1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <title>Writing an LLVM Pass</title>
6 <link rel="stylesheet" href="llvm.css" type="text/css">
10 <div class="doc_title">
15 <li><a href="#introduction">Introduction - What is a pass?</a></li>
16 <li><a href="#quickstart">Quick Start - Writing hello world</a>
18 <li><a href="#makefile">Setting up the build environment</a></li>
19 <li><a href="#basiccode">Basic code required</a></li>
20 <li><a href="#running">Running a pass with <tt>opt</tt>
21 or <tt>analyze</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="#Pass">The <tt>Pass</tt> class</a>
28 <li><a href="#run">The <tt>run</tt> method</a></li>
30 <li><a href="#FunctionPass">The <tt>FunctionPass</tt> class</a>
32 <li><a href="#doInitialization_mod">The <tt>doInitialization(Module
33 &)</tt> method</a></li>
34 <li><a href="#runOnFunction">The <tt>runOnFunction</tt> method</a></li>
35 <li><a href="#doFinalization_mod">The <tt>doFinalization(Module
36 &)</tt> method</a></li>
38 <li><a href="#BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
40 <li><a href="#doInitialization_fn">The <tt>doInitialization(Function
41 &)</tt> method</a></li>
42 <li><a href="#runOnBasicBlock">The <tt>runOnBasicBlock</tt>
44 <li><a href="#doFinalization_fn">The <tt>doFinalization(Function
45 &)</tt> method</a></li>
47 <li><a href="#MachineFunctionPass">The <tt>MachineFunctionPass</tt>
50 <li><a href="#runOnMachineFunction">The
51 <tt>runOnMachineFunction(MachineFunction &)</tt> method</a></li>
54 <li><a href="#registration">Pass Registration</a>
56 <li><a href="#print">The <tt>print</tt> method</a></li>
58 <li><a href="#interaction">Specifying interactions between passes</a>
60 <li><a href="#getAnalysisUsage">The <tt>getAnalysisUsage</tt>
62 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
63 <li><a href="#AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a></li>
64 <li><a href="#AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a></li>
65 <li><a href="#getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a></li>
67 <li><a href="#analysisgroup">Implementing Analysis Groups</a>
69 <li><a href="#agconcepts">Analysis Group Concepts</a></li>
70 <li><a href="#registerag">Using <tt>RegisterAnalysisGroup</tt></a></li>
72 <li><a href="#passmanager">What PassManager does</a>
74 <li><a href="#releaseMemory">The <tt>releaseMemory</tt> method</a></li>
76 <li><a href="#debughints">Using GDB with dynamically loaded passes</a>
78 <li><a href="#breakpoint">Setting a breakpoint in your pass</a></li>
79 <li><a href="#debugmisc">Miscellaneous Problems</a></li>
81 <li><a href="#future">Future extensions planned</a>
83 <li><a href="#SMP">Multithreaded LLVM</a></li>
84 <li><a href="#ModuleSource">A new <tt>ModuleSource</tt> interface</a></li>
85 <li><a href="#PassFunctionPass"><tt>Pass</tt>es requiring
86 <tt>FunctionPass</tt>es</a></li>
90 <div class="doc_author">
91 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
94 <!-- *********************************************************************** -->
95 <div class="doc_section">
96 <a name="introduction">Introduction - What is a pass?</a>
98 <!-- *********************************************************************** -->
100 <div class="doc_text">
102 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
103 passes are where the interesting parts of the compiler exist. Passes perform
104 the transformations and optimizations that make up the compiler, they build
105 the analysis results that are used by these transformations, and they are, above
106 all, a structuring technique for compiler code.</p>
108 <p>All LLVM passes are subclasses of the <tt><a
109 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">Pass</a></tt>
110 class, which implement functionality by overriding virtual methods inherited
111 from <tt>Pass</tt>. Depending on how your pass works, you may be able to
112 inherit from the <tt><a
113 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1FunctionPass.html">FunctionPass</a></tt>
115 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1BasicBlockPass.html">BasicBlockPass</a></tt>,
116 which gives the system more information about what your pass does, and how it
117 can be combined with other passes. One of the main features of the LLVM Pass
118 Framework is that it schedules passes to run in an efficient way based on the
119 constraints that your pass has.</p>
121 <p>We start by showing you how to construct a pass, everything from setting up
122 the code, to compiling, loading, and executing it. After the basics are down,
123 more advanced features are discussed.</p>
127 <!-- *********************************************************************** -->
128 <div class="doc_section">
129 <a name="quickstart">Quick Start - Writing hello world</a>
131 <!-- *********************************************************************** -->
133 <div class="doc_text">
135 <p>Here we describe how to write the "hello world" of passes. The "Hello" pass
136 is designed to simply print out the name of non-external functions that exist in
137 the program being compiled. It does not modify the program at all, just
138 inspects it. The source code and files for this pass are available in the LLVM
139 source tree in the <tt>lib/Transforms/Hello</tt> directory.</p>
143 <!-- ======================================================================= -->
144 <div class="doc_subsection">
145 <a name="makefile">Setting up the build environment</a>
148 <div class="doc_text">
150 <p>First thing you need to do is create a new directory somewhere in the LLVM
151 source base. For this example, we'll assume that you made
152 "<tt>lib/Transforms/Hello</tt>". The first thing you must do is set up a build
153 script (Makefile) that will compile the source code for the new pass. To do
154 this, copy this into "<tt>Makefile</tt>":</p>
159 # Makefile for hello pass
161 # Path to top level of LLVM heirarchy
164 # Name of the library to build
167 # Build a dynamically loadable shared object
170 # Include the makefile implementation stuff
171 include $(LEVEL)/Makefile.common
174 <p>This makefile specifies that all of the <tt>.cpp</tt> files in the current
175 directory are to be compiled and linked together into a
176 <tt>lib/Debug/libhello.so</tt> shared object that can be dynamically loaded by
177 the <tt>opt</tt> or <tt>analyze</tt> tools. If your operating system uses a
178 suffix other than .so (such as windows of Mac OS/X), the appropriate extension
181 <p>Now that we have the build scripts set up, we just need to write the code for
186 <!-- ======================================================================= -->
187 <div class="doc_subsection">
188 <a name="basiccode">Basic code required</a>
191 <div class="doc_text">
193 <p>Now that we have a way to compile our new pass, we just have to write it.
