| blob | ed6b4d662a1ba082ad25b9c5a724695f76fc67e0 |
3 <html>
4 <head>
8 </head>
9 <body>
12 Writing an LLVM Pass
13 </div>
15 <ol>
18 <ul>
22 </ul></li>
24 <ul>
27 <ul>
29 </ul></li>
31 <ul>
37 </ul></li>
39 <ul>
45 </ul></li>
47 <ul>
53 </ul></li>
55 <ul>
59 method</a></li>
62 </ul></li>
64 class</a>
65 <ul>
68 </ul></li>
69 </ul>
71 <ul>
73 </ul></li>
75 <ul>
77 method</a></li>
78 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
83 </ul></li>
85 <ul>
88 </ul></li>
91 <ul>
93 </ul></li>
95 <ul>
98 </ul></li>
100 <ul>
103 </ul></li>
105 <ul>
107 </ul></li>
108 </ol>
113 </div>
115 <!-- *********************************************************************** -->
118 </div>
119 <!-- *********************************************************************** -->
123 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
124 passes are where most of the interesting parts of the compiler exist. Passes
125 perform the transformations and optimizations that make up the compiler, they
126 build the analysis results that are used by these transformations, and they are,
127 above all, a structuring technique for compiler code.</p>
131 class, which implement functionality by overriding virtual methods inherited
138 more information about what your pass does, and how it can be combined with
139 other passes. One of the main features of the LLVM Pass Framework is that it
140 schedules passes to run in an efficient way based on the constraints that your
141 pass meets (which are indicated by which class they derive from).</p>
143 <p>We start by showing you how to construct a pass, everything from setting up
144 the code, to compiling, loading, and executing it. After the basics are down,
145 more advanced features are discussed.</p>
147 </div>
149 <!-- *********************************************************************** -->
152 </div>
153 <!-- *********************************************************************** -->
158 is designed to simply print out the name of non-external functions that exist in
159 the program being compiled. It does not modify the program at all, it just
160 inspects it. The source code and files for this pass are available in the LLVM
163 </div>
165 <!-- ======================================================================= -->
168 </div>
172 <p>First, you need to create a new directory somewhere in the LLVM source
173 base. For this example, we'll assume that you made
175 (Makefile) that will compile the source code for the new pass. To do this,
177 <hr/>
180 # Makefile for hello pass
182 # Path to top level of LLVM heirarchy
183 LEVEL = ../../..
185 # Name of the library to build
186 LIBRARYNAME = Hello
188 # Make the shared library become a loadable module so the tools can
189 # dlopen/dlsym on the resulting library.
190 LOADABLE_MODULE = 1
192 # Tell the build system which LLVM libraries your pass needs. You'll probably
193 # need at least LLVMSystem.a, LLVMSupport.a, LLVMCore.a but possibly several
194 # others too.
195 LLVMLIBS = LLVMCore.a LLVMSupport.a LLVMSystem.a
197 # Include the makefile implementation stuff
198 include $(LEVEL)/Makefile.common
199 </pre></div>
202 directory are to be compiled and linked together into a
205 If your operating system uses a suffix other than .so (such as windows or
206 Mac OS/X), the appropriate extension will be used.</p>
208 <p>Now that we have the build scripts set up, we just need to write the code for
209 the pass itself.</p>
211 </div>
213 <!-- ======================================================================= -->
216 </div>
220 <p>Now that we have a way to compile our new pass, we just have to write it.
221 Start out with:</p>
226 </pre></div>
230 we are operating on <tt><a
236 </pre></div>
237 <p>... which is required because the functions from the include files
238 live in the llvm namespace.
239 </p>
245 </pre></div>
247 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
248 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
249 things declared inside of the anonymous namespace only visible to the current
250 file. If you're not familiar with them, consult a decent C++ book for more
251 information.</p>
257 </pre></div><p>
261 The different builtin pass subclasses are described in detail <a
264 time.</p>
267 static char ID;
268 Hello() : FunctionPass(&ID) {}
269 </pre></div><p>
271 <p> This declares pass identifier used by LLVM to identify pass. This allows LLVM to
272 avoid using expensive C++ runtime information.</p>
278 }
280 </pre></div>
283 which overloads an abstract virtual method inherited from <a
285 to do our thing, so we just print out our message with the name of each
286 function.</p>
289 char Hello::ID = 0;
290 </pre></div>
292 <p> We initialize pass ID here. LLVM uses ID's address to identify pass so
293 initialization value is not important.</p>
297 false /* Only looks at CFG */,
298 false /* Analysis Pass */);
