Blame - docs/WritingAnLLVMPass.html - fp2-dev/platform/external/llvm

blob: 353fe6e366fdc8814272dd8d81260cbc85df55d6 [file] [log] [blame]

Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	1	<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
				2	<html><head><title>Writing an LLVM Pass</title></head>
				3
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	4	<body bgcolor=white>
				5
				6	<table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				7	<tr><td>  <font size=+3 color="#EEEEFF" face="Georgia,Palatino,Times,Roman"><b>Writing an LLVM Pass</b></font></td>
				8	</tr></table>
				9
				10
				11	<ol>
				12	<li><a href="#introduction">Introduction - What is a pass?</a>
				13	<li><a href="#quickstart">Quick Start - Writing hello world</a>
				14	<ul>
				15	<li><a href="#makefile">Setting up the build environment</a>
				16	<li><a href="#basiccode">Basic code required</a>
				17	<li><a href="#running">Running a pass with <tt>opt</tt>
				18	or <tt>analyze</tt></a>
				19	</ul>
				20	<li><a href="#passtype">Pass classes and requirements</a>
				21	<ul>
				22	<li><a href="#Pass">The <tt>Pass</tt> class</a>
				23	<ul>
				24	<li><a href="#run">The <tt>run</tt> method</a>
				25	</ul>
				26	<li><a href="#FunctionPass">The <tt>FunctionPass</tt> class</a>
				27	<ul>
				28	<li><a href="#doInitialization">The <tt>doInitialization</tt> method</a>
				29	<li><a href="#runOnFunction">The <tt>runOnFunction</tt> method</a>
				30	<li><a href="#doFinalization">The <tt>doFinalization</tt> method</a>
				31	</ul>
				32	<li><a href="#BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
				33	<ul>
				34	<li><a href="#runOnBasicBlock">The <tt>runOnBasicBlock</tt> method</a>
				35	</ul>
				36	</ul>
				37	<li><a href="#registration">Pass Registration</a>
				38	<ul>
				39	<li><a href="#print">The <tt>print</tt> method</a>
				40	</ul>
				41	<li><a href="#interaction">Specifying interactions between passes</a>
				42	<ul>
				43	<li><a href="#getAnalysisUsage">The <tt>getAnalysisUsage</tt> method</a>
				44	<li><a href="#getAnalysis">The <tt>getAnalysis</tt> method</a>
				45	</ul>
Chris Lattner	7991070	2002-08-22 19:21:04 +0000	[diff] [blame]	46	<li><a href="#analysisgroup">Implementing Analysis Groups</a>
				47	<ul>
				48	<li><a href="#agconcepts">Analysis Group Concepts</a>
				49	<li><a href="#registerag">Using <tt>RegisterAnalysisGroup</tt></a>
				50	</ul>
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	51	<li><a href="#passmanager">What PassManager does</a>
				52	<ul>
				53	<li><a href="#releaseMemory">The <tt>releaseMemory</tt> method</a>
				54	</ul>
Chris Lattner	480e2ef	2002-09-06 02:02:58 +0000	[diff] [blame]	55	<li><a href="#debughints">Using GDB with dynamically loaded passes</a>
				56	<ul>
				57	<li><a href="#breakpoint">Setting a breakpoint in your pass
				58	<li><a href="#debugmisc">Miscellaneous Problems
				59	</ul>
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	60	<li><a href="#future">Future extensions planned</a>
				61	<ul>
				62	<li><a href="#SMP">Multithreaded LLVM</a>
				63	<li><a href="#ModuleSource">A new <tt>ModuleSource</tt> interface</a>
				64	<li><a href="#PassFunctionPass"><tt>Pass</tt>'s requiring <tt>FunctionPass</tt>'s</a>
				65	</ul>
Chris Lattner	38c633d	2002-08-08 20:23:41 +0000	[diff] [blame]	66
				67	<p><b>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></b><p>
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	68	</ol><p>
				69
				70
				71
				72	<!-- *********************************************************************** -->
				73	<table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				74	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
				75	<a name="introduction">Introduction - What is a pass?
				76	</b></font></td></tr></table><ul>
				77	<!-- *********************************************************************** -->
				78
				79	The LLVM Pass Framework is an important part of the LLVM system, because LLVM
				80	passes are where the interesting parts of the compiler exist. Passes perform
				81	the transformations and optimizations that make up the compiler, they build
				82	the analysis results that are used by these transformations, and they are, above
				83	all, a structuring technique for compiler code.<p>
				84
				85	All LLVM passes are subclasses of the <tt><a
				86	href="http://llvm.cs.uiuc.edu/doxygen/classPass.html">Pass</a></tt> class, which
				87	implement functionality by overriding virtual methods inherited from
				88	<tt>Pass</tt>. Depending on how your pass works, you may be able to inherit
				89	from the <tt><a
				90	href="http://llvm.cs.uiuc.edu/doxygen/structFunctionPass.html">FunctionPass</a></tt>
				91	or <tt><a
				92	href="http://llvm.cs.uiuc.edu/doxygen/structBasicBlockPass.html">BasicBlockPass</a></tt>,
				93	which gives the system more information about what your pass does, and how it
				94	can be combined with other passes. One of the main features of the LLVM Pass
				95	Framework is that it schedules passes to run in an efficient way based on the
				96	constraints that your pass has.<p>
				97
				98	We start by showing you how to construct a pass, everything from setting up the
				99	code, to compiling, loading, and executing it. After the basics are down, more
				100	advanced features are discussed.<p>
				101
				102
				103	<!-- *********************************************************************** -->
				104	</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				105	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
				106	<a name="quickstart">Quick Start - Writing hello world
				107	</b></font></td></tr></table><ul>
				108	<!-- *********************************************************************** -->
				109
				110	Here we describe how to write the "hello world" of passes. The "Hello" pass is
				111	designed to simply print out the name of non-external functions that exist in
				112	the program being compiled. It does not modify the program at all, just
				113	inspects it. The source code and files for this pass are available in the LLVM
				114	source tree in the <tt>lib/Transforms/Hello</tt> directory.<p>
				115
				116
				117	<!-- ======================================================================= -->
				118	</ul><table width="100%" bgcolor="#441188" border=0 cellpadding=4 cellspacing=0>
				119	<tr><td> </td><td width="100%">
				120	<font color="#EEEEFF" face="Georgia,Palatino"><b>
				121	<a name="makefile">Setting up the build environment
				122	</b></font></td></tr></table><ul>
				123
				124	First thing you need to do is create a new directory somewhere in the LLVM
				125	source base. For this example, we'll assume that you made
				126	"<tt>lib/Transforms/Hello</tt>". The first thing you must do is set up a build
				127	script (Makefile) that will compile the source code for the new pass. To do
Chris Lattner	17a4c3e	2002-08-14 20:06:13 +0000	[diff] [blame]	128	this, copy this into "<tt>Makefile</tt>" (be very careful that there are no
				129	extra space characters at the end of the lines though... that seems to confuse
				130	<tt>gmake</tt>):<p>
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	131
				132	</ul><hr><ul><pre>
				133	# Makefile for hello pass
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	134
Chris Lattner	17a4c3e	2002-08-14 20:06:13 +0000	[diff] [blame]	135	# Path to top level of LLVM heirarchy
Chris Lattner	7ce83e5	2002-08-14 20:07:01 +0000	[diff] [blame]	136	LEVEL = ../../..
