tt> method is called on every
822 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
823 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
824 representation of the function. If you want to get at the LLVM <tt>Function</tt>
825 for the <tt>MachineFunction</tt> you're working on, use
826 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
827 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
828 <tt>MachineFunctionPass</tt>.</p>
832 <!-- *********************************************************************** -->
833 <div class="doc_section">
834 <a name="registration">Pass registration</a>
836 <!-- *********************************************************************** -->
838 <div class="doc_text">
840 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
841 pass registration works, and discussed some of the reasons that it is used and
842 what it does. Here we discuss how and why passes are registered.</p>
844 <p>As we saw above, passes are registered with the <b><tt>RegisterPass</tt></b>
845 template, which requires you to pass at least two
846 parameters. The first parameter is the name of the pass that is to be used on
847 the command line to specify that the pass should be added to a program (for
848 example, with <tt>opt</tt> or <tt>bugpoint</tt>). The second argument is the
849 name of the pass, which is to be used for the <tt>--help</tt> output of
851 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
853 <p>If you want your pass to be easily dumpable, you should
854 implement the virtual <tt>print</tt> method:</p>
858 <!-- _______________________________________________________________________ -->
859 <div class="doc_subsubsection">
860 <a name="print">The <tt>print</tt> method</a>
863 <div class="doc_text">
865 <div class="doc_code"><pre>
866 <b>virtual void</b> print(llvm::OStream &O, <b>const</b> Module *M) <b>const</b>;
869 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
870 a human readable version of the analysis results. This is useful for debugging
871 an analysis itself, as well as for other people to figure out how an analysis
872 works. Use the <tt>opt -analyze</tt> argument to invoke this method.</p>
874 <p>The <tt>llvm::OStream</tt> parameter specifies the stream to write the results on,
875 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
876 program that has been analyzed. Note however that this pointer may be null in
877 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
878 debugger), so it should only be used to enhance debug output, it should not be
883 <!-- *********************************************************************** -->
884 <div class="doc_section">
885 <a name="interaction">Specifying interactions between passes</a>
887 <!-- *********************************************************************** -->
889 <div class="doc_text">
891 <p>One of the main responsibilities of the <tt>PassManager</tt> is the make sure
892 that passes interact with each other correctly. Because <tt>PassManager</tt>
893 tries to <a href="#passmanager">optimize the execution of passes</a> it must
894 know how the passes interact with each other and what dependencies exist between
895 the various passes. To track this, each pass can declare the set of passes that
896 are required to be executed before the current pass, and the passes which are
897 invalidated by the current pass.</p>
899 <p>Typically this functionality is used to require that analysis results are
900 computed before your pass is run. Running arbitrary transformation passes can
901 invalidate the computed analysis results, which is what the invalidation set
902 specifies. If a pass does not implement the <tt><a
903 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
904 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
908 <!-- _______________________________________________________________________ -->
909 <div class="doc_subsubsection">
910 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
913 <div class="doc_text">
915 <div class="doc_code"><pre>
916 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
919 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
920 invalidated sets may be specified for your transformation. The implementation
921 should fill in the <tt><a
922 href="http://llvm.org/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
923 object with information about which passes are required and not invalidated. To
924 do this, a pass may call any of the following methods on the AnalysisUsage
928 <!-- _______________________________________________________________________ -->
929 <div class="doc_subsubsection">
930 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
933 <div class="doc_text">
935 If your pass requires a previous pass to be executed (an analysis for example),
936 it can use one of these methods to arrange for it to be run before your pass.
937 LLVM has many different types of analyses and passes that can be required,
938 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
939 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
940 be no critical edges in the CFG when your pass has been run.
944 Some analyses chain to other analyses to do their job. For example, an <a
945 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
946 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
947 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
948 used instead of the <tt>addRequired</tt> method. This informs the PassManager
949 that the transitively required pass should be alive as long as the requiring
954 <!-- _______________________________________________________________________ -->
955 <div class="doc_subsubsection">
956 <a name="AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a>
959 <div class="doc_text">
961 One of the jobs of the PassManager is to optimize how and when analyses are run.
962 In particular, it attempts to avoid recomputing data unless it needs to. For
963 this reason, passes are allowed to declare that they preserve (i.e., they don't
964 invalidate) an existing analysis if it's available. For example, a simple
965 constant folding pass would not modify the CFG, so it can't possibly affect the
966 results of dominator analysis. By default, all passes are assumed to invalidate
971 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
972 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
973 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
974 not modify the LLVM program at all (which is true for analyses), and the
975 <tt>setPreservesCFG</tt> method can be used by transformations that change
976 instructions in the program but do not modify the CFG or terminator instructions
977 (note that this property is implicitly set for <a
978 href="#BasicBlockPass">BasicBlockPass</a>'s).