197 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
198 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Function_8h-source.html">llvm/Function.h</a>"
201 <p>Which are needed because we are writing a <tt><a
202 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">Pass</a></tt>, and
203 we are operating on <tt><a
204 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Function.html">Function</a></tt>'s.</p>
212 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
213 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
214 things declared inside of the anonymous namespace only visible to the current
215 file. If you're not familiar with them, consult a decent C++ book for more
218 <p>Next, we declare our pass itself:</p>
221 <b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
224 <p>This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
225 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1FunctionPass.html">FunctionPass</a></tt>.
226 The different builtin pass subclasses are described in detail <a
227 href="#passtype">later</a>, but for now, know that <a
228 href="#FunctionPass"><tt>FunctionPass</tt></a>'s operate a function at a
232 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
233 std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
236 }; <i>// end of struct Hello</i>
239 <p>We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method,
240 which overloads an abstract virtual method inherited from <a
241 href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
242 to do our thing, so we just print out our message with the name of each
246 RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
247 } <i>// end of anonymous namespace</i>
250 <p>Lastly, we register our class <tt>Hello</tt>, giving it a command line
251 argument "<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>". There are
252 several different ways of <a href="#registration">registering your pass</a>,
253 depending on what it is to be used for. For "optimizations" we use the
254 <tt>RegisterOpt</tt> template.</p>
256 <p>As a whole, the <tt>.cpp</tt> file looks like:</p>
259 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
260 <b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Function_8h-source.html">llvm/Function.h</a>"
263 <b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
264 <b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
265 std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
270 RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
274 <p>Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
275 command in the local directory and you should get a new
276 "<tt>lib/Debug/libhello.so</tt> file. Note that everything in this file is
277 contained in an anonymous namespace: this reflects the fact that passes are self
278 contained units that do not need external interfaces (although they can have
279 them) to be useful.</p>
283 <!-- ======================================================================= -->
284 <div class="doc_subsection">
285 <a name="running">Running a pass with <tt>opt</tt> or <tt>analyze</tt></a>
288 <div class="doc_text">
290 <p>Now that you have a brand new shiny shared object file, we can use the
291 <tt>opt</tt> command to run an LLVM program through your pass. Because you
292 registered your pass with the <tt>RegisterOpt</tt> template, you will be able to
293 use the <tt>opt</tt> tool to access it, once loaded.</p>
295 <p>To test it, follow the example at the end of the <a
296 href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
297 LLVM. We can now run the bytecode file (<tt>hello.bc</tt>) for the program
298 through our transformation like this (or course, any bytecode file will
302 $ opt -load ../../../lib/Debug/libhello.so -hello < hello.bc > /dev/null
308 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
309 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
310 argument (which is one reason you need to <a href="#registration">register your
311 pass</a>). Because the hello pass does not modify the program in any
312 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
313 <tt>/dev/null</tt>).</p>
315 <p>To see what happened to the other string you registered, try running
316 <tt>opt</tt> with the <tt>--help</tt> option:</p>
319 $ opt -load ../../../lib/Debug/libhello.so --help
320 OVERVIEW: llvm .bc -> .bc modular optimizer
322 USAGE: opt [options] <input bytecode>
325 Optimizations available:
327 -funcresolve - Resolve Functions
328 -gcse - Global Common Subexpression Elimination
329 -globaldce - Dead Global Elimination
330 <b>-hello - Hello World Pass</b>
331 -indvars - Canonicalize Induction Variables
332 -inline - Function Integration/Inlining
333 -instcombine - Combine redundant instructions
337 <p>The pass name get added as the information string for your pass, giving some
338 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
339 would go ahead and make it do the cool transformations you want. Once you get
340 it all working and tested, it may become useful to find out how fast your pass
341 is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
342 line option (<tt>--time-passes</tt>) that allows you to get information about
343 the execution time of your pass along with the other passes you queue up. For
347 $ opt -load ../../../lib/Debug/libhello.so -hello -time-passes < hello.bc > /dev/null
351 ===============================================================================
352 ... Pass execution timing report ...