Chris Lattner	17a4c3e	2002-08-14 20:06:13 +0000	[diff] [blame]	137
				138	# Name of the library to build
Chris Lattner	7ce83e5	2002-08-14 20:07:01 +0000	[diff] [blame]	139	LIBRARYNAME = hello
Chris Lattner	17a4c3e	2002-08-14 20:06:13 +0000	[diff] [blame]	140
				141	# Build a dynamically loadable shared object
				142	SHARED_LIBRARY = 1
				143
				144	# Include the makefile implementation stuff
				145	include $(LEVEL)/Makefile.common
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	146	</pre></ul><hr><ul><p>
				147
				148	This makefile specifies that all of the <tt>.cpp</tt> files in the current
				149	directory are to be compiled and linked together into a
				150	<tt>lib/Debug/libhello.so</tt> shared object that can be dynamically loaded by
				151	the <tt>opt</tt> or <tt>analyze</tt> tools.<p>
				152
				153	Now that we have the build scripts set up, we just need to write the code for
				154	the pass itself.<p>
				155
				156
				157	<!-- ======================================================================= -->
				158	</ul><table width="100%" bgcolor="#441188" border=0 cellpadding=4 cellspacing=0>
				159	<tr><td> </td><td width="100%">
				160	<font color="#EEEEFF" face="Georgia,Palatino"><b>
				161	<a name="basiccode">Basic code required
				162	</b></font></td></tr></table><ul>
				163
				164	Now that we have a way to compile our new pass, we just have to write it. Start
				165	out with:<p>
				166
				167	<pre>
				168	<b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
				169	<b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Function_8h-source.html">llvm/Function.h</a>"
				170	</pre>
				171
				172	Which are needed because we are writing a <tt><a
				173	href="http://llvm.cs.uiuc.edu/doxygen/classPass.html">Pass</a></tt>, and we are
				174	operating on <tt><a
				175	href="http://llvm.cs.uiuc.edu/doxygen/classFunction.html">Function</a></tt>'s.<p>
				176
				177	Next we have:<p>
				178
				179	<pre>
				180	<b>namespace</b> {
				181	</pre><p>
				182
				183	... which starts out an anonymous namespace. Anonymous namespaces are to C++
				184	what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
				185	things declared inside of the anonymous namespace only visible to the current
				186	file. If you're not familiar with them, consult a decent C++ book for more
				187	information.<p>
				188
				189	Next, we declare our pass itself:<p>
				190
				191	<pre>
				192	<b>struct</b> Hello : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
				193	</pre><p>
				194
				195	This declares a "<tt>Hello</tt>" class that is a subclass of <tt><a
				196	href="http://llvm.cs.uiuc.edu/doxygen/structFunctionPass.html">FunctionPass</a></tt>.
Chris Lattner	d6ea926	2002-09-09 03:48:46 +0000	[diff] [blame^]	197	The different builtin pass subclasses are described in detail <a
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	198	href="#passtype">later</a>, but for now, know that <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s
				199	operate a function at a time.<p>
				200
				201	<pre>
				202	<b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
				203	std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
				204	<b>return false</b>;
				205	}
				206	}; <i>// end of struct Hello</i>
				207	</pre>
				208
				209	We declare a "<a href="#runOnFunction"><tt>runOnFunction</tt></a>" method, which
				210	overloads an abstract virtual method inherited from <a
				211	href="#FunctionPass"><tt>FunctionPass</tt></a>. This is where we are supposed
				212	to do our thing, so we just print out our message with the name of each
				213	function.<p>
				214
				215	<pre>
				216	RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
				217	} <i>// end of anonymous namespace</i>
				218	</pre><p>
				219
				220	Lastly, we register our class <tt>Hello</tt>, giving it a command line argument
				221	"<tt>hello</tt>", and a name "<tt>Hello World Pass</tt>". There are several
				222	different ways of <a href="#registration">registering your pass</a>, depending
				223	on what it is to be used for. For "optimizations" we use the
				224	<tt>RegisterOpt</tt> template.<p>
				225
				226	As a whole, the <tt>.cpp</tt> file looks like:<p>
				227
				228	<pre>
				229	<b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Pass_8h-source.html">llvm/Pass.h</a>"
				230	<b>#include</b> "<a href="http://llvm.cs.uiuc.edu/doxygen/Function_8h-source.html">llvm/Function.h</a>"
				231
				232	<b>namespace</b> {
				233	<b>struct Hello</b> : <b>public</b> <a href="#FunctionPass">FunctionPass</a> {
				234	<b>virtual bool</b> <a href="#runOnFunction">runOnFunction</a>(Function &F) {
				235	std::cerr << "<i>Hello: </i>" << F.getName() << "\n";
				236	<b>return false</b>;
				237	}
				238	};
				239
				240	RegisterOpt<Hello> X("<i>hello</i>", "<i>Hello World Pass</i>");
				241	}
				242	</pre><p>
				243
				244	Now that it's all together, compile the file with a simple "<tt>gmake</tt>"
				245	command in the local directory and you should get a new
				246	"<tt>lib/Debug/libhello.so</tt> file. Note that everything in this file is
				247	contained in an anonymous namespace: this reflects the fact that passes are self
				248	contained units that do not need external interfaces (although they can have
				249	them) to be useful.<p>
				250
				251
				252	<!-- ======================================================================= -->
				253	</ul><table width="100%" bgcolor="#441188" border=0 cellpadding=4 cellspacing=0>
				254	<tr><td> </td><td width="100%">
				255	<font color="#EEEEFF" face="Georgia,Palatino"><b>
				256	<a name="running">Running a pass with <tt>opt</tt> or <tt>analyze</tt>
				257	</b></font></td></tr></table><ul>
				258
				259	Now that you have a brand new shiny <tt>.so</tt> file, we can use the
				260	<tt>opt</tt> command to run an LLVM program through your pass. Because you
				261	registered your pass with the <tt>RegisterOpt</tt> template, you will be able to
				262	use the <tt>opt</tt> tool to access it, once loaded.<p>
				263
				264	To test it, follow the example at the end of the <a
				265	href="GettingStarted.html">Getting Started Guide</a> to compile "Hello World" to
				266	LLVM. We can now run the bytecode file (<tt>hello.bc</tt>) for the program
				267	through our transformation like this (or course, any bytecode file will
				268	work):<p>
				269
				270	<pre>
				271	$ opt -load ../../../lib/Debug/libhello.so -hello < hello.bc > /dev/null
				272	Hello: __main
				273	Hello: puts
				274	Hello: main
				275	</pre><p>
				276
				277	The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your pass
				278	as a shared object, which makes '<tt>-hello</tt>' a valid command line argument
				279	(which is one reason you need to <a href="#registration">register your
				280	pass</a>). Because the hello pass does not modify the program in any
				281	interesting way, we just throw away the result of <tt>opt</tt> (sending it to
				282	<tt>/dev/null</tt>).<p>
				283
				284	To see what happened to the other string you registered, try running
				285	<tt>opt</tt> with the <tt>--help</tt> option:<p>
				286
				287	<pre>
				288	$ opt -load ../../../lib/Debug/libhello.so --help
				289	OVERVIEW: llvm .bc -> .bc modular optimizer
				290
				291	USAGE: opt [options] <input bytecode>
				292
				293	OPTIONS:
				294	Optimizations available:
				295	...