982 <tt>addPreserved</tt> is particularly useful for transformations like
983 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
984 and dominator related analyses if they exist, so it can preserve them, despite
985 the fact that it hacks on the CFG.
989 <!-- _______________________________________________________________________ -->
990 <div class="doc_subsubsection">
991 <a name="AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a>
994 <div class="doc_text">
996 <div class="doc_code"><pre>
997 <i>// This is an example implementation from an analysis, which does not modify
998 // the program at all, yet has a prerequisite.</i>
999 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1000 AU.setPreservesAll();
1001 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1007 <div class="doc_code"><pre>
1008 <i>// This example modifies the program, but does not modify the CFG</i>
1009 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1010 AU.setPreservesCFG();
1011 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1017 <!-- _______________________________________________________________________ -->
1018 <div class="doc_subsubsection">
1019 <a name="getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a>
1022 <div class="doc_text">
1024 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1025 your class, providing you with access to the passes that you declared that you
1026 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
1027 method. It takes a single template argument that specifies which pass class you
1028 want, and returns a reference to that pass. For example:</p>
1030 <div class="doc_code"><pre>
1031 bool LICM::runOnFunction(Function &F) {
1032 LoopInfo &LI = getAnalysis<LoopInfo>();
1037 <p>This method call returns a reference to the pass desired. You may get a
1038 runtime assertion failure if you attempt to get an analysis that you did not
1039 declare as required in your <a
1040 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
1041 method can be called by your <tt>run*</tt> method implementation, or by any
1042 other local method invoked by your <tt>run*</tt> method.</p>
1045 If your pass is capable of updating analyses if they exist (e.g.,
1046 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1047 <tt>getAnalysisToUpdate</tt> method, which returns a pointer to the analysis if
1048 it is active. For example:</p>
1050 <div class="doc_code"><pre>
1052 if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
1053 <i>// A DominatorSet is active. This code will update it.</i>
1060 <!-- *********************************************************************** -->
1061 <div class="doc_section">
1062 <a name="analysisgroup">Implementing Analysis Groups</a>
1064 <!-- *********************************************************************** -->
1066 <div class="doc_text">
1068 <p>Now that we understand the basics of how passes are defined, how the are
1069 used, and how they are required from other passes, it's time to get a little bit
1070 fancier. All of the pass relationships that we have seen so far are very
1071 simple: one pass depends on one other specific pass to be run before it can run.
1072 For many applications, this is great, for others, more flexibility is
1075 <p>In particular, some analyses are defined such that there is a single simple
1076 interface to the analysis results, but multiple ways of calculating them.
1077 Consider alias analysis for example. The most trivial alias analysis returns
1078 "may alias" for any alias query. The most sophisticated analysis a
1079 flow-sensitive, context-sensitive interprocedural analysis that can take a
1080 significant amount of time to execute (and obviously, there is a lot of room
1081 between these two extremes for other implementations). To cleanly support
1082 situations like this, the LLVM Pass Infrastructure supports the notion of
1083 Analysis Groups.</p>
1087 <!-- _______________________________________________________________________ -->
1088 <div class="doc_subsubsection">
1089 <a name="agconcepts">Analysis Group Concepts</a>
1092 <div class="doc_text">
1094 <p>An Analysis Group is a single simple interface that may be implemented by
1095 multiple different passes. Analysis Groups can be given human readable names
1096 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1097 class. An analysis group may have one or more implementations, one of which is
1098 the "default" implementation.</p>
1100 <p>Analysis groups are used by client passes just like other passes are: the
1101 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1102 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
1103 scans the available passes to see if any implementations of the analysis group
1104 are available. If none is available, the default implementation is created for
1105 the pass to use. All standard rules for <A href="#interaction">interaction
1106 between passes</a> still apply.</p>
1108 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
1109 passes, all analysis group implementations must be registered, and must use the
1110 <A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
1111 implementation pool. Also, a default implementation of the interface
1112 <b>must</b> be registered with <A
1113 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
1115 <p>As a concrete example of an Analysis Group in action, consider the <a
1116 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1117 analysis group. The default implementation of the alias analysis interface (the
1119 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1120 pass) just does a few simple checks that don't require significant analysis to
1121 compute (such as: two different globals can never alias each other, etc).