300 </pre></div>
303 giving it a command line
305 Last two RegisterPass arguments are optional. Their default value is false.
306 If a pass walks CFG without modifying it then third argument is set to true.
307 If a pass is an analysis pass, for example dominator tree pass, then true
308 is supplied as fourth argument. </p>
321 static char ID;
322 Hello() : FunctionPass(&ID) {}
327 }
328 };
330 char Hello::ID = 0;
332 }
333 </pre></div>
336 command in the local directory and you should get a new
337 "<tt>Debug/lib/Hello.so</tt> file. Note that everything in this file is
338 contained in an anonymous namespace: this reflects the fact that passes are self
339 contained units that do not need external interfaces (although they can have
340 them) to be useful.</p>
342 </div>
344 <!-- ======================================================================= -->
347 </div>
351 <p>Now that you have a brand new shiny shared object file, we can use the
352 <tt>opt</tt> command to run an LLVM program through your pass. Because you
353 registered your pass with the <tt>RegisterPass</tt> template, you will be able to
354 use the <tt>opt</tt> tool to access it, once loaded.</p>
356 <p>To test it, follow the example at the end of the <a
358 LLVM. We can now run the bitcode file (<tt>hello.bc</tt>) for the program
359 through our transformation like this (or course, any bitcode file will
360 work):</p>
363 $ opt -load ../../../Debug/lib/Hello.so -hello < hello.bc > /dev/null
364 Hello: __main
365 Hello: puts
366 Hello: main
367 </pre></div>
369 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
370 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
372 pass</a>). Because the hello pass does not modify the program in any
373 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
374 <tt>/dev/null</tt>).</p>
376 <p>To see what happened to the other string you registered, try running
377 <tt>opt</tt> with the <tt>--help</tt> option:</p>
380 $ opt -load ../../../Debug/lib/Hello.so --help
381 OVERVIEW: llvm .bc -> .bc modular optimizer
383 USAGE: opt [options] <input bitcode>
385 OPTIONS:
386 Optimizations available:
387 ...
388 -funcresolve - Resolve Functions
389 -gcse - Global Common Subexpression Elimination
390 -globaldce - Dead Global Elimination
391 <b>-hello - Hello World Pass</b>
392 -indvars - Canonicalize Induction Variables
393 -inline - Function Integration/Inlining
394 -instcombine - Combine redundant instructions
395 ...
396 </pre></div>
398 <p>The pass name get added as the information string for your pass, giving some
399 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
400 would go ahead and make it do the cool transformations you want. Once you get
401 it all working and tested, it may become useful to find out how fast your pass
403 line option (<tt>--time-passes</tt>) that allows you to get information about
404 the execution time of your pass along with the other passes you queue up. For
405 example:</p>
408 $ opt -load ../../../Debug/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
409 Hello: __main
410 Hello: puts
411 Hello: main
412 ===============================================================================
413 ... Pass execution timing report ...
414 ===============================================================================
415 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
417 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
418 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bitcode Writer
419 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
420 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
421 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
422 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
423 </pre></div>
425 <p>As you can see, our implementation above is pretty fast :). The additional
426 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
427 that the LLVM emitted by your pass is still valid and well formed LLVM, which
428 hasn't been broken somehow.</p>
430 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
431 about some more details of how they work and how to use them.</p>
433 </div>
435 <!-- *********************************************************************** -->
438 </div>
439 <!-- *********************************************************************** -->
443 <p>One of the first things that you should do when designing a new pass is to
444 decide what class you should subclass for your pass. The <a
447 did not discuss why or when this should occur. Here we talk about the classes
448 available, from the most general to the most specific.</p>
450 <p>When choosing a superclass for your Pass, you should choose the <b>most
451 specific</b> class possible, while still being able to meet the requirements
452 listed. This gives the LLVM Pass Infrastructure information necessary to
453 optimize how passes are run, so that the resultant compiler isn't unneccesarily
454 slow.</p>
456 </div>
458 <!-- ======================================================================= -->
461 </div>
467 class. This pass type is used for passes that do not have to be run, do not
468 change state, and never need to be updated. This is not a normal type of
469 transformation or analysis, but can provide information about the current
470 compiler configuration.</p>
472 <p>Although this pass class is very infrequently used, it is important for
473 providing information about the current target machine being compiled for, and
474 other static information that can affect the various transformations.</p>
476 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
479 </div>
481 <!-- ======================================================================= -->
484 </div>
490 class is the most general of all superclasses that you can use. Deriving from
491 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
492 refering to function bodies in no predictable order, or adding and removing
493 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
494 subclasses, no optimization can be done for their execution.</p>
496 <p>A module pass can use function level passes (e.g. dominators) using
497 the getAnalysis interface
498 <tt>getAnalysis<DominatorTree>(llvm::Function *)</tt> to provide the
499 function to retrieve analysis result for, if the function pass does not require
500 any module or immutable passes. Note that this can only be done for functions for which the