				296	-funcresolve - Resolve Functions
				297	-gcse - Global Common Subexpression Elimination
				298	-globaldce - Dead Global Elimination
				299	<b>-hello - Hello World Pass</b>
				300	-indvars - Cannonicalize Induction Variables
				301	-inline - Function Integration/Inlining
				302	-instcombine - Combine redundant instructions
				303	...
				304	</pre><p>
				305
				306	The pass name get added as the information string for your pass, giving some
				307	documentation to users of <tt>opt</tt>. Now that you have a working pass, you
				308	would go ahead and make it do the cool transformations you want. Once you get
				309	it all working and tested, it may become useful to find out how fast your pass
				310	is. The <a href="#passManager"><tt>PassManager</tt></a> provides a nice command
				311	line option (<tt>--time-passes</tt>) that allows you to get information about
				312	the execution time of your pass along with the other passes you queue up. For
				313	example:<p>
				314
				315	<pre>
				316	$ opt -load ../../../lib/Debug/libhello.so -hello -time-passes < hello.bc > /dev/null
				317	Hello: __main
				318	Hello: puts
				319	Hello: main
				320	===============================================================================
				321	... Pass execution timing report ...
				322	===============================================================================
				323	Total Execution Time: 0.02 seconds (0.0479059 wall clock)
				324
				325	---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
				326	0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bytecode Writer
				327	0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
				328	0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
				329	<b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
				330	0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
				331	</pre><p>
				332
				333	As you can see, our implementation above is pretty fast :). The additional
				334	passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
				335	that the LLVM emitted by your pass is still valid and well formed LLVM, which
				336	hasn't been broken somehow.
				337
				338	Now that you have seen the basics of the mechanics behind passes, we can talk
				339	about some more details of how they work and how to use them.<p>
				340
				341
				342
				343	<!-- *********************************************************************** -->
				344	</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				345	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
				346	<a name="passtype">Pass classes and requirements
				347	</b></font></td></tr></table><ul>
				348	<!-- *********************************************************************** -->
				349
				350	One of the first things that you should do when designing a new pass is to
				351	decide what class you should subclass for your pass. The <a
				352	href="#basiccode">Hello World</a> example uses the <tt><a
				353	href="#FunctionPass">FunctionPass</a></tt> class for its implementation, but we
				354	did not discuss why or when this should occur. Here we talk about the classes
				355	available, from the most general to the most specific.<p>
				356
Chris Lattner	7991070	2002-08-22 19:21:04 +0000	[diff] [blame]	357	When choosing a superclass for your Pass, you should choose the <b>most
				358	specific</b> class possible, while still being able to meet the requirements
				359	listed. This gives the LLVM Pass Infrastructure information neccesary to
				360	optimize how passes are run, so that the resultant compiler isn't unneccesarily
				361	slow.<p>
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	362
				363
				364
				365	<!-- ======================================================================= -->
				366	</ul><table width="100%" bgcolor="#441188" border=0 cellpadding=4 cellspacing=0>
				367	<tr><td> </td><td width="100%">
				368	<font color="#EEEEFF" face="Georgia,Palatino"><b>
				369	<a name="Pass">The <tt>Pass</tt> class
				370	</b></font></td></tr></table><ul>
				371
				372	The "<tt><a href="http://llvm.cs.uiuc.edu/doxygen/classPass.html">Pass</a></tt>"
				373	class is the most general of all superclasses that you can use. Deriving from
				374	<tt>Pass</tt> indicates that your pass uses the entire program as a unit,
				375	refering to function bodies in no predictable order, or adding and removing
				376	functions. Because nothing is known about the behavior of direct <tt>Pass</tt>
				377	subclasses, no optimization can be done for their execution.<p>
				378
				379	To write a correct <tt>Pass</tt> subclass, derive from <tt>Pass</tt> and
				380	overload the <tt>run</tt> method with the following signature:<p>
				381
				382	<!-- _______________________________________________________________________ -->
				383	</ul><h4><a name="run"><hr size=0>The <tt>run</tt> method</h4><ul>
				384
				385
				386	<pre>
				387	<b>virtual bool</b> run(Module &M) = 0;
				388	</pre><p>
				389
				390	The <tt>run</tt> method performs the interesting work of the pass, and should
				391	return true if the module was modified by the transformation, false
				392	otherwise.<p>
				393
				394
				395
				396	<!-- ======================================================================= -->
				397	</ul><table width="100%" bgcolor="#441188" border=0 cellpadding=4 cellspacing=0>
				398	<tr><td> </td><td width="100%">
				399	<font color="#EEEEFF" face="Georgia,Palatino"><b>
				400	<a name="FunctionPass">The <tt>FunctionPass</tt> class
				401	</b></font></td></tr></table><ul>
				402
				403	In contrast to direct <tt>Pass</tt> subclasses, direct <tt><a
				404	href="http://llvm.cs.uiuc.edu/doxygen/classPass.html">FunctionPass</a></tt>
				405	subclasses do have a predictable, local behavior that can be expected by the
				406	system. All <tt>FunctionPass</tt> execute on each function in the program
				407	independant of all of the other functions in the program.
				408	<tt>FunctionPass</tt>'s do not require that they are executed in a particular
				409	order, and <tt>FunctionPass</tt>'s do not modify external functions.<p>
				410
				411	To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:<p>
				412
				413	<ol>
				414	<li>Modify a Function other than the one currently being processed.
				415	<li>Add or remove Function's from the current Module.
				416	<li>Add or remove global variables from the current Module.