1122 Passes that use the <tt><a
1123 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1124 interface (for example the <tt><a
1125 href="http://llvm.org/doxygen/structGCSE.html">gcse</a></tt> pass), do
1126 not care which implementation of alias analysis is actually provided, they just
1127 use the designated interface.</p>
1129 <p>From the user's perspective, commands work just like normal. Issuing the
1130 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1131 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1132 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1133 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1134 hypothetical example) instead.</p>
1138 <!-- _______________________________________________________________________ -->
1139 <div class="doc_subsubsection">
1140 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1143 <div class="doc_text">
1145 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1146 group itself as well as add pass implementations to the analysis group. First,
1147 an analysis should be registered, with a human readable name provided for it.
1148 Unlike registration of passes, there is no command line argument to be specified
1149 for the Analysis Group Interface itself, because it is "abstract":</p>
1151 <div class="doc_code"><pre>
1152 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1155 <p>Once the analysis is registered, passes can declare that they are valid
1156 implementations of the interface by using the following code:</p>
1158 <div class="doc_code"><pre>
1160 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1161 RegisterPass<FancyAA>
1162 B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
1164 //<i> Declare that we implement the AliasAnalysis interface</i>
1165 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> C(B);
1169 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1170 uses the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
1171 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1172 analysis group. Every implementation of an analysis group should join using
1173 this template. A single pass may join multiple different analysis groups with
1176 <div class="doc_code"><pre>
1178 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1179 RegisterPass<<a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1180 D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
1182 //<i> Declare that we implement the AliasAnalysis interface</i>
1183 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <b>true</b>> E(D);
1187 <p>Here we show how the default implementation is specified (using the extra
1188 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1189 one default implementation available at all times for an Analysis Group to be
1190 used. Here we declare that the <tt><a
1191 href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1192 pass is the default implementation for the interface.</p>
1196 <!-- *********************************************************************** -->
1197 <div class="doc_section">
1198 <a name="passStatistics">Pass Statistics</a>
1200 <!-- *********************************************************************** -->
1202 <div class="doc_text">
1204 href="http://llvm.org/doxygen/Statistic_8h-source.html"><tt>Statistic</tt></a>
1205 class is designed to be an easy way to expose various success
1206 metrics from passes. These statistics are printed at the end of a
1207 run, when the -stats command line option is enabled on the command
1208 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1213 <!-- *********************************************************************** -->
1214 <div class="doc_section">
1215 <a name="passmanager">What PassManager does</a>
1217 <!-- *********************************************************************** -->
1219 <div class="doc_text">
1222 href="http://llvm.org/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1224 href="http://llvm.org/doxygen/classllvm_1_1PassManager.html">class</a>
1225 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1226 are set up correctly, and then schedules passes to run efficiently. All of the
1227 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1230 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1231 time of a series of passes:</p>
1234 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1235 recomputing analysis results as much as possible. This means keeping track of
1236 which analyses are available already, which analyses get invalidated, and which
1237 analyses are needed to be run for a pass. An important part of work is that the
1238 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1239 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1240 results as soon as they are no longer needed.</li>
1242 <li><b>Pipeline the execution of passes on the program</b> - The
1243 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1244 of a series of passes by pipelining the passes together. This means that, given
1245 a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1246 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1247 the first function, then all of the <a
1248 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1249 etc... until the entire program has been run through the passes.