501 analysis ran, e.g. in the case of dominators you should only ask for the
502 DominatorTree for function definitions, not declarations.</p>
504 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
505 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
506 following signature:</p>
508 </div>
510 <!-- _______________________________________________________________________ -->
513 </div>
518 <b>virtual bool</b> runOnModule(Module &M) = 0;
519 </pre></div>
521 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
522 It should return true if the module was modified by the transformation and
523 false otherwise.</p>
525 </div>
527 <!-- ======================================================================= -->
530 </div>
536 is used by passes that need to traverse the program bottom-up on the call graph
537 (callees before callers). Deriving from CallGraphSCCPass provides some
538 mechanics for building and traversing the CallGraph, but also allows the system
539 to optimize execution of CallGraphSCCPass's. If your pass meets the
540 requirements outlined below, and doesn't meet the requirements of a <tt><a
543 <tt>CallGraphSCCPass</tt>.</p>
545 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
547 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
549 <ol>
551 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
552 the current SCC.</li>
554 <li>... <em>not allowed</em> to inspect any Function's other than those in the
555 current SCC and the direct callees of the SCC.</li>
557 <li>... <em>required</em> to preserve the current CallGraph object, updating it
558 to reflect any changes made to the program.</li>
560 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
561 though they may change the contents of an SCC.</li>
563 <li>... <em>allowed</em> to add or remove global variables from the current
564 Module.</li>
566 <li>... <em>allowed</em> to maintain state across invocations of
568 </ol>
570 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
571 because it has to handle SCCs with more than one node in it. All of the virtual
572 methods described below should return true if they modified the program, or
573 false if they didn't.</p>
575 </div>
577 <!-- _______________________________________________________________________ -->
580 method</a>
581 </div>
586 <b>virtual bool</b> doInitialization(CallGraph &CG);
587 </pre></div>
589 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
590 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
591 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
592 is designed to do simple initialization type of stuff that does not depend on
593 the SCCs being processed. The <tt>doInitialization</tt> method call is not
594 scheduled to overlap with any other pass executions (thus it should be very
595 fast).</p>
597 </div>
599 <!-- _______________________________________________________________________ -->
602 </div>
607 <b>virtual bool</b> runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;
608 </pre></div>
610 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
611 should return true if the module was modified by the transformation, false
612 otherwise.</p>
614 </div>
616 <!-- _______________________________________________________________________ -->
619 &)</tt> method</a>
620 </div>
625 <b>virtual bool</b> doFinalization(CallGraph &CG);
626 </pre></div>
628 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
629 called when the pass framework has finished calling <a
631 program being compiled.</p>
633 </div>
635 <!-- ======================================================================= -->
638 </div>
642 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
644 subclasses do have a predictable, local behavior that can be expected by the
645 system. All <tt>FunctionPass</tt> execute on each function in the program
646 independent of all of the other functions in the program.
647 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
648 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
650 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
652 <ol>
653 <li>Modify a Function other than the one currently being processed.</li>
654 <li>Add or remove Function's from the current Module.</li>
655 <li>Add or remove global variables from the current Module.</li>
656 <li>Maintain state across invocations of
658 </ol>
660 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
662 may overload three virtual methods to do their work. All of these methods
663 should return true if they modified the program, or false if they didn't.</p>
665 </div>
667 <!-- _______________________________________________________________________ -->
670 method</a>
671 </div>
676 <b>virtual bool</b> doInitialization(Module &M);
677 </pre></div>
679 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
680 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
681 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
682 is designed to do simple initialization type of stuff that does not depend on
683 the functions being processed. The <tt>doInitialization</tt> method call is not
684 scheduled to overlap with any other pass executions (thus it should be very
685 fast).</p>
687 <p>A good example of how this method should be used is the <a
689 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
690 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
691 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
692 free functions that it needs, adding prototypes to the module if necessary.</p>
694 </div>
696 <!-- _______________________________________________________________________ -->
699 </div>
704 <b>virtual bool</b> runOnFunction(Function &F) = 0;
705 </pre></div><p>
707 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
708 the transformation or analysis work of your pass. As usual, a true value should
709 be returned if the function is modified.</p>
711 </div>
713 <!-- _______________________________________________________________________ -->
716 &)</tt> method</a>
717 </div>
722 <b>virtual bool</b> doFinalization(Module &M);
723 </pre></div>
725 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
726 called when the pass framework has finished calling <a
728 program being compiled.</p>
730 </div>
732 <!-- ======================================================================= -->
735 </div>