353 ===============================================================================
354 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
356 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
357 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bytecode Writer
358 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
359 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
360 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
361 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
364 <p>As you can see, our implementation above is pretty fast :). The additional
365 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
366 that the LLVM emitted by your pass is still valid and well formed LLVM, which
367 hasn't been broken somehow.</p>
369 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
370 about some more details of how they work and how to use them.</p>
374 <!-- *********************************************************************** -->
375 <div class="doc_section">
376 <a name="passtype">Pass classes and requirements</a>
378 <!-- *********************************************************************** -->
380 <div class="doc_text">
382 <p>One of the first things that you should do when designing a new pass is to
383 decide what class you should subclass for your pass. The <a
384 href="#basiccode">Hello World</a> example uses the <tt><a
385 href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
386 did not discuss why or when this should occur. Here we talk about the classes
387 available, from the most general to the most specific.</p>
389 <p>When choosing a superclass for your Pass, you should choose the <b>most
390 specific</b> class possible, while still being able to meet the requirements
391 listed. This gives the LLVM Pass Infrastructure information necessary to
392 optimize how passes are run, so that the resultant compiler isn't unneccesarily
397 <!-- ======================================================================= -->
398 <div class="doc_subsection">
399 <a name="ImmutablePass">The <tt>ImmutablePass</tt> class</a>
402 <div class="doc_text">
404 <p>The most plain and boring type of pass is the "<tt><a
405 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1ImmutablePass.html">ImmutablePass</a></tt>"
406 class. This pass type is used for passes that do not have to be run, do not
407 change state, and never need to be updated. This is not a normal type of
408 transformation or analysis, but can provide information about the current
409 compiler configuration.</p>
411 <p>Although this pass class is very infrequently used, it is important for
412 providing information about the current target machine being compiled for, and
413 other static information that can affect the various transformations.</p>
415 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
416 invalidated, and are never "run".</p>
420 <!-- ======================================================================= -->
421 <div class="doc_subsection">
422 <a name="Pass">The <tt>Pass</tt> class</a>
425 <div class="doc_text">
428 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">Pass</a></tt>"
429 class is the most general of all superclasses that you can use. Deriving from
430 <tt>Pass</tt> indicates that your pass uses the entire program as a unit,
431 refering to function bodies in no predictable order, or adding and removing
432 functions. Because nothing is known about the behavior of direct <tt>Pass</tt>
433 subclasses, no optimization can be done for their execution.</p>
435 <p>To write a correct <tt>Pass</tt> subclass, derive from <tt>Pass</tt> and
436 overload the <tt>run</tt> method with the following signature:</p>
440 <!-- _______________________________________________________________________ -->
441 <div class="doc_subsubsection">
442 <a name="run">The <tt>run</tt> method</a>
445 <div class="doc_text">
448 <b>virtual bool</b> run(Module &M) = 0;
451 <p>The <tt>run</tt> method performs the interesting work of the pass, and should
452 return true if the module was modified by the transformation, false
457 <!-- ======================================================================= -->
458 <div class="doc_subsection">
459 <a name="FunctionPass">The <tt>FunctionPass</tt> class</a>
462 <div class="doc_text">
464 <p>In contrast to direct <tt>Pass</tt> subclasses, direct <tt><a
465 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1Pass.html">FunctionPass</a></tt>
466 subclasses do have a predictable, local behavior that can be expected by the
467 system. All <tt>FunctionPass</tt> execute on each function in the program
468 independent of all of the other functions in the program.
469 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
470 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
472 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
475 <li>Modify a Function other than the one currently being processed.</li>
476 <li>Add or remove Function's from the current Module.</li>
477 <li>Add or remove global variables from the current Module.</li>
478 <li>Maintain state across invocations of
479 <a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)</li>
482 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
483 href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
484 may overload three virtual methods to do their work. All of these methods
485 should return true if they modified the program, or false if they didn't.</p>
489 <!-- _______________________________________________________________________ -->
490 <div class="doc_subsubsection">
491 <a name="doInitialization_mod">The <tt>doInitialization(Module &)</tt>
495 <div class="doc_text">
498 <b>virtual bool</b> doInitialization(Module &M);
501 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
502 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
503 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
504 is designed to do simple initialization type of stuff that does not depend on
505 the functions being processed. The <tt>doInitialization</tt> method call is not
506 scheduled to overlap with any other pass executions (thus it should be very
509 <p>A good example of how this method should be used is the <a
510 href="http://llvm.cs.uiuc.edu/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
511 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
512 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
513 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
514 free functions that it needs, adding prototypes to the module if necessary.</p>
518 <!-- _______________________________________________________________________ -->
519 <div class="doc_subsubsection">
520 <a name="runOnFunction">The <tt>runOnFunction</tt> method</a>
523 <div class="doc_text">
526 <b>virtual bool</b> runOnFunction(Function &F) = 0;
529 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
530 the transformation or analysis work of your pass. As usual, a true value should
531 be returned if the function is modified.</p>
535 <!-- _______________________________________________________________________ -->
536 <div class="doc_subsubsection">
537 <a name="doFinalization_mod">The <tt>doFinalization(Module
538 &)</tt> method</a>
541 <div class="doc_text">
544 <b>virtual bool</b> doFinalization(Module &M);
547 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
548 called when the pass framework has finished calling <a
549 href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
550 program being compiled.</p>
554 <!-- ======================================================================= -->
555 <div class="doc_subsection">
556 <a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
559 <div class="doc_text">
561 <p><tt>BasicBlockPass</tt>'s are just like <a
562 href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
563 their scope of inspection and modification to a single basic block at a time.