				417	<li>Maintain state across invocations of
				418	<a href="#runOnFunction"><tt>runOnFunction</tt></a> (including global data)
				419	</ol><p>
				420
				421	Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
				422	href="#basiccode">Hello World</a> pass for example). <tt>FunctionPass</tt>'s
				423	may overload three virtual methods to do their work. All of these methods
				424	should return true if they modified the program, or false if they didn't.<p>
				425
				426	<!-- _______________________________________________________________________ -->
				427	</ul><h4><a name="doInitialization"><hr size=0>The <tt>doInitialization</tt>
				428	method</h4><ul>
				429
				430	<pre>
				431	<b>virtual bool</b> doInitialization(Module &M);
				432	</pre><p>
				433
				434	The <tt>doIninitialize</tt> method is allowed to do most of the things that
				435	<tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
				436	functions, get pointers to functions, etc. The <tt>doInitialize</tt> method is
				437	designed to do simple initialization type of stuff that does not depend on the
				438	functions being processed. The <tt>doInitialization</tt> function call is not
				439	scheduled to overlap with any other pass executions.<p>
				440
				441	A good example of how this method should be used is the <a
				442	href="http://llvm.cs.uiuc.edu/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations</a>
				443	pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
				444	platform dependant <tt>malloc()</tt> and <tt>free()</tt> function calls. It
				445	uses the <tt>doInitialization</tt> method to get a reference to the malloc and
				446	free functions that it needs, adding prototypes to the module if neccesary.<p>
				447
				448	<!-- _______________________________________________________________________ -->
				449	</ul><h4><a name="runOnFunction"><hr size=0>The <tt>runOnFunction</tt> method</h4><ul>
				450
				451	<pre>
				452	<b>virtual bool</b> runOnFunction(Function &F) = 0;
				453	</pre><p>
				454
				455	The <tt>runOnFunction</tt> method must be implemented by your subclass to do the
				456	transformation or analysis work of your pass. As usual, a true value should be
				457	returned if the function is modified.<p>
				458
				459	<!-- _______________________________________________________________________ -->
				460	</ul><h4><a name="doFinalization"><hr size=0>The <tt>doFinalization</tt> method</h4><ul>
				461
				462	<pre>
				463	<b>virtual bool</b> doFinalization(Module &M);
				464	</pre</p>
				465
				466	The <tt>doFinalization</tt> method is an infrequently used method that is called
				467	when the pass framework has finished calling <a
				468	href="#runOnFunction"><tt>runOnFunction</tt></a> for every function in the
				469	program being compiled.<p>
				470
				471
				472
				473	<!-- ======================================================================= -->
				474	</ul><table width="100%" bgcolor="#441188" border=0 cellpadding=4 cellspacing=0>
				475	<tr><td> </td><td width="100%">
				476	<font color="#EEEEFF" face="Georgia,Palatino"><b>
				477	<a name="BasicBlockPass">The <tt>BasicBlockPass</tt> class</a>
				478	</b></font></td></tr></table><ul>
				479
				480	<tt>BasicBlockPass</tt>'s are just like <a
				481	href="#FunctionPass"><tt>FunctionPass</tt></a>'s, except that they must limit
				482	their scope of inspection and modification to a single basic block at a time.
				483	As such, they are <b>not</b> allowed to do any of the following:<p>
				484
				485	<ol>
				486	<li>Modify or inspect any basic blocks outside of the current one
				487	<li>Maintain state across invocations of
				488	<a href="#runOnBasicBlock"><tt>runOnBasicBlock</tt></a>
				489	<li>Modify the constrol flow graph (by altering terminator instructions)
				490	<li>Any of the things verboten for
				491	<a href="#FunctionPass"><tt>FunctionPass</tt></a>'s.
				492	</ol><p>
				493
				494	<tt>BasicBlockPass</tt>'s are useful for traditional local and "peephole"
				495	optimizations. They may override the same <a
				496	href="#doInitialization"><tt>doInitialization</tt></a> and <a
				497	href="#doFinalization"><tt>doFinalization</tt></a> methods that <a
				498	href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have a
				499	<tt>runOnBasicBlock</tt> method:<p>
				500
				501	<!-- _______________________________________________________________________ -->
				502	</ul><h4><a name="runOnBasicBlock"><hr size=0>The <tt>runOnBasicBlock</tt> method</h4><ul>
				503
				504	<pre>
				505	<b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
				506	</pre><p>
				507
				508	Override this function to do the work of the <tt>BasicBlockPass</tt>. This
				509	function is not allowed to inspect or modify basic blocks other than the
				510	parameter, and are not allowed to modify the CFG. A true value must be returned
				511	if the basic block is modified.<p>
				512
				513
				514	<!-- *********************************************************************** -->
				515	</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				516	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
				517	<a name="registration">Pass registration
				518	</b></font></td></tr></table><ul>
				519	<!-- *********************************************************************** -->
				520
				521	In the <a href="#basiccode">Hello World</a> example pass we illustrated how pass
				522	registration works, and discussed some of the reasons that it is used and what
				523	it does. Here we discuss how and why passes are registered.<p>
				524
				525	Passes can be registered in several different ways. Depending on the general
				526	classification of the pass, you should use one of the following templates to
				527	register the pass:<p>
				528
				529	<ul>
				530	<li><b><tt>RegisterOpt</tt></b> - This template should be used when you are
				531	registering a pass that logically should be available for use in the
				532	'<tt>opt</tt>' utility.<p>
				533
				534	<li><b><tt>RegisterAnalysis</tt></b> - This template should be used when you are
				535	registering a pass that logically should be available for use in the
				536	'<tt>analysis</tt>' utility.<p>
				537
				538	<li><b><tt>RegisterLLC</tt></b> - This template should be used when you are
				539	registering a pass that logically should be available for use in the
				540	'<tt>llc</tt>' utility.<p>
				541
				542	<li><b><tt>RegisterPass</tt></b> - This is the generic form of the
				543	<tt>Register*</tt> templates that should be used if you want your pass listed by
				544	multiple or no utilities. This template takes an extra third argument that
				545	specifies which tools it should be listed in. See the <a
				546	href="http://llvm.cs.uiuc.edu/doxygen/PassSupport_8h-source.html">PassSupport.h</a>
				547	file for more information.<p>
				548	</ul><p>
				549
				550	Regardless of how you register your pass, you must specify at least two
				551	parameters. The first parameter is the name of the pass that is to be used on
				552	the command line to specify that the pass should be added to a program (for
				553	example <tt>opt</tt> or <tt>analyze</tt>). The second argument is the name of
				554	the pass, which is to be used for the <tt>--help</tt> output of programs, as
				555	well as for debug output generated by the <tt>--debug-pass</tt> option.<p>
				556
				557	If you pass is constructed by its default constructor, you only ever have to
				558	pass these two arguments. If, on the other hand, you require other information
				559	(like target specific information), you must pass an additional argument. This
				560	argument is a pointer to a function used to create the pass. For an example of
				561	how this works, look at the <a
				562	href="http://llvm.cs.uiuc.edu/doxygen/LowerAllocations_8cpp-source.html">LowerAllocations.cpp</a>
				563	file.<p>
				564
				565	If a pass is registered to be used by the <tt>analyze</tt> utility, you should
				566	implement the virtual <tt>print</tt> method:<p>
				567
				568	<!-- _______________________________________________________________________ -->
				569	</ul><h4><a name="print"><hr size=0>The <tt>print</tt> method</h4><ul>
				570
				571	<pre>
				572	<b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
				573	</pre><p>
				574
				575	The <tt>print</tt> method must be implemented by "analyses" in order to print a
				576	human readable version of the analysis results. This is useful for debugging an
				577	analysis itself, as well as for other people to figure out how an analysis
				578	works. The <tt>analyze</tt> tool uses this method to generate its output.<p>
				579
				580	The <tt>ostream</tt> parameter specifies the stream to write the results on, and
				581	the <tt>Module</tt> parameter gives a pointer to the top level module of the
				582	program that has been analyzed. Note however that this pointer may be null in
				583	certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
				584	debugger), so it should only be used to enhance debug output, it should not be