739 <p> All <tt>LoopPass</tt> execute on each loop in the function independent of
740 all of the other loops in the function. <tt>LoopPass</tt> processes loops in
741 loop nest order such that outer most loop is processed last. </p>
743 <p> <tt>LoopPass</tt> subclasses are allowed to update loop nest using
744 <tt>LPPassManager</tt> interface. Implementing a loop pass is usually
745 straightforward. <tt>Looppass</tt>'s may overload three virtual methods to
746 do their work. All these methods should return true if they modified the
747 program, or false if they didn't. </p>
748 </div>
750 <!-- _______________________________________________________________________ -->
753 LPPassManager &)</tt>
754 method</a>
755 </div>
760 <b>virtual bool</b> doInitialization(Loop *, LPPassManager &LPM);
761 </pre></div>
763 <p>The <tt>doInitialization</tt> method is designed to do simple initialization
764 type of stuff that does not depend on the functions being processed. The
765 <tt>doInitialization</tt> method call is not scheduled to overlap with any
766 other pass executions (thus it should be very fast). LPPassManager
767 interface should be used to access Function or Module level analysis
768 information.</p>
770 </div>
773 <!-- _______________________________________________________________________ -->
776 </div>
781 <b>virtual bool</b> runOnLoop(Loop *, LPPassManager &LPM) = 0;
782 </pre></div><p>
784 <p>The <tt>runOnLoop</tt> method must be implemented by your subclass to do
785 the transformation or analysis work of your pass. As usual, a true value should
786 be returned if the function is modified. <tt>LPPassManager</tt> interface
787 should be used to update loop nest.</p>
789 </div>
791 <!-- _______________________________________________________________________ -->
794 </div>
799 <b>virtual bool</b> doFinalization();
800 </pre></div>
802 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
803 called when the pass framework has finished calling <a
805 program being compiled. </p>
807 </div>
811 <!-- ======================================================================= -->
814 </div>
818 <p><tt>BasicBlockPass</tt>'s are just like <a
820 their scope of inspection and modification to a single basic block at a time.
821 As such, they are <b>not</b> allowed to do any of the following:</p>
823 <ol>
824 <li>Modify or inspect any basic blocks outside of the current one</li>
825 <li>Maintain state across invocations of
827 <li>Modify the control flow graph (by altering terminator instructions)</li>
828 <li>Any of the things forbidden for
830 </ol>
833 optimizations. They may override the same <a
836 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
838 </div>
840 <!-- _______________________________________________________________________ -->
843 &)</tt> method</a>
844 </div>
849 <b>virtual bool</b> doInitialization(Function &F);
850 </pre></div>
852 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
853 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
854 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
855 to do simple initialization that does not depend on the
856 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
857 scheduled to overlap with any other pass executions (thus it should be very
858 fast).</p>
860 </div>
862 <!-- _______________________________________________________________________ -->
865 </div>
870 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
871 </pre></div>
873 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
874 function is not allowed to inspect or modify basic blocks other than the
875 parameter, and are not allowed to modify the CFG. A true value must be returned
876 if the basic block is modified.</p>
878 </div>
880 <!-- _______________________________________________________________________ -->
883 method</a>
884 </div>
889 <b>virtual bool</b> doFinalization(Function &F);
890 </pre></div>
892 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
893 called when the pass framework has finished calling <a
895 program being compiled. This can be used to perform per-function
896 finalization.</p>
898 </div>
900 <!-- ======================================================================= -->
903 </div>
907 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
908 executes on the machine-dependent representation of each LLVM function in the
909 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
910 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
911 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
912 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
914 <ol>
915 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
916 <li>Modify a MachineFunction other than the one currently being processed.</li>
917 <li>Add or remove MachineFunctions from the current Module.</li>
918 <li>Add or remove global variables from the current Module.</li>
919 <li>Maintain state across invocations of <a
921 data)</li>
922 </ol>
924 </div>
926 <!-- _______________________________________________________________________ -->
929 &MF)</tt> method</a>
930 </div>
935 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
936 </pre></div>
938 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
939 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
940 work of your <tt>MachineFunctionPass</tt>.</p>
942 <p>The <tt>runOnMachineFunction</tt> method is called on every
943 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
944 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
945 representation of the function. If you want to get at the LLVM <tt>Function</tt>
946 for the <tt>MachineFunction</tt> you're working on, use
947 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
948 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
949 <tt>MachineFunctionPass</tt>.</p>
951 </div>
953 <!-- *********************************************************************** -->
956 </div>
957 <!-- *********************************************************************** -->
962 pass registration works, and discussed some of the reasons that it is used and
963 what it does. Here we discuss how and why passes are registered.</p>
965 <p>As we saw above, passes are registered with the <b><tt>RegisterPass</tt></b>
966 template, which requires you to pass at least two
967 parameters. The first parameter is the name of the pass that is to be used on