564 As such, they are <b>not</b> allowed to do any of the following:</p>
567 <li>Modify or inspect any basic blocks outside of the current one</li>
568 <li>Maintain state across invocations of
569 <a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a></li>
570 <li>Modify the constrol flow graph (by altering terminator instructions)</li>
571 <li>Any of the things verboten for
572 <a href="#FunctionPass"><tt>FunctionPass</tt></a>es.</li>
575 <p><tt>BasicBlockPass</tt>es are useful for traditional local and "peephole"
576 optimizations. They may override the same <a
577 href="#doInitialization_mod"><tt>doInitialization(Module &)</tt></a> and <a
578 href="#doFinalization_mod"><tt>doFinalization(Module &)</tt></a> methods that <a
579 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
583 <!-- _______________________________________________________________________ -->
584 <div class="doc_subsubsection">
585 <a name="doInitialization_fn">The <tt>doInitialization(Function
586 &)</tt> method</a>
589 <div class="doc_text">
592 <b>virtual bool</b> doInitialization(Function &F);
595 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
596 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
597 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
598 to do simple initialization type of stuff that does not depend on the
599 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
600 scheduled to overlap with any other pass executions (thus it should be very
605 <!-- _______________________________________________________________________ -->
606 <div class="doc_subsubsection">
607 <a name="runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
610 <div class="doc_text">
613 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
616 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
617 function is not allowed to inspect or modify basic blocks other than the
618 parameter, and are not allowed to modify the CFG. A true value must be returned
619 if the basic block is modified.</p>
623 <!-- _______________________________________________________________________ -->
624 <div class="doc_subsubsection">
625 <a name="doFinalization_fn">The <tt>doFinalization(Function &)</tt>
629 <div class="doc_text">
632 <b>virtual bool</b> doFinalization(Function &F);
635 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
636 called when the pass framework has finished calling <a
637 href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a> for every BasicBlock in the
638 program being compiled. This can be used to perform per-function
643 <!-- ======================================================================= -->
644 <div class="doc_subsection">
645 <a name="MachineFunctionPass">The <tt>MachineFunctionPass</tt> class</a>
648 <div class="doc_text">
650 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
651 executes on the machine-dependent representation of each LLVM function in the
652 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
653 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
654 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
655 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
658 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
659 <li>Modify a MachineFunction other than the one currently being processed.</li>
660 <li>Add or remove MachineFunctions from the current Module.</li>
661 <li>Add or remove global variables from the current Module.</li>
662 <li>Maintain state across invocations of <a
663 href="#runOnMachineFunction"><tt>runOnMachineFunction</tt></a> (including global
669 <!-- _______________________________________________________________________ -->
670 <div class="doc_subsubsection">
671 <a name="runOnMachineFunction">The <tt>runOnMachineFunction(MachineFunction
672 &MF)</tt> method</a>
675 <div class="doc_text">
678 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
681 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
682 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
683 work of your <tt>MachineFunctionPass</tt>.</p>
685 <p>The <tt>runOnMachineFunction</tt> method is called on every
686 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
687 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
688 representation of the function. If you want to get at the LLVM <tt>Function</tt>
689 for the <tt>MachineFunction</tt> you're working on, use
690 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
691 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
692 <tt>MachineFunctionPass</tt>.</p>
696 <!-- *********************************************************************** -->
697 <div class="doc_section">
698 <a name="registration">Pass registration</a>
700 <!-- *********************************************************************** -->
702 <div class="doc_text">
704 <p>In the <a href="#basiccode">Hello World</a> example pass we illustrated how
705 pass registration works, and discussed some of the reasons that it is used and
706 what it does. Here we discuss how and why passes are registered.</p>
708 <p>Passes can be registered in several different ways. Depending on the general
709 classification of the pass, you should use one of the following templates to
710 register the pass:</p>
713 <li><b><tt>RegisterOpt</tt></b> - This template should be used when you are
714 registering a pass that logically should be available for use in the
715 '<tt>opt</tt>' utility.</li>
717 <li><b><tt>RegisterAnalysis</tt></b> - This template should be used when you are
718 registering a pass that logically should be available for use in the
719 '<tt>analyze</tt>' utility.</li>
721 <li><b><tt>RegisterPass</tt></b> - This is the generic form of the
722 <tt>Register*</tt> templates that should be used if you want your pass listed by
723 multiple or no utilities. This template takes an extra third argument that
724 specifies which tools it should be listed in. See the <a
725 href="http://llvm.cs.uiuc.edu/doxygen/PassSupport_8h-source.html">PassSupport.h</a>
726 file for more information.</li>
730 <p>Regardless of how you register your pass, you must specify at least two
731 parameters. The first parameter is the name of the pass that is to be used on
732 the command line to specify that the pass should be added to a program (for
733 example <tt>opt</tt> or <tt>analyze</tt>). The second argument is the name of
734 the pass, which is to be used for the <tt>--help</tt> output of programs, as
735 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
737 <p>If a pass is registered to be used by the <tt>analyze</tt> utility, you
738 should implement the virtual <tt>print</tt> method:</p>
742 <!-- _______________________________________________________________________ -->
743 <div class="doc_subsubsection">
744 <a name="print">The <tt>print</tt> method</a>
747 <div class="doc_text">
750 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
753 <p>The <tt>print</tt> method must be implemented by "analyses" in order to print
754 a human readable version of the analysis results. This is useful for debugging
755 an analysis itself, as well as for other people to figure out how an analysis
756 works. The <tt>analyze</tt> tool uses this method to generate its output.</p>
758 <p>The <tt>ostream</tt> parameter specifies the stream to write the results on,
759 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
760 program that has been analyzed. Note however that this pointer may be null in
761 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
762 debugger), so it should only be used to enhance debug output, it should not be
767 <!-- *********************************************************************** -->
768 <div class="doc_section">
769 <a name="interaction">Specifying interactions between passes</a>
771 <!-- *********************************************************************** -->
773 <div class="doc_text">
775 <p>One of the main responsibilities of the <tt>PassManager</tt> is the make sure
776 that passes interact with each other correctly. Because <tt>PassManager</tt>
777 tries to <a href="#passmanager">optimize the execution of passes</a> it must
778 know how the passes interact with each other and what dependencies exist between
779 the various passes. To track this, each pass can declare the set of passes that
780 are required to be executed before the current pass, and the passes which are
781 invalidated by the current pass.</p>
783 <p>Typically this functionality is used to require that analysis results are
784 computed before your pass is run. Running arbitrary transformation passes can
785 invalidate the computed analysis results, which is what the invalidation set
786 specifies. If a pass does not implement the <tt><a
787 href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
788 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
792 <!-- _______________________________________________________________________ -->
793 <div class="doc_subsubsection">
794 <a name="getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
797 <div class="doc_text">
800 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
803 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
804 invalidated sets may be specified for your transformation. The implementation
805 should fill in the <tt><a
806 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AnalysisUsage.html">AnalysisUsage</a></tt>
807 object with information about which passes are required and not invalidated. To
808 do this, a pass may call any of the following methods on the AnalysisUsage
812 <!-- _______________________________________________________________________ -->
813 <div class="doc_subsubsection">
814 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
817 <div class="doc_text">
819 If you pass requires a previous pass to be executed (an analysis for example),
820 it can use one of these methods to arrange for it to be run before your pass.