				585	depended on.<p>
				586
				587
				588	<!-- *********************************************************************** -->
				589	</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				590	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
				591	<a name="interaction">Specifying interactions between passes
				592	</b></font></td></tr></table><ul>
				593	<!-- *********************************************************************** -->
				594
				595	One of the main responsibilities of the <tt>PassManager</tt> is the make sure
				596	that passes interact with each other correctly. Because <tt>PassManager</tt>
				597	tries to <a href="#passmanager">optimize the execution of passes</a> it must
				598	know how the passes interact with each other and what dependencies exist between
				599	the various passes. To track this, each pass can declare the set of passes that
				600	are required to be executed before the current pass, and the passes which are
				601	invalidated by the current pass.<p>
				602
				603	Typically this functionality is used to require that analysis results are
				604	computed before your pass is run. Running arbitrary transformation passes can
				605	invalidate the computed analysis results, which is what the invalidation set
				606	specifies. If a pass does not implement the <tt><a
				607	href="#getAnalysisUsage">getAnalysisUsage</a></tt> method, it defaults to not
				608	having any prerequisite passes, and invalidating <b>all</b> other passes.<p>
				609
				610
				611	<!-- _______________________________________________________________________ -->
				612	</ul><h4><a name="getAnalysisUsage"><hr size=0>The <tt>getAnalysisUsage</tt> method</h4><ul>
				613
				614	<pre>
				615	<b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
				616	</pre><p>
				617
				618	By implementing the <tt>getAnalysisUsage</tt> method, the required and
				619	invalidated sets may be specified for your transformation. The implementation
				620	should fill in the <tt><a
				621	href="http://llvm.cs.uiuc.edu/doxygen/classAnalysisUsage.html">AnalysisUsage</a></tt>
				622	object with information about which passes are required and not invalidated. To do this, the following set methods are provided by the <tt><a
				623	href="http://llvm.cs.uiuc.edu/doxygen/classAnalysisUsage.html">AnalysisUsage</a></tt> class:<p>
				624
				625	<pre>
				626	<i>// addRequires - Add the specified pass to the required set for your pass.</i>
				627	<b>template</b><<b>class</b> PassClass>
				628	AnalysisUsage &AnalysisUsage::addRequired();
				629
				630	<i>// addPreserved - Add the specified pass to the set of analyses preserved by
				631	// this pass</i>
				632	<b>template</b><<b>class</b> PassClass>
				633	AnalysisUsage &AnalysisUsage::addPreserved();
				634
				635	<i>// setPreservesAll - Call this if the pass does not modify its input at all</i>
				636	<b>void</b> AnalysisUsage::setPreservesAll();
				637
				638	<i>// preservesCFG - This function should be called by the pass, iff they do not:
				639	//
				640	// 1. Add or remove basic blocks from the function
				641	// 2. Modify terminator instructions in any way.
				642	//
				643	// This is automatically implied for <a href="#BasicBlockPass">BasicBlockPass</a>'s
				644	//</i>
				645	<b>void</b> AnalysisUsage::preservesCFG();
				646	</pre><p>
				647
				648	Some examples of how to use these methods are:<p>
				649
				650	<pre>
				651	<i>// This is an example implementation from an analysis, which does not modify
				652	// the program at all, yet has a prerequisite.</i>
				653	<b>void</b> <a href="http://llvm.cs.uiuc.edu/doxygen/structPostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
				654	AU.setPreservesAll();
				655	AU.addRequired<<a href="http://llvm.cs.uiuc.edu/doxygen/structPostDominatorTree.html">PostDominatorTree</a>>();
				656	}
				657	</pre><p>
				658
				659	and:<p>
				660
				661	<pre>
				662	<i>// This example modifies the program, but does not modify the CFG</i>
				663	<b>void</b> <a href="http://llvm.cs.uiuc.edu/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
				664	AU.preservesCFG();
				665	AU.addRequired<<a href="http://llvm.cs.uiuc.edu/doxygen/classLoopInfo.html">LoopInfo</a>>();
				666	}
				667	</pre><p>
				668
				669	<!-- _______________________________________________________________________ -->
				670	</ul><h4><a name="getAnalysis"><hr size=0>The <tt>getAnalysis<></tt> method</h4><ul>
				671
				672	The <tt>Pass::getAnalysis<></tt> method is inherited by your class,
				673	providing you with access to the passes that you declared that you required with
				674	the <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method. It takes
				675	a single template argument that specifies which pass class you want, and returns
				676	a reference to that pass.<p>
				677
				678	<pre>
				679	<b>template</b><<b>typename</b> PassClass>
				680	AnalysisType &getAnalysis();
				681	</pre><p>
				682
				683	This method call returns a reference to the pass desired. You may get a runtime
				684	assertion failure if you attempt to get an analysis that you did not declare as
				685	required in your <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a>
				686	implementation. This method can be called by your <tt>run*</tt> method
				687	implementation, or by any other local method invoked by your <tt>run*</tt>
				688	method.<p>
				689
Chris Lattner	7991070	2002-08-22 19:21:04 +0000	[diff] [blame]	690	<!-- *********************************************************************** -->
				691	</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				692	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
				693	<a name="analysisgroup">Implementing Analysis Groups
				694	</b></font></td></tr></table><ul>
				695	<!-- *********************************************************************** -->
				696
				697	Now that we understand the basics of how passes are defined, how the are used,
				698	and how they are required from other passes, it's time to get a little bit
				699	fancier. All of the pass relationships that we have seen so far are very
				700	simple: one pass depends on one other specific pass to be run before it can run.
				701	For many applications, this is great, for others, more flexibility is
				702	required.<p>
				703
				704	In particular, some analyses are defined such that there is a single simple
				705	interface to the analysis results, but multiple ways of calculating them.
				706	Consider alias analysis for example. The most trivial alias analysis returns
				707	"may alias" for any alias query. The most sophisticated analysis a
				708	flow-sensitive, context-sensitive interprocedural analysis that can take a
				709	significant amount of time to execute (and obviously, there is a lot of room
				710	between these two extremes for other implementations). To cleanly support
				711	situations like this, the LLVM Pass Infrastructure supports the notion of
				712	Analysis Groups.<p>
				713
				714	<!-- _______________________________________________________________________ -->
				715	</ul><h4><a name="agconcepts"><hr size=0>Analysis Group Concepts</h4><ul>
				716
				717	An Analysis Group is a single simple interface that may be implemented by
				718	multiple different passes. Analysis Groups can be given human readable names
				719	just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
				720	class. An analysis group may have one or more implementations, one of which is
				721	the "default" implementation.<p>
				722
				723	Analysis groups are used by client passes just like other passes are: the
				724	<tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
				725	In order to resolve this requirement, the <a href="#passmanager">PassManager</a>
				726	scans the available passes to see if any implementations of the analysis group
				727	are available. If none is available, the default implementation is created for
				728	the pass to use. All standard rules for <A href="#interaction">interaction
				729	between passes</a> still apply.<p>
				730
				731	Although <a href="#registration">Pass Registration</a> is optional for normal
				732	passes, all analysis group implementations must be registered, and must use the
				733	<A href="#registerag"><tt>RegisterAnalysisGroup</tt></a> template to join the
				734	implementation pool. Also, a default implementation of the interface
				735	<b>must</b> be registered with <A
				736	href="#registerag"><tt>RegisterAnalysisGroup</tt></a>.<p>
				737
				738	As a concrete example of an Analysis Group in action, consider the <a
				739	href="http://llvm.cs.uiuc.edu/doxygen/structAliasAnalysis.html">AliasAnalysis</a>
				740	analysis group. The default implementation of the alias analysis interface (the
				741	<tt><a
				742	href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">basicaa</a></tt>
				743	pass) just does a few simple checks that don't require significant analysis to
				744	compute (such as: two different globals can never alias each other, etc).