968 the command line to specify that the pass should be added to a program (for
969 example, with <tt>opt</tt> or <tt>bugpoint</tt>). The second argument is the
970 name of the pass, which is to be used for the <tt>--help</tt> output of
971 programs, as
972 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
974 <p>If you want your pass to be easily dumpable, you should
975 implement the virtual <tt>print</tt> method:</p>
977 </div>
979 <!-- _______________________________________________________________________ -->
982 </div>
987 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
988 </pre></div>
991 a human readable version of the analysis results. This is useful for debugging
992 an analysis itself, as well as for other people to figure out how an analysis
993 works. Use the <tt>opt -analyze</tt> argument to invoke this method.</p>
995 <p>The <tt>llvm::OStream</tt> parameter specifies the stream to write the results on,
996 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
997 program that has been analyzed. Note however that this pointer may be null in
998 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
999 debugger), so it should only be used to enhance debug output, it should not be
1000 depended on.</p>
1002 </div>
1004 <!-- *********************************************************************** -->
1007 </div>
1008 <!-- *********************************************************************** -->
1012 <p>One of the main responsibilities of the <tt>PassManager</tt> is to make sure
1013 that passes interact with each other correctly. Because <tt>PassManager</tt>
1015 know how the passes interact with each other and what dependencies exist between
1016 the various passes. To track this, each pass can declare the set of passes that
1017 are required to be executed before the current pass, and the passes which are
1018 invalidated by the current pass.</p>
1020 <p>Typically this functionality is used to require that analysis results are
1021 computed before your pass is run. Running arbitrary transformation passes can
1022 invalidate the computed analysis results, which is what the invalidation set
1023 specifies. If a pass does not implement the <tt><a
1025 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
1027 </div>
1029 <!-- _______________________________________________________________________ -->
1032 </div>
1037 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
1038 </pre></div>
1040 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
1041 invalidated sets may be specified for your transformation. The implementation
1042 should fill in the <tt><a
1044 object with information about which passes are required and not invalidated. To
1045 do this, a pass may call any of the following methods on the AnalysisUsage
1046 object:</p>
1047 </div>
1049 <!-- _______________________________________________________________________ -->
1051 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
1052 </div>
1055 <p>
1056 If your pass requires a previous pass to be executed (an analysis for example),
1057 it can use one of these methods to arrange for it to be run before your pass.
1058 LLVM has many different types of analyses and passes that can be required,
1059 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
1060 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
1061 be no critical edges in the CFG when your pass has been run.
1062 </p>
1064 <p>
1065 Some analyses chain to other analyses to do their job. For example, an <a
1068 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
1069 used instead of the <tt>addRequired</tt> method. This informs the PassManager
1070 that the transitively required pass should be alive as long as the requiring
1071 pass is.
1072 </p>
1073 </div>
1075 <!-- _______________________________________________________________________ -->
1078 </div>
1081 <p>
1082 One of the jobs of the PassManager is to optimize how and when analyses are run.
1083 In particular, it attempts to avoid recomputing data unless it needs to. For
1084 this reason, passes are allowed to declare that they preserve (i.e., they don't
1085 invalidate) an existing analysis if it's available. For example, a simple
1086 constant folding pass would not modify the CFG, so it can't possibly affect the
1087 results of dominator analysis. By default, all passes are assumed to invalidate
1088 all others.
1089 </p>
1091 <p>
1092 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
1093 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
1094 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
1095 not modify the LLVM program at all (which is true for analyses), and the
1096 <tt>setPreservesCFG</tt> method can be used by transformations that change
1097 instructions in the program but do not modify the CFG or terminator instructions
1098 (note that this property is implicitly set for <a
1100 </p>
1102 <p>
1103 <tt>addPreserved</tt> is particularly useful for transformations like
1104 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
1105 and dominator related analyses if they exist, so it can preserve them, despite
1106 the fact that it hacks on the CFG.
1107 </p>
1108 </div>
1110 <!-- _______________________________________________________________________ -->
1113 </div>
1118 <i>// This is an example implementation from an analysis, which does not modify
1119 // the program at all, yet has a prerequisite.</i>
1120 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1121 AU.setPreservesAll();
1122 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1123 }
1124 </pre></div>
1126 <p>and:</p>
1129 <i>// This example modifies the program, but does not modify the CFG</i>
1130 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1131 AU.setPreservesCFG();
1132 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1133 }
1134 </pre></div>
1136 </div>
1138 <!-- _______________________________________________________________________ -->
1141 <tt>getAnalysisIfAvailable<></tt> methods</a>
1142 </div>
1146 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1147 your class, providing you with access to the passes that you declared that you
1149 method. It takes a single template argument that specifies which pass class you
1150 want, and returns a reference to that pass. For example:</p>
1153 bool LICM::runOnFunction(Function &F) {
1154 LoopInfo &LI = getAnalysis<LoopInfo>();
1155 ...