1251 <p>This improves the cache behavior of the compiler, because it is only touching
1252 the LLVM program representation for a single function at a time, instead of
1253 traversing the entire program. It reduces the memory consumption of compiler,
1254 because, for example, only one <a
1255 href="http://llvm.org/doxygen/classllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1256 needs to be calculated at a time. This also makes it possible some <a
1257 href="#SMP">interesting enhancements</a> in the future.</p></li>
1261 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1262 much information it has about the behaviors of the passes it is scheduling. For
1263 example, the "preserved" set is intentionally conservative in the face of an
1264 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1265 Not implementing when it should be implemented will have the effect of not
1266 allowing any analysis results to live across the execution of your pass.</p>
1268 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1269 options that is useful for debugging pass execution, seeing how things work, and
1270 diagnosing when you should be preserving more analyses than you currently are
1271 (To get information about all of the variants of the <tt>--debug-pass</tt>
1272 option, just type '<tt>opt --help-hidden</tt>').</p>
1274 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1275 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1276 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1278 <div class="doc_code"><pre>
1279 $ opt -load ../../../Debug/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1281 Function Pass Manager
1282 Dominator Set Construction
1283 Immediate Dominators Construction
1284 Global Common Subexpression Elimination
1285 -- Immediate Dominators Construction
1286 -- Global Common Subexpression Elimination
1287 Natural Loop Construction
1288 Loop Invariant Code Motion
1289 -- Natural Loop Construction
1290 -- Loop Invariant Code Motion
1292 -- Dominator Set Construction
1298 <p>This output shows us when passes are constructed and when the analysis
1299 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1300 GCSE uses dominator and immediate dominator information to do its job. The LICM
1301 pass uses natural loop information, which uses dominator sets, but not immediate
1302 dominators. Because immediate dominators are no longer useful after the GCSE
1303 pass, it is immediately destroyed. The dominator sets are then reused to
1304 compute natural loop information, which is then used by the LICM pass.</p>
1306 <p>After the LICM pass, the module verifier runs (which is automatically added
1307 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1308 resultant LLVM code is well formed. After it finishes, the dominator set
1309 information is destroyed, after being computed once, and shared by three
1312 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1313 World</a> pass in between the two passes:</p>
1315 <div class="doc_code"><pre>
1316 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1318 Function Pass Manager
1319 Dominator Set Construction
1320 Immediate Dominators Construction
1321 Global Common Subexpression Elimination
1322 <b>-- Dominator Set Construction</b>
1323 -- Immediate Dominators Construction
1324 -- Global Common Subexpression Elimination
1325 <b> Hello World Pass
1327 Dominator Set Construction</b>
1328 Natural Loop Construction
1329 Loop Invariant Code Motion
1330 -- Natural Loop Construction
1331 -- Loop Invariant Code Motion
1333 -- Dominator Set Construction
1342 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1343 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1344 we need to add the following <a
1345 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1347 <div class="doc_code"><pre>
1348 <i>// We don't modify the program, so we preserve all analyses</i>
1349 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1350 AU.setPreservesAll();
1354 <p>Now when we run our pass, we get this output:</p>
1356 <div class="doc_code"><pre>
1357 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1358 Pass Arguments: -gcse -hello -licm
1360 Function Pass Manager
1361 Dominator Set Construction
1362 Immediate Dominators Construction
1363 Global Common Subexpression Elimination
1364 -- Immediate Dominators Construction
1365 -- Global Common Subexpression Elimination
1368 Natural Loop Construction
1369 Loop Invariant Code Motion
1370 -- Loop Invariant Code Motion
1371 -- Natural Loop Construction
1373 -- Dominator Set Construction
1382 <p>Which shows that we don't accidentally invalidate dominator information
1383 anymore, and therefore do not have to compute it twice.</p>
1387 <!-- _______________________________________________________________________ -->
1388 <div class="doc_subsubsection">
1389 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1392 <div class="doc_text">
1394 <div class="doc_code"><pre>
1395 <b>virtual void</b> releaseMemory();
1398 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1399 results, and how long to keep them around for. Because the lifetime of the pass
1400 object itself is effectively the entire duration of the compilation process, we
1401 need some way to free analysis results when they are no longer useful. The
1402 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1404 <p>If you are writing an analysis or any other pass that retains a significant
1405 amount of state (for use by another pass which "requires" your pass and uses the
1406 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1407 <tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
1408 internal state. This method is called after the <tt>run*</tt> method for the
1409 class, before the next call of <tt>run*</tt> in your pass.</p>
1413 <!-- *********************************************************************** -->
1414 <div class="doc_section">
1415 <a name="registering">Registering dynamically loaded passes</a>