				745	Passes that use the <tt><a
				746	href="http://llvm.cs.uiuc.edu/doxygen/structAliasAnalysis.html">AliasAnalysis</a></tt>
				747	interface (for example the <tt><a
				748	href="http://llvm.cs.uiuc.edu/doxygen/classGCSE.html">gcse</a></tt> pass), do not care which implementation
				749	of alias analysis is actually provided, they just use the designated
				750	interface.<p>
				751
				752	From the user's perspective, commands work just like normal. Issuing the
				753	command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
				754	instantiated and added to the pass sequence. Issuing the command '<tt>opt
				755	-somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
				756	<tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
				757	hypothetical example) instead.<p>
				758
				759
				760	<!-- _______________________________________________________________________ -->
				761	</ul><h4><a name="registerag"><hr size=0>Using <tt>RegisterAnalysisGroup</tt></h4><ul>
				762
				763	The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
				764	group itself as well as add pass implementations to the analysis group. First,
				765	an analysis should be registered, with a human readable name provided for it.
				766	Unlike registration of passes, there is no command line argument to be specified
				767	for the Analysis Group Interface itself, because it is "abstract":<p>
				768
				769	<pre>
				770	<b>static</b> RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/structAliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
				771	</pre><p>
				772
				773	Once the analysis is registered, passes can declare that they are valid
				774	implementations of the interface by using the following code:<p>
				775
				776	<pre>
				777	<b>namespace</b> {
				778	//<i> Analysis Group implementations <b>must</b> be registered normally...</i>
				779	RegisterOpt<FancyAA>
				780	B("<i>somefancyaa</i>", "<i>A more complex alias analysis implementation</i>");
				781
				782	//<i> Declare that we implement the AliasAnalysis interface</i>
				783	RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/structAliasAnalysis.html">AliasAnalysis</a>, FancyAA> C;
				784	}
				785	</pre><p>
				786
				787	This just shows a class <tt>FancyAA</tt> that is registered normally, then uses
				788	the <tt>RegisterAnalysisGroup</tt> template to "join" the <tt><a
				789	href="http://llvm.cs.uiuc.edu/doxygen/structAliasAnalysis.html">AliasAnalysis</a></tt>
				790	analysis group. Every implementation of an analysis group should join using
				791	this template. A single pass may join multiple different analysis groups with
				792	no problem.<p>
				793
				794	<pre>
				795	<b>namespace</b> {
				796	//<i> Analysis Group implementations <b>must</b> be registered normally...</i>
				797	RegisterOpt<<a href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
				798	D("<i>basicaa</i>", "<i>Basic Alias Analysis (default AA impl)</i>");
				799
				800	//<i> Declare that we implement the AliasAnalysis interface</i>
				801	RegisterAnalysisGroup<<a href="http://llvm.cs.uiuc.edu/doxygen/structAliasAnalysis.html">AliasAnalysis</a>, <a href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>, <b>true</b>> E;
				802	}
				803	</pre><p>
				804
				805	Here we show how the default implementation is specified (using the extra
				806	argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
				807	one default implementation available at all times for an Analysis Group to be
				808	used. Here we declare that the <tt><a
				809	href="http://llvm.cs.uiuc.edu/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
				810	pass is the default implementation for the interface.<p>
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	811
				812
				813	<!-- *********************************************************************** -->
				814	</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				815	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
				816	<a name="passmanager">What PassManager does
				817	</b></font></td></tr></table><ul>
				818	<!-- *********************************************************************** -->
				819
				820	The <a
				821	href="http://llvm.cs.uiuc.edu/doxygen/PassManager_8h-source.html"><tt>PassManager</tt></a>
				822	<a href="http://llvm.cs.uiuc.edu/doxygen/classPassManager.html">class</a> takes
				823	a list of passes, ensures their <a href="#interaction">prerequisites</a> are set
				824	up correctly, and then schedules passes to run efficiently. All of the LLVM
				825	tools that run passes use the <tt>PassManager</tt> for execution of these
				826	passes.<p>
				827
				828	The <tt>PassManager</tt> does two main things to try to reduce the execution
				829	time of a series of passes:<p>
				830
				831	<ol>
				832	<li><b>Share analysis results</b> - The PassManager attempts to avoid
				833	recomputing analysis results as much as possible. This means keeping track of
				834	which analyses are available already, which analyses get invalidated, and which
				835	analyses are needed to be run for a pass. An important part of work is that the
				836	<tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
				837	it to <a href="#releaseMemory">free memory</a> allocated to holding analysis
				838	results as soon as they are no longer needed.<p>
				839
				840	<li><b>Pipeline the execution of passes on the program</b> - The
				841	<tt>PassManager</tt> attempts to get better cache and memory usage behavior out
				842	of a series of passes by pipelining the passes together. This means that, given
				843	a series of consequtive <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s, it
				844	will execute all of the <a href="#FunctionPass"><tt>FunctionPass</tt></a>'s on
				845	the first function, then all of the <a
				846	href="#FunctionPass"><tt>FunctionPass</tt></a>'s on the second function,
				847	etc... until the entire program has been run through the passes.<p>
				848
				849	This improves the cache behavior of the compiler, because it is only touching
				850	the LLVM program representation for a single function at a time, instead of
				851	traversing the entire program. It reduces the memory consumption of compiler,
				852	because, for example, only one <a
				853	href="http://llvm.cs.uiuc.edu/doxygen/structDominatorSet.html"><tt>DominatorSet</tt></a>
				854	needs to be calculated at a time. This also makes it possible some <a
				855	href="#SMP">interesting enhancements</a> in the future.<p>
				856
				857	</ol><p>
				858
				859	The effectiveness of the <tt>PassManager</tt> is influenced directly by how much
				860	information it has about the behaviors of the passes it is scheduling. For
				861	example, the "preserved" set is intentionally conservative in the face of an
				862	unimplemented <a href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method.