1156 }
1157 </pre></div>
1159 <p>This method call returns a reference to the pass desired. You may get a
1160 runtime assertion failure if you attempt to get an analysis that you did not
1161 declare as required in your <a
1163 method can be called by your <tt>run*</tt> method implementation, or by any
1164 other local method invoked by your <tt>run*</tt> method.
1166 A module level pass can use function level analysis info using this interface.
1167 For example:</p>
1170 bool ModuleLevelPass::runOnModule(Module &M) {
1171 ...
1172 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
1173 ...
1174 }
1175 </pre></div>
1177 <p>In above example, runOnFunction for DominatorTree is called by pass manager
1178 before returning a reference to the desired pass.</p>
1180 <p>
1181 If your pass is capable of updating analyses if they exist (e.g.,
1182 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1183 <tt>getAnalysisIfAvailable</tt> method, which returns a pointer to the analysis
1184 if it is active. For example:</p>
1187 ...
1188 if (DominatorSet *DS = getAnalysisIfAvailable<DominatorSet>()) {
1189 <i>// A DominatorSet is active. This code will update it.</i>
1190 }
1191 ...
1192 </pre></div>
1194 </div>
1196 <!-- *********************************************************************** -->
1199 </div>
1200 <!-- *********************************************************************** -->
1204 <p>Now that we understand the basics of how passes are defined, how they are
1205 used, and how they are required from other passes, it's time to get a little bit
1206 fancier. All of the pass relationships that we have seen so far are very
1207 simple: one pass depends on one other specific pass to be run before it can run.
1208 For many applications, this is great, for others, more flexibility is
1209 required.</p>
1211 <p>In particular, some analyses are defined such that there is a single simple
1212 interface to the analysis results, but multiple ways of calculating them.
1213 Consider alias analysis for example. The most trivial alias analysis returns
1215 flow-sensitive, context-sensitive interprocedural analysis that can take a
1216 significant amount of time to execute (and obviously, there is a lot of room
1217 between these two extremes for other implementations). To cleanly support
1218 situations like this, the LLVM Pass Infrastructure supports the notion of
1219 Analysis Groups.</p>
1221 </div>
1223 <!-- _______________________________________________________________________ -->
1226 </div>
1230 <p>An Analysis Group is a single simple interface that may be implemented by
1231 multiple different passes. Analysis Groups can be given human readable names
1232 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1233 class. An analysis group may have one or more implementations, one of which is
1236 <p>Analysis groups are used by client passes just like other passes are: the
1237 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1239 scans the available passes to see if any implementations of the analysis group
1240 are available. If none is available, the default implementation is created for
1242 between passes</a> still apply.</p>
1245 passes, all analysis group implementations must be registered, and must use the
1247 implementation pool. Also, a default implementation of the interface
1248 <b>must</b> be registered with <A
1251 <p>As a concrete example of an Analysis Group in action, consider the <a
1253 analysis group. The default implementation of the alias analysis interface (the
1254 <tt><a
1256 pass) just does a few simple checks that don't require significant analysis to
1257 compute (such as: two different globals can never alias each other, etc).
1258 Passes that use the <tt><a
1260 interface (for example the <tt><a
1262 not care which implementation of alias analysis is actually provided, they just
1263 use the designated interface.</p>
1265 <p>From the user's perspective, commands work just like normal. Issuing the
1266 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1267 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1268 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1269 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1270 hypothetical example) instead.</p>
1272 </div>
1274 <!-- _______________________________________________________________________ -->
1277 </div>
1281 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1282 group itself as well as add pass implementations to the analysis group. First,
1283 an analysis should be registered, with a human readable name provided for it.
1284 Unlike registration of passes, there is no command line argument to be specified
1288 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1289 </pre></div>
1291 <p>Once the analysis is registered, passes can declare that they are valid
1292 implementations of the interface by using the following code:</p>
1295 <b>namespace</b> {
1296 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1297 RegisterPass<FancyAA>
1300 //<i> Declare that we implement the AliasAnalysis interface</i>
1301 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> C(B);
1302 }
1303 </pre></div>
1305 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1308 analysis group. Every implementation of an analysis group should join using
1309 this template. A single pass may join multiple different analysis groups with
1310 no problem.</p>
1313 <b>namespace</b> {
1314 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1315 RegisterPass<<a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1318 //<i> Declare that we implement the AliasAnalysis interface</i>
1319 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <b>true</b>> E(D);
1320 }
1321 </pre></div>
1323 <p>Here we show how the default implementation is specified (using the extra
1324 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1325 one default implementation available at all times for an Analysis Group to be
1326 used. Only default implementation can derive from <tt>ImmutablePass</tt>.