821 LLVM has many different types of analyses and passes that can be required,
822 spaning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
823 requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
824 be no critical edges in the CFG when your pass has been run.
828 Some analyses chain to other analyses to do their job. For example, an <a
829 href="AliasAnalysis.html">AliasAnalysis</a> implementation is required to <a
830 href="AliasAnalysis.html#chaining">chain</a> to other alias analysis passes. In
831 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
832 used instead of the <tt>addRequired</tt> method. This informs the PassManager
833 that the transitively required pass should be alive as long as the requiring
838 <!-- _______________________________________________________________________ -->
839 <div class="doc_subsubsection">
840 <a name="AU::addPreserved">The <tt>AnalysisUsage::addPreserved<></tt> method</a>
843 <div class="doc_text">
845 One of the jobs of the PassManager is to optimize how and when analyses are run.
846 In particular, it attempts to avoid recomputing data unless it needs to. For
847 this reason, passes are allowed to declare that they preserve (i.e., they don't
848 invalidate) an existing analysis if it's available. For example, a simple
849 constant folding pass would not modify the CFG, so it can't possible effect the
850 results of dominator analysis. By default, all passes are assumed to invalidate
855 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
856 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
857 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
858 not modify the LLVM program at all (which is true for analyses), and the
859 <tt>setPreservesCFG</tt> method can be used by transformations that change
860 instructions in the program but do not modify the CFG or terminator instructions
861 (note that this property is implicitly set for <a
862 href="#BasicBlockPass">BasicBlockPass</a>'s).
866 <tt>addPreserved</tt> is particularly useful for transformations like
867 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
868 and dominator related analyses if they exist, so it can preserve them, despite
869 the fact that it hacks on the CFG.
873 <!-- _______________________________________________________________________ -->
874 <div class="doc_subsubsection">
875 <a name="AU::examples">Example implementations of <tt>getAnalysisUsage</tt></a>
878 <div class="doc_text">
881 <i>// This is an example implementation from an analysis, which does not modify
882 // the program at all, yet has a prerequisite.</i>
883 <b>void</b> <a href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
884 AU.setPreservesAll();
885 AU.addRequired<<a href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
892 <i>// This example modifies the program, but does not modify the CFG</i>
893 <b>void</b> <a href="http://llvm.cs.uiuc.edu/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
894 AU.setPreservesCFG();
895 AU.addRequired<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
901 <!-- _______________________________________________________________________ -->
902 <div class="doc_subsubsection">
903 <a name="getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a>
906 <div class="doc_text">
908 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
909 your class, providing you with access to the passes that you declared that you
910 required with the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
911 method. It takes a single template argument that specifies which pass class you
912 want, and returns a reference to that pass. For example:</p>
915 bool LICM::runOnFunction(Function &F) {
916 LoopInfo &LI = getAnalysis<LoopInfo>();
921 <p>This method call returns a reference to the pass desired. You may get a
922 runtime assertion failure if you attempt to get an analysis that you did not
923 declare as required in your <a
924 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> implementation. This
925 method can be called by your <tt>run*</tt> method implementation, or by any
926 other local method invoked by your <tt>run*</tt> method.</p>
929 If your pass is capable of updating analyses if they exist (e.g.,
930 <tt>BreakCriticalEdges</tt>, as described above), you can use the
931 <tt>getAnalysisToUpdate</tt> method, which returns a pointer to the analysis if
932 it is active. For example:</p>
936 if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
937 <i>// A DominatorSet is active. This code will update it.</i>
944 <!-- *********************************************************************** -->
945 <div class="doc_section">
946 <a name="analysisgroup">Implementing Analysis Groups</a>
948 <!-- *********************************************************************** -->
950 <div class="doc_text">
952 <p>Now that we understand the basics of how passes are defined, how the are
953 used, and how they are required from other passes, it's time to get a little bit
954 fancier. All of the pass relationships that we have seen so far are very
955 simple: one pass depends on one other specific pass to be run before it can run.