1417 <!-- *********************************************************************** -->
1419 <div class="doc_text">
1421 <p><i>Size matters</i> when constructing production quality tools using llvm,
1422 both for the purposes of distribution, and for regulating the resident code size
1423 when running on the target system. Therefore, it becomes desirable to
1424 selectively use some passes, while omitting others and maintain the flexibility
1425 to change configurations later on. You want to be able to do all this, and,
1426 provide feedback to the user. This is where pass registration comes into
1429 <p>The fundamental mechanisms for pass registration are the
1430 <tt>MachinePassRegistry</tt> class and subclasses of
1431 <tt>MachinePassRegistryNode</tt>.</p>
1433 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1434 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1435 communicates additions and deletions to the command line interface.</p>
1437 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1438 information provided about a particular pass. This information includes the
1439 command line name, the command help string and the address of the function used
1440 to create an instance of the pass. A global static constructor of one of these
1441 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1442 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1443 in the tool will be registered at start up. A dynamically loaded pass will
1444 register on load and unregister at unload.</p>
1448 <!-- _______________________________________________________________________ -->
1449 <div class="doc_subsection">
1450 <a name="registering_existing">Using existing registries</a>
1453 <div class="doc_text">
1455 <p>There are predefined registries to track instruction scheduling
1456 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1457 machine passes. Here we will describe how to <i>register</i> a register
1458 allocator machine pass.</p>
1460 <p>Implement your register allocator machine pass. In your register allocator
1461 .cpp file add the following include;</p>
1463 <div class="doc_code"><pre>
1464 #include "llvm/CodeGen/RegAllocRegistry.h"
1467 <p>Also in your register allocator .cpp file, define a creator function in the
1470 <div class="doc_code"><pre>
1471 FunctionPass *createMyRegisterAllocator() {
1472 return new MyRegisterAllocator();
1476 <p>Note that the signature of this function should match the type of
1477 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1478 "installing" declaration, in the form;</p>
1480 <div class="doc_code"><pre>
1481 static RegisterRegAlloc myRegAlloc("myregalloc",
1482 " my register allocator help string",
1483 createMyRegisterAllocator);
1486 <p>Note the two spaces prior to the help string produces a tidy result on the
1489 <div class="doc_code"><pre>
1492 -regalloc - Register allocator to use: (default = linearscan)
1493 =linearscan - linear scan register allocator
1494 =local - local register allocator
1495 =simple - simple register allocator
1496 =myregalloc - my register allocator help string
1500 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1501 an option. Registering instruction schedulers is similar except use the
1502 <tt>RegisterScheduler</tt> class. Note that the
1503 <tt>RegisterScheduler::FunctionPassCtor</tt> is significantly different from
1504 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1506 <p>To force the load/linking of your register allocator into the llc/lli tools,
1507 add your creator function's global declaration to "Passes.h" and add a "pseudo"
1508 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1513 <!-- _______________________________________________________________________ -->
1514 <div class="doc_subsection">
1515 <a name="registering_new">Creating new registries</a>
1518 <div class="doc_text">
1520 <p>The easiest way to get started is to clone one of the existing registries; we
1521 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1522 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1524 <p>Then you need to declare the registry. Example: if your pass registry is
1525 <tt>RegisterMyPasses</tt> then define;</p>
1527 <div class="doc_code"><pre>
1528 MachinePassRegistry RegisterMyPasses::Registry;
1531 <p>And finally, declare the command line option for your passes. Example:</p>
1533 <div class="doc_code"><pre>
1534 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1535 RegisterPassParser<RegisterMyPasses> >
1537 cl::init(&createDefaultMyPass),
1538 cl::desc("my pass option help"));
1541 <p>Here the command option is "mypass", with createDefaultMyPass as the default
1546 <!-- *********************************************************************** -->
1547 <div class="doc_section">
1548 <a name="debughints">Using GDB with dynamically loaded passes</a>
1550 <!-- *********************************************************************** -->
1552 <div class="doc_text">
1554 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1555 should be. First of all, you can't set a breakpoint in a shared object that has
1556 not been loaded yet, and second of all there are problems with inlined functions
1557 in shared objects. Here are some suggestions to debugging your pass with
1560 <p>For sake of discussion, I'm going to assume that you are debugging a
1561 transformation invoked by <tt>opt</tt>, although nothing described here depends
1566 <!-- _______________________________________________________________________ -->
1567 <div class="doc_subsubsection">
1568 <a name="breakpoint">Setting a breakpoint in your pass</a>
1571 <div class="doc_text">
1573 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1575 <div class="doc_code"><pre>
1578 Copyright 2000 Free Software Foundation, Inc.
1579 GDB is free software, covered by the GNU General Public License, and you are
1580 welcome to change it and/or distribute copies of it under certain conditions.