1327 Here we declare that the
1328 <tt><a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1329 pass is the default implementation for the interface.</p>
1331 </div>
1333 <!-- *********************************************************************** -->
1336 </div>
1337 <!-- *********************************************************************** -->
1340 <p>The <a
1342 class is designed to be an easy way to expose various success
1343 metrics from passes. These statistics are printed at the end of a
1344 run, when the -stats command line option is enabled on the command
1345 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1347 </div>
1350 <!-- *********************************************************************** -->
1353 </div>
1354 <!-- *********************************************************************** -->
1358 <p>The <a
1360 <a
1363 are set up correctly, and then schedules passes to run efficiently. All of the
1364 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1365 passes.</p>
1367 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1368 time of a series of passes:</p>
1370 <ol>
1371 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1372 recomputing analysis results as much as possible. This means keeping track of
1373 which analyses are available already, which analyses get invalidated, and which
1374 analyses are needed to be run for a pass. An important part of work is that the
1375 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1377 results as soon as they are no longer needed.</li>
1379 <li><b>Pipeline the execution of passes on the program</b> - The
1380 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1381 of a series of passes by pipelining the passes together. This means that, given
1384 the first function, then all of the <a
1386 etc... until the entire program has been run through the passes.
1388 <p>This improves the cache behavior of the compiler, because it is only touching
1389 the LLVM program representation for a single function at a time, instead of
1390 traversing the entire program. It reduces the memory consumption of compiler,
1391 because, for example, only one <a
1393 needs to be calculated at a time. This also makes it possible to implement
1394 some <a
1397 </ol>
1399 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1400 much information it has about the behaviors of the passes it is scheduling. For
1403 Not implementing when it should be implemented will have the effect of not
1404 allowing any analysis results to live across the execution of your pass.</p>
1406 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1407 options that is useful for debugging pass execution, seeing how things work, and
1408 diagnosing when you should be preserving more analyses than you currently are
1409 (To get information about all of the variants of the <tt>--debug-pass</tt>
1410 option, just type '<tt>opt --help-hidden</tt>').</p>
1412 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1414 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1417 $ opt -load ../../../Debug/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1418 Module Pass Manager
1419 Function Pass Manager
1420 Dominator Set Construction
1421 Immediate Dominators Construction
1422 Global Common Subexpression Elimination
1423 -- Immediate Dominators Construction
1424 -- Global Common Subexpression Elimination
1425 Natural Loop Construction
1426 Loop Invariant Code Motion
1427 -- Natural Loop Construction
1428 -- Loop Invariant Code Motion
1429 Module Verifier
1430 -- Dominator Set Construction
1431 -- Module Verifier
1432 Bitcode Writer
1433 --Bitcode Writer
1434 </pre></div>
1436 <p>This output shows us when passes are constructed and when the analysis
1437 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1438 GCSE uses dominator and immediate dominator information to do its job. The LICM
1439 pass uses natural loop information, which uses dominator sets, but not immediate
1440 dominators. Because immediate dominators are no longer useful after the GCSE
1441 pass, it is immediately destroyed. The dominator sets are then reused to
1442 compute natural loop information, which is then used by the LICM pass.</p>
1444 <p>After the LICM pass, the module verifier runs (which is automatically added
1445 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1446 resultant LLVM code is well formed. After it finishes, the dominator set
1447 information is destroyed, after being computed once, and shared by three
1448 passes.</p>
1451 World</a> pass in between the two passes:</p>
1454 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1455 Module Pass Manager
1456 Function Pass Manager
1457 Dominator Set Construction
1458 Immediate Dominators Construction
1459 Global Common Subexpression Elimination
1460 <b>-- Dominator Set Construction</b>
1461 -- Immediate Dominators Construction
1462 -- Global Common Subexpression Elimination
1463 <b> Hello World Pass
1464 -- Hello World Pass
1465 Dominator Set Construction</b>
1466 Natural Loop Construction
1467 Loop Invariant Code Motion
1468 -- Natural Loop Construction
1469 -- Loop Invariant Code Motion
1470 Module Verifier
1471 -- Dominator Set Construction
1472 -- Module Verifier
1473 Bitcode Writer
1474 --Bitcode Writer
1475 Hello: __main
1476 Hello: puts
1477 Hello: main
1478 </pre></div>
1481 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1482 we need to add the following <a
1486 <i>// We don't modify the program, so we preserve all analyses</i>