956 For many applications, this is great, for others, more flexibility is
959 <p>In particular, some analyses are defined such that there is a single simple
960 interface to the analysis results, but multiple ways of calculating them.
961 Consider alias analysis for example. The most trivial alias analysis returns
962 "may alias" for any alias query. The most sophisticated analysis a
963 flow-sensitive, context-sensitive interprocedural analysis that can take a
964 significant amount of time to execute (and obviously, there is a lot of room
965 between these two extremes for other implementations). To cleanly support
966 situations like this, the LLVM Pass Infrastructure supports the notion of
971 <!-- _______________________________________________________________________ -->
972 <div class="doc_subsubsection">
973 <a name="agconcepts">Analysis Group Concepts</a>
976 <div class="doc_text">
978 <p>An Analysis Group is a single simple interface that may be implemented by
979 multiple different passes. Analysis Groups can be given human readable names
980 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
981 class. An analysis group may have one or more implementations, one of which is
982 the "default" implementation.</p>
984 <p>Analysis groups are used by client passes just like other passes are: the
985 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
986 In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
987 scans the available passes to see if any implementations of the analysis group
988 are available. If none is available, the default implementation is created for
989 the pass to use. All standard rules for <A href="#interaction">interaction
990 between passes</a> still apply.</p>
992 <p>Although <a href="#registration">Pass Registration</a> is optional for normal
993 passes, all analysis group implementations must be registered, and must use the
994 <A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
995 implementation pool. Also, a default implementation of the interface
996 <b>must</b> be registered with <A
997 href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.</p>
999 <p>As a concrete example of an Analysis Group in action, consider the <a
1000 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>
1001 analysis group. The default implementation of the alias analysis interface (the
1003 href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
1004 pass) just does a few simple checks that don't require significant analysis to
1005 compute (such as: two different globals can never alias each other, etc).
1006 Passes that use the <tt><a
1007 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1008 interface (for example the <tt><a
1009 href="http://llvm.cs.uiuc.edu/doxygen/structGCSE.html">gcse</a></tt> pass), do
1010 not care which implementation of alias analysis is actually provided, they just
1011 use the designated interface.</p>
1013 <p>From the user's perspective, commands work just like normal. Issuing the
1014 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1015 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1016 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1017 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1018 hypothetical example) instead.</p>
1022 <!-- _______________________________________________________________________ -->
1023 <div class="doc_subsubsection">
1024 <a name="registerag">Using <tt>RegisterAnalysisGroup</tt></a>
1027 <div class="doc_text">
1029 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1030 group itself as well as add pass implementations to the analysis group. First,
1031 an analysis should be registered, with a human readable name provided for it.
1032 Unlike registration of passes, there is no command line argument to be specified
1033 for the Analysis Group Interface itself, because it is "abstract":</p>
1036 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1039 <p>Once the analysis is registered, passes can declare that they are valid
1040 implementations of the interface by using the following code:</p>
1044 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1045 RegisterOpt<FancyAA>
1046 B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
1048 //<i> Declare that we implement the AliasAnalysis interface</i>
1049 RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, FancyAA> C;
1053 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1054 uses the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
1055 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1056 analysis group. Every implementation of an analysis group should join using
1057 this template. A single pass may join multiple different analysis groups with
1062 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1063 RegisterOpt<<a href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1064 D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
1066 //<i> Declare that we implement the AliasAnalysis interface</i>
1067 RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <a href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>, <b>true</b>> E;
1071 <p>Here we show how the default implementation is specified (using the extra
1072 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1073 one default implementation available at all times for an Analysis Group to be
1074 used. Here we declare that the <tt><a
1075 href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1076 pass is the default implementation for the interface.</p>
1080 <!-- *********************************************************************** -->
1081 <div class="doc_section">
1082 <a name="passmanager">What PassManager does</a>
1084 <!-- *********************************************************************** -->
1086 <div class="doc_text">
1089 href="http://llvm.cs.uiuc.edu/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
1091 href="http://llvm.cs.uiuc.edu/doxygen/classllvm_1_1PassManager.html">class</a>
1092 takes a list of passes, ensures their <a href="#interaction">prerequisites</a>
1093 are set up correctly, and then schedules passes to run efficiently. All of the
1094 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1097 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1098 time of a series of passes:</p>
1101 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1102 recomputing analysis results as much as possible. This means keeping track of
1103 which analyses are available already, which analyses get invalidated, and which
1104 analyses are needed to be run for a pass. An important part of work is that the
1105 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1106 it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
1107 results as soon as they are no longer needed.</li>
1109 <li><b>Pipeline the execution of passes on the program</b> - The
1110 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1111 of a series of passes by pipelining the passes together. This means that, given
1112 a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
1113 will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
1114 the first function, then all of the <a
1115 href="#FunctionPass"><tt>FunctionPass</tt></a>es on the second function,
1116 etc... until the entire program has been run through the passes.
1118 <p>This improves the cache behavior of the compiler, because it is only touching
1119 the LLVM program representation for a single function at a time, instead of
1120 traversing the entire program. It reduces the memory consumption of compiler,
1121 because, for example, only one <a
1122 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1DominatorSet.html"><tt>DominatorSet</tt></a>
1123 needs to be calculated at a time. This also makes it possible some <a
1124 href="#SMP">interesting enhancements</a> in the future.</p></li>
1128 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1129 much information it has about the behaviors of the passes it is scheduling. For
1130 example, the "preserved" set is intentionally conservative in the face of an
1131 unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
1132 Not implementing when it should be implemented will have the effect of not
1133 allowing any analysis results to live across the execution of your pass.</p>
1135 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1136 options that is useful for debugging pass execution, seeing how things work, and
1137 diagnosing when you should be preserving more analyses than you currently are
1138 (To get information about all of the variants of the <tt>--debug-pass</tt>
1139 option, just type '<tt>opt --help-hidden</tt>').</p>
1141 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1142 how our <a href="#basiccode">Hello World</a> pass interacts with other passes.