				863	Not implementing when it should be implemented will have the effect of not
				864	allowing any analysis results to live across the execution of your pass.<p>
				865
				866	The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
				867	options that is useful for debugging pass execution, seeing how things work, and
				868	diagnosing when you should be preserving more analyses than you currently are
				869	(To get information about all of the variants of the <tt>--debug-pass</tt>
				870	option, just type '<tt>opt --help-hidden</tt>').<p>
				871
				872	By using the <tt>--debug-pass=Structure</tt> option, for example, we can see how
				873	our <a href="#basiccode">Hello World</a> pass interacts with other passes. Lets
				874	try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:<p>
				875
				876	<pre>
				877	$ opt -load ../../../lib/Debug/libhello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
				878	Module Pass Manager
				879	Function Pass Manager
				880	Dominator Set Construction
				881	Immediate Dominators Construction
				882	Global Common Subexpression Elimination
				883	-- Immediate Dominators Construction
				884	-- Global Common Subexpression Elimination
				885	Natural Loop Construction
				886	Loop Invariant Code Motion
				887	-- Natural Loop Construction
				888	-- Loop Invariant Code Motion
				889	Module Verifier
				890	-- Dominator Set Construction
				891	-- Module Verifier
				892	Bytecode Writer
				893	--Bytecode Writer
				894	</pre><p>
				895
				896	This output shows us when passes are constructed and when the analysis results
				897	are known to be dead (prefixed with '<tt>--</tt>'). Here we see that GCSE uses
				898	dominator and immediate dominator information to do its job. The LICM pass uses
				899	natural loop information, which uses dominator sets, but not immediate
				900	dominators. Because immediate dominators are no longer useful after the GCSE
				901	pass, it is immediately destroyed. The dominator sets are then reused to
				902	compute natural loop information, which is then used by the LICM pass.<p>
				903
				904	After the LICM pass, the module verifier runs (which is automatically added by
				905	the '<tt>opt</tt>' tool), which uses the dominator set to check that the
				906	resultant LLVM code is well formed. After it finishes, the dominator set
				907	information is destroyed, after being computed once, and shared by three
				908	passes.<p>
				909
				910	Lets see how this changes when we run the <a href="#basiccode">Hello World</a>
				911	pass in between the two passes:<p>
				912
				913	<pre>
				914	$ opt -load ../../../lib/Debug/libhello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
				915	Module Pass Manager
				916	Function Pass Manager
				917	Dominator Set Construction
				918	Immediate Dominators Construction
				919	Global Common Subexpression Elimination
				920	<b>-- Dominator Set Construction</b>
				921	-- Immediate Dominators Construction
				922	-- Global Common Subexpression Elimination
				923	<b> Hello World Pass
				924	-- Hello World Pass
				925	Dominator Set Construction</b>
				926	Natural Loop Construction
				927	Loop Invariant Code Motion
				928	-- Natural Loop Construction
				929	-- Loop Invariant Code Motion
				930	Module Verifier
				931	-- Dominator Set Construction
				932	-- Module Verifier
				933	Bytecode Writer
				934	--Bytecode Writer
				935	Hello: __main
				936	Hello: puts
				937	Hello: main
				938	</pre><p>
				939
				940	Here we see that the <a href="#basiccode">Hello World</a> pass has killed the
				941	Dominator Set pass, even though it doesn't modify the code at all! To fix this,
				942	we need to add the following <a
				943	href="#getAnalysisUsage"><tt>getAnalysisUsage</tt></a> method to our pass:<p>
				944
				945	<pre>
				946	<i>// We don't modify the program, so we preserve all analyses</i>
				947	<b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
				948	AU.setPreservesAll();
				949	}
				950	</pre><p>
				951
				952	Now when we run our pass, we get this output:<p>
				953
				954	<pre>
				955	$ opt -load ../../../lib/Debug/libhello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
				956	Pass Arguments: -gcse -hello -licm
				957	Module Pass Manager
				958	Function Pass Manager
				959	Dominator Set Construction
				960	Immediate Dominators Construction
				961	Global Common Subexpression Elimination
				962	-- Immediate Dominators Construction
				963	-- Global Common Subexpression Elimination
				964	Hello World Pass
				965	-- Hello World Pass
				966	Natural Loop Construction
				967	Loop Invariant Code Motion
				968	-- Loop Invariant Code Motion
				969	-- Natural Loop Construction
				970	Module Verifier
				971	-- Dominator Set Construction
				972	-- Module Verifier
				973	Bytecode Writer
				974	--Bytecode Writer
				975	Hello: __main
				976	Hello: puts
				977	Hello: main
				978	</pre><p>
				979
				980	Which shows that we don't accidentally invalidate dominator information
				981	anymore, and therefore do not have to compute it twice.<p>
				982
				983
				984	<!-- _______________________________________________________________________ -->
				985	</ul><h4><a name="releaseMemory"><hr size=0>The <tt>releaseMemory</tt> method</h4><ul>
				986
				987	<pre>
				988	<b>virtual void</b> releaseMemory();
				989	</pre><p>
				990
				991	The <tt>PassManager</tt> automatically determines when to compute analysis
				992	results, and how long to keep them around for. Because the lifetime of the pass
				993	object itself is effectively the entire duration of the compilation process, we
				994	need some way to free analysis results when they are no longer useful. The
				995	<tt>releaseMemory</tt> virtual method is the way to do this.<p>
				996
				997	If you are writing an analysis or any other pass that retains a significant
				998	amount of state (for use by another pass which "requires" your pass and uses the
				999	<a href="#getAnalysis">getAnalysis</a> method) you should implement
				1000	<tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
				1001	internal state. This method is called after the <tt>run*</tt> method for the
				1002	class, before the next call of <tt>run*</tt> in your pass.<p>
				1003
				1004
				1005	<!-- *********************************************************************** -->
				1006	</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				1007	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
Chris Lattner	480e2ef	2002-09-06 02:02:58 +0000	[diff] [blame]	1008	<a name="debughints">Using GDB with dynamically loaded passes
				1009	</b></font></td></tr></table><ul>
				1010	<!-- *********************************************************************** -->
				1011
				1012	Unfortunately, using GDB with dynamically loaded passes is not as easy as it
				1013	should be. First of all, you can't set a breakpoint in a shared object that has
				1014	not been loaded yet, and second of all there are problems with inlined functions
				1015	in shared objects. Here are some suggestions to debugging your pass with
				1016	GDB.<p>
				1017
				1018	For sake of discussion, I'm going to assume that you are debugging a
				1019	transformation invoked by <tt>opt</tt>, although nothing described here depends
				1020	on that.<p>
				1021
				1022	<!-- _______________________________________________________________________ -->
				1023	</ul><h4><a name="breakpoint"><hr size=0>Setting a breakpoint in your pass</h4><ul>
				1024
				1025	First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:<p>
				1026
				1027	<pre>
				1028	$ <b>gdb opt</b>
				1029	GNU gdb 5.0
				1030	Copyright 2000 Free Software Foundation, Inc.
				1031	GDB is free software, covered by the GNU General Public License, and you are
				1032	welcome to change it and/or distribute copies of it under certain conditions.