1581 Type "show copying" to see the conditions.
1582 There is absolutely no warranty for GDB. Type "show warranty" for details.
1583 This GDB was configured as "sparc-sun-solaris2.6"...
1587 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1588 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1589 (the shared object isn't loaded until runtime), we must execute the process, and
1590 have it stop before it invokes our pass, but after it has loaded the shared
1591 object. The most foolproof way of doing this is to set a breakpoint in
1592 <tt>PassManager::run</tt> and then run the process with the arguments you
1595 <div class="doc_code"><pre>
1596 (gdb) <b>break llvm::PassManager::run</b>
1597 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1598 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]</b>
1599 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
1600 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1601 70 bool PassManager::run(Module &M) { return PM->run(M); }
1605 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1606 now free to set breakpoints in your pass so that you can trace through execution
1607 or do other standard debugging stuff.</p>
1611 <!-- _______________________________________________________________________ -->
1612 <div class="doc_subsubsection">
1613 <a name="debugmisc">Miscellaneous Problems</a>
1616 <div class="doc_text">
1618 <p>Once you have the basics down, there are a couple of problems that GDB has,
1619 some with solutions, some without.</p>
1622 <li>Inline functions have bogus stack information. In general, GDB does a
1623 pretty good job getting stack traces and stepping through inline functions.
1624 When a pass is dynamically loaded however, it somehow completely loses this
1625 capability. The only solution I know of is to de-inline a function (move it
1626 from the body of a class to a .cpp file).</li>
1628 <li>Restarting the program breaks breakpoints. After following the information
1629 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1630 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1631 and you start getting errors about breakpoints being unsettable. The only way I
1632 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1633 already set in your pass, run the program, and re-set the breakpoints once
1634 execution stops in <tt>PassManager::run</tt>.</li>
1638 <p>Hopefully these tips will help with common case debugging situations. If
1639 you'd like to contribute some tips of your own, just contact <a
1640 href="mailto:sabre@nondot.org">Chris</a>.</p>
1644 <!-- *********************************************************************** -->
1645 <div class="doc_section">
1646 <a name="future">Future extensions planned</a>
1648 <!-- *********************************************************************** -->
1650 <div class="doc_text">
1652 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1653 some nifty stuff, there are things we'd like to add in the future. Here is
1654 where we are going:</p>
1658 <!-- _______________________________________________________________________ -->
1659 <div class="doc_subsubsection">
1660 <a name="SMP">Multithreaded LLVM</a>
1663 <div class="doc_text">
1665 <p>Multiple CPU machines are becoming more common and compilation can never be
1666 fast enough: obviously we should allow for a multithreaded compiler. Because of
1667 the semantics defined for passes above (specifically they cannot maintain state
1668 across invocations of their <tt>run*</tt> methods), a nice clean way to
1669 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1670 to create multiple instances of each pass object, and allow the separate
1671 instances to be hacking on different parts of the program at the same time.</p>
1673 <p>This implementation would prevent each of the passes from having to implement
1674 multithreaded constructs, requiring only the LLVM core to have locking in a few
1675 places (for global resources). Although this is a simple extension, we simply
1676 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1677 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1681 <!-- _______________________________________________________________________ -->
1682 <div class="doc_subsubsection">
1683 <a name="PassFunctionPass"><tt>ModulePass</tt>es requiring <tt>FunctionPass</tt>es</a>
1686 <div class="doc_text">
1688 <p>Currently it is illegal for a <a href="#ModulePass"><tt>ModulePass</tt></a>
1689 to require a <a href="#FunctionPass"><tt>FunctionPass</tt></a>. This is because
1690 there is only one instance of the <a
1691 href="#FunctionPass"><tt>FunctionPass</tt></a> object ever created, thus nowhere
1692 to store information for all of the functions in the program at the same time.
1693 Although this has come up a couple of times before, this has always been worked
1694 around by factoring one big complicated pass into a global and an
1695 interprocedural part, both of which are distinct. In the future, it would be
1696 nice to have this though.</p>
1698 <p>Note that it is no problem for a <a
1699 href="#FunctionPass"><tt>FunctionPass</tt></a> to require the results of a <a
1700 href="#ModulePass"><tt>ModulePass</tt></a>, only the other way around.</p>
1704 <!-- *********************************************************************** -->
1707 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1708 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1709 <a href="http://validator.w3.org/check/referer"><img
1710 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!" /></a>
1712 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1713 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1714 Last modified: $Date$