1143 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1146 $ opt -load ../../../lib/Debug/libhello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1148 Function Pass Manager
1149 Dominator Set Construction
1150 Immediate Dominators Construction
1151 Global Common Subexpression Elimination
1152 -- Immediate Dominators Construction
1153 -- Global Common Subexpression Elimination
1154 Natural Loop Construction
1155 Loop Invariant Code Motion
1156 -- Natural Loop Construction
1157 -- Loop Invariant Code Motion
1159 -- Dominator Set Construction
1165 <p>This output shows us when passes are constructed and when the analysis
1166 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1167 GCSE uses dominator and immediate dominator information to do its job. The LICM
1168 pass uses natural loop information, which uses dominator sets, but not immediate
1169 dominators. Because immediate dominators are no longer useful after the GCSE
1170 pass, it is immediately destroyed. The dominator sets are then reused to
1171 compute natural loop information, which is then used by the LICM pass.</p>
1173 <p>After the LICM pass, the module verifier runs (which is automatically added
1174 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1175 resultant LLVM code is well formed. After it finishes, the dominator set
1176 information is destroyed, after being computed once, and shared by three
1179 <p>Lets see how this changes when we run the <a href="#basiccode">Hello
1180 World</a> pass in between the two passes:</p>
1183 $ opt -load ../../../lib/Debug/libhello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1185 Function Pass Manager
1186 Dominator Set Construction
1187 Immediate Dominators Construction
1188 Global Common Subexpression Elimination
1189 <b>-- Dominator Set Construction</b>
1190 -- Immediate Dominators Construction
1191 -- Global Common Subexpression Elimination
1192 <b> Hello World Pass
1194 Dominator Set Construction</b>
1195 Natural Loop Construction
1196 Loop Invariant Code Motion
1197 -- Natural Loop Construction
1198 -- Loop Invariant Code Motion
1200 -- Dominator Set Construction
1209 <p>Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
1210 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1211 we need to add the following <a
1212 href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:</p>
1215 <i>// We don't modify the program, so we preserve all analyses</i>
1216 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1217 AU.setPreservesAll();
1221 <p>Now when we run our pass, we get this output:</p>
1224 $ opt -load ../../../lib/Debug/libhello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1225 Pass Arguments: -gcse -hello -licm
1227 Function Pass Manager
1228 Dominator Set Construction
1229 Immediate Dominators Construction
1230 Global Common Subexpression Elimination
1231 -- Immediate Dominators Construction
1232 -- Global Common Subexpression Elimination
1235 Natural Loop Construction
1236 Loop Invariant Code Motion
1237 -- Loop Invariant Code Motion
1238 -- Natural Loop Construction
1240 -- Dominator Set Construction
1249 <p>Which shows that we don't accidentally invalidate dominator information
1250 anymore, and therefore do not have to compute it twice.</p>
1254 <!-- _______________________________________________________________________ -->
1255 <div class="doc_subsubsection">
1256 <a name="releaseMemory">The <tt>releaseMemory</tt> method</a>
1259 <div class="doc_text">
1262 <b>virtual void</b> releaseMemory();
1265 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1266 results, and how long to keep them around for. Because the lifetime of the pass
1267 object itself is effectively the entire duration of the compilation process, we
1268 need some way to free analysis results when they are no longer useful. The
1269 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1271 <p>If you are writing an analysis or any other pass that retains a significant
1272 amount of state (for use by another pass which "requires" your pass and uses the
1273 <a href="#getAnalysis">getAnalysis</a> method) you should implement
1274 <tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
1275 internal state. This method is called after the <tt>run*</tt> method for the
1276 class, before the next call of <tt>run*</tt> in your pass.</p>
1280 <!-- *********************************************************************** -->
1281 <div class="doc_section">
1282 <a name="debughints">Using GDB with dynamically loaded passes</a>
1284 <!-- *********************************************************************** -->
1286 <div class="doc_text">
1288 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1289 should be. First of all, you can't set a breakpoint in a shared object that has
1290 not been loaded yet, and second of all there are problems with inlined functions
1291 in shared objects. Here are some suggestions to debugging your pass with
1294 <p>For sake of discussion, I'm going to assume that you are debugging a
1295 transformation invoked by <tt>opt</tt>, although nothing described here depends
1300 <!-- _______________________________________________________________________ -->
1301 <div class="doc_subsubsection">
1302 <a name="breakpoint">Setting a breakpoint in your pass</a>
1305 <div class="doc_text">
1307 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1312 Copyright 2000 Free Software Foundation, Inc.
1313 GDB is free software, covered by the GNU General Public License, and you are
1314 welcome to change it and/or distribute copies of it under certain conditions.