				1033	Type "show copying" to see the conditions.
				1034	There is absolutely no warranty for GDB. Type "show warranty" for details.
				1035	This GDB was configured as "sparc-sun-solaris2.6"...
				1036	(gdb)
				1037	</pre><p>
				1038
				1039	Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
				1040	time to load. Be patient. Since we cannot set a breakpoint in our pass yet
				1041	(the shared object isn't loaded until runtime), we must execute the process, and
				1042	have it stop before it invokes our pass, but after it has loaded the shared
				1043	object. The most foolproof way of doing this is to set a breakpoint in
				1044	<tt>PassManager::run</tt> and then run the process with the arguments you
				1045	want:<p>
				1046
				1047	<pre>
				1048	(gdb) <b>break PassManager::run</b>
				1049	Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
				1050	(gdb) <b>run test.bc -load /shared/lattner/cvs/llvm/lib/Debug/[libname].so -[passoption]</b>
				1051	Starting program: /shared/lattner/cvs/llvm/tools/Debug/opt test.bc
				1052	-load /shared/lattner/cvs/llvm/lib/Debug/[libname].so -[passoption]
				1053	Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
				1054	70 bool PassManager::run(Module &M) { return PM->run(M); }
				1055	(gdb)
				1056	</pre></p>
				1057
				1058	Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are now
				1059	free to set breakpoints in your pass so that you can trace through execution or
				1060	do other standard debugging stuff.<p>
				1061
				1062
				1063	<!-- _______________________________________________________________________ -->
				1064	</ul><h4><a name="debugmisc"><hr size=0>Miscellaneous Problems</h4><ul>
				1065
				1066	Once you have the basics down, there are a couple of problems that GDB has, some
				1067	with solutions, some without.<p>
				1068
				1069	<ul>
				1070	<li>Inline functions have bogus stack information. In general, GDB does a
				1071	pretty good job getting stack traces and stepping through inline functions.
				1072	When a pass is dynamically loaded however, it somehow completely loses this
				1073	capability. The only solution I know of is to de-inline a function (move it
				1074	from the body of a class to a .cpp file).<p>
				1075
				1076	<li>Restarting the program breaks breakpoints. After following the information
				1077	above, you have succeeded in getting some breakpoints planted in your pass. Nex
				1078	thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
				1079	and you start getting errors about breakpoints being unsettable. The only way I
				1080	have found to "fix" this problem is to <tt>delete</tt> the breakpoints that are
				1081	already set in your pass, run the program, and re-set the breakpoints once
				1082	execution stops in <tt>PassManager::run</tt>.<p>
				1083
				1084	</ul>
				1085
				1086	Hopefully these tips will help with common case debugging situations. If you'd
				1087	like to contribute some tips of your own, just contact <a
				1088	href="mailto:sabre@nondot.org">Chris</a>.<p>
				1089
				1090
				1091	<!-- *********************************************************************** -->
				1092	</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
				1093	<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	1094	<a name="future">Future extensions planned
				1095	</b></font></td></tr></table><ul>
				1096	<!-- *********************************************************************** -->
				1097
				1098	Although the LLVM Pass Infrastructure is very capable as it stands, and does
				1099	some nifty stuff, there are things we'd like to add in the future. Here is
				1100	where we are going:<p>
				1101
				1102	<!-- _______________________________________________________________________ -->
				1103	</ul><h4><a name="SMP"><hr size=0>Multithreaded LLVM</h4><ul>
				1104
				1105	Multiple CPU machines are becoming more command and compilation can never be
				1106	fast enough: obviously we should allow for a multithreaded compiler. Because of
				1107	the semantics defined for passes above (specifically they cannot maintain state
				1108	across invocations of their <tt>run*</tt> methods), a nice clean way to
				1109	implement a multithreaded compiler would be for the <tt>PassManager</tt> class
				1110	to create multiple instances of each pass object, and allow the seperate
				1111	instances to be hacking on different parts of the program at the same time.<p>
				1112
				1113	This implementation would prevent each of the passes from having to implement
				1114	multithreaded constructs, requiring only the LLVM core to have locking in a few
				1115	places (for global resources). Although this is a simple extension, we simply
				1116	haven't had time (or multiprocessor machines, thus a reason) to implement this.
				1117	Despite that, we have kept the LLVM passes SMP ready, and you should too.<p>
				1118
				1119
				1120	<!-- _______________________________________________________________________ -->
				1121	</ul><h4><a name="ModuleSource"><hr size=0>A new <tt>ModuleSource</tt> interface</h4><ul>
				1122
				1123	Currently, the <tt>PassManager</tt>'s <tt>run</tt> method takes a <tt><a
				1124	href="http://llvm.cs.uiuc.edu/doxygen/classModule.html">Module</a></tt> as
				1125	input, and runs all of the passes on this module. The problem with this
				1126	approach is that none of the <tt>PassManager</tt> features can be used for
				1127	timing and debugging the actual <b>loading</b> of the module from disk or
				1128	standard input.<p>
				1129
				1130	To solve this problem, eventually the <tt>PassManger</tt> class will accept a
				1131	<tt>ModuleSource</tt> object instead of a Module itself. When complete, this
				1132	will also allow for streaming of functions out of the bytecode representation,
				1133	allowing us to avoid holding the entire program in memory at once if we only are
				1134	dealing with <a href="#FunctionPass">FunctionPass</a>'s.<p>
				1135
				1136	As part of a different issue, eventually the bytecode loader will be extended to
				1137	allow on-demand loading of functions from the bytecode representation, in order
				1138	to better support the runtime reoptimizer. The bytecode format is already
				1139	capable of this, the loader just needs to be reworked a bit.<p>
				1140
				1141
				1142	<!-- _______________________________________________________________________ -->
				1143	</ul><h4><a name="PassFunctionPass"><hr size=0><tt>Pass</tt>'s requiring <tt>FunctionPass</tt>'s</h4><ul>
				1144
				1145	Currently it is illegal for a <a href="#Pass"><tt>Pass</tt></a> to require a <a
				1146	href="#FunctionPass"><tt>FunctionPass</tt></a>. This is because there is only
				1147	one instance of the <a href="#FunctionPass"><tt>FunctionPass</tt></a> object
				1148	ever created, thus nowhere to store information for all of the functions in the
				1149	program at the same time. Although this has come up a couple of times before,
				1150	this has always been worked around by factoring one big complicated pass into a
				1151	global and an interprocedural part, both of which are distinct. In the future,
				1152	it would be nice to have this though.<p>
				1153
				1154	Note that it is no problem for a <a
				1155	href="#FunctionPass"><tt>FunctionPass</tt></a> to require the results of a <a
				1156	href="#Pass"><tt>Pass</tt></a>, only the other way around.<p>
				1157
				1158
				1159	<!-- *********************************************************************** -->
				1160	</ul>
				1161	<!-- *********************************************************************** -->
				1162
				1163	<hr><font size-1>
Chris Lattner	480e2ef	2002-09-06 02:02:58 +0000	[diff] [blame]	1164	<address><a href="mailto:sabre@nondot.org">Chris Lattner</a></address>
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	1165	<!-- Created: Tue Aug 6 15:00:33 CDT 2002 -->
				1166	<!-- hhmts start -->
Chris Lattner	480e2ef	2002-09-06 02:02:58 +0000	[diff] [blame]	1167	Last modified: Thu Sep 5 15:06:01 CDT 2002
Chris Lattner	c6bb824	2002-08-08 20:11:18 +0000	[diff] [blame]	1168	<!-- hhmts end -->
				1169	</font></body></html>