1315 Type "show copying" to see the conditions.
1316 There is absolutely no warranty for GDB. Type "show warranty" for details.
1317 This GDB was configured as "sparc-sun-solaris2.6"...
1321 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1322 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1323 (the shared object isn't loaded until runtime), we must execute the process, and
1324 have it stop before it invokes our pass, but after it has loaded the shared
1325 object. The most foolproof way of doing this is to set a breakpoint in
1326 <tt>PassManager::run</tt> and then run the process with the arguments you
1330 (gdb) <b>break PassManager::run</b>
1331 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1332 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/lib/Debug/[libname].so -[passoption]</b>
1333 Starting program: opt test.bc -load $(LLVMTOP)/llvm/lib/Debug/[libname].so -[passoption]
1334 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1335 70 bool PassManager::run(Module &M) { return PM->run(M); }
1339 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1340 now free to set breakpoints in your pass so that you can trace through execution
1341 or do other standard debugging stuff.</p>
1345 <!-- _______________________________________________________________________ -->
1346 <div class="doc_subsubsection">
1347 <a name="debugmisc">Miscellaneous Problems</a>
1350 <div class="doc_text">
1352 <p>Once you have the basics down, there are a couple of problems that GDB has,
1353 some with solutions, some without.</p>
1356 <li>Inline functions have bogus stack information. In general, GDB does a
1357 pretty good job getting stack traces and stepping through inline functions.
1358 When a pass is dynamically loaded however, it somehow completely loses this
1359 capability. The only solution I know of is to de-inline a function (move it
1360 from the body of a class to a .cpp file).</li>
1362 <li>Restarting the program breaks breakpoints. After following the information
1363 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1364 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1365 and you start getting errors about breakpoints being unsettable. The only way I
1366 have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
1367 already set in your pass, run the program, and re-set the breakpoints once
1368 execution stops in <tt>PassManager::run</tt>.</li>
1372 <p>Hopefully these tips will help with common case debugging situations. If
1373 you'd like to contribute some tips of your own, just contact <a
1374 href="mailto:sabre@nondot.org">Chris</a>.</p>
1378 <!-- *********************************************************************** -->
1379 <div class="doc_section">
1380 <a name="future">Future extensions planned</a>
1382 <!-- *********************************************************************** -->
1384 <div class="doc_text">
1386 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1387 some nifty stuff, there are things we'd like to add in the future. Here is
1388 where we are going:</p>
1392 <!-- _______________________________________________________________________ -->
1393 <div class="doc_subsubsection">
1394 <a name="SMP">Multithreaded LLVM</a>
1397 <div class="doc_text">
1399 <p>Multiple CPU machines are becoming more common and compilation can never be
1400 fast enough: obviously we should allow for a multithreaded compiler. Because of
1401 the semantics defined for passes above (specifically they cannot maintain state
1402 across invocations of their <tt>run*</tt> methods), a nice clean way to
1403 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1404 to create multiple instances of each pass object, and allow the separate
1405 instances to be hacking on different parts of the program at the same time.</p>
1407 <p>This implementation would prevent each of the passes from having to implement
1408 multithreaded constructs, requiring only the LLVM core to have locking in a few
1409 places (for global resources). Although this is a simple extension, we simply
1410 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1411 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1415 <!-- _______________________________________________________________________ -->
1416 <div class="doc_subsubsection">
1417 <a name="ModuleSource">A new <tt>ModuleSource</tt> interface</a>
1420 <div class="doc_text">
1422 <p>Currently, the <tt>PassManager</tt>'s <tt>run</tt> method takes a <tt><a
1423 href="http://llvm.cs.uiuc.edu/doxygen/structllvm_1_1Module.html">Module</a></tt>
1424 as input, and runs all of the passes on this module. The problem with this
1425 approach is that none of the <tt>PassManager</tt> features can be used for
1426 timing and debugging the actual <b>loading</b> of the module from disk or
1429 <p>To solve this problem, eventually the <tt>PassManager</tt> class will accept
1430 a <tt>ModuleSource</tt> object instead of a Module itself. When complete, this
1431 will also allow for streaming of functions out of the bytecode representation,
1432 allowing us to avoid holding the entire program in memory at once if we only are
1433 dealing with <a href="#FunctionPass">FunctionPass</a>es.</p>
1435 <p>As part of a different issue, eventually the bytecode loader will be extended
1436 to allow on-demand loading of functions from the bytecode representation, in
1437 order to better support the runtime reoptimizer. The bytecode format is already
1438 capable of this, the loader just needs to be reworked a bit.</p>
1442 <!-- _______________________________________________________________________ -->
1443 <div class="doc_subsubsection">
1444 <a name="PassFunctionPass"><tt>Pass</tt>es requiring <tt>FunctionPass</tt>es</a>
1447 <div class="doc_text">
1449 <p>Currently it is illegal for a <a href="#Pass"><tt>Pass</tt></a> to require a
1450 <a href="#FunctionPass"><tt>FunctionPass</tt></a>. This is because there is
1451 only one instance of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>
1452 object ever created, thus nowhere to store information for all of the functions
1453 in the program at the same time. Although this has come up a couple of times
1454 before, this has always been worked around by factoring one big complicated pass
1455 into a global and an interprocedural part, both of which are distinct. In the
1456 future, it would be nice to have this though.</p>
1458 <p>Note that it is no problem for a <a
1459 href="#FunctionPass"><tt>FunctionPass</tt></a> to require the results of a <a
1460 href="#Pass"><tt>Pass</tt></a>, only the other way around.</p>
1464 <!-- *********************************************************************** -->
1467 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1468 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1469 <a href="http://validator.w3.org/check/referer"><img
1470 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!" /></a>
1472 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1473 <a href="http://llvm.cs.uiuc.edu">The LLVM Compiler Infrastructure</a><br>
1474 Last modified: $Date$