1487 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1488 AU.setPreservesAll();
1489 }
1490 </pre></div>
1492 <p>Now when we run our pass, we get this output:</p>
1495 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1496 Pass Arguments: -gcse -hello -licm
1497 Module Pass Manager
1498 Function Pass Manager
1499 Dominator Set Construction
1500 Immediate Dominators Construction
1501 Global Common Subexpression Elimination
1502 -- Immediate Dominators Construction
1503 -- Global Common Subexpression Elimination
1504 Hello World Pass
1505 -- Hello World Pass
1506 Natural Loop Construction
1507 Loop Invariant Code Motion
1508 -- Loop Invariant Code Motion
1509 -- Natural Loop Construction
1510 Module Verifier
1511 -- Dominator Set Construction
1512 -- Module Verifier
1513 Bitcode Writer
1514 --Bitcode Writer
1515 Hello: __main
1516 Hello: puts
1517 Hello: main
1518 </pre></div>
1520 <p>Which shows that we don't accidentally invalidate dominator information
1521 anymore, and therefore do not have to compute it twice.</p>
1523 </div>
1525 <!-- _______________________________________________________________________ -->
1528 </div>
1533 <b>virtual void</b> releaseMemory();
1534 </pre></div>
1536 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1537 results, and how long to keep them around for. Because the lifetime of the pass
1538 object itself is effectively the entire duration of the compilation process, we
1539 need some way to free analysis results when they are no longer useful. The
1540 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1542 <p>If you are writing an analysis or any other pass that retains a significant
1545 <tt>releaseMemory</tt> to, well, release the memory allocated to maintain this
1546 internal state. This method is called after the <tt>run*</tt> method for the
1547 class, before the next call of <tt>run*</tt> in your pass.</p>
1549 </div>
1551 <!-- *********************************************************************** -->
1554 </div>
1555 <!-- *********************************************************************** -->
1559 <p><i>Size matters</i> when constructing production quality tools using llvm,
1560 both for the purposes of distribution, and for regulating the resident code size
1561 when running on the target system. Therefore, it becomes desirable to
1562 selectively use some passes, while omitting others and maintain the flexibility
1563 to change configurations later on. You want to be able to do all this, and,
1564 provide feedback to the user. This is where pass registration comes into
1565 play.</p>
1567 <p>The fundamental mechanisms for pass registration are the
1568 <tt>MachinePassRegistry</tt> class and subclasses of
1569 <tt>MachinePassRegistryNode</tt>.</p>
1571 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1572 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1573 communicates additions and deletions to the command line interface.</p>
1575 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1576 information provided about a particular pass. This information includes the
1577 command line name, the command help string and the address of the function used
1578 to create an instance of the pass. A global static constructor of one of these
1579 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1580 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1581 in the tool will be registered at start up. A dynamically loaded pass will
1582 register on load and unregister at unload.</p>
1584 </div>
1586 <!-- _______________________________________________________________________ -->
1589 </div>
1593 <p>There are predefined registries to track instruction scheduling
1594 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1595 machine passes. Here we will describe how to <i>register</i> a register
1596 allocator machine pass.</p>
1598 <p>Implement your register allocator machine pass. In your register allocator
1599 .cpp file add the following include;</p>
1603 </pre></div>
1605 <p>Also in your register allocator .cpp file, define a creator function in the
1606 form; </p>
1609 FunctionPass *createMyRegisterAllocator() {
1610 return new MyRegisterAllocator();
1611 }
1612 </pre></div>
1614 <p>Note that the signature of this function should match the type of
1615 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1621 createMyRegisterAllocator);
1622 </pre></div>
1624 <p>Note the two spaces prior to the help string produces a tidy result on the
1625 --help query.</p>
1628 $ llc --help
1629 ...
1630 -regalloc - Register allocator to use: (default = linearscan)
1631 =linearscan - linear scan register allocator
1632 =local - local register allocator
1633 =simple - simple register allocator
1634 =myregalloc - my register allocator help string
1635 ...
1636 </pre></div>
1638 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1639 an option. Registering instruction schedulers is similar except use the
1640 <tt>RegisterScheduler</tt> class. Note that the
1641 <tt>RegisterScheduler::FunctionPassCtor</tt> is significantly different from
1642 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1644 <p>To force the load/linking of your register allocator into the llc/lli tools,
1646 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1648 </div>
1651 <!-- _______________________________________________________________________ -->
1654 </div>
1658 <p>The easiest way to get started is to clone one of the existing registries; we
1659 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1660 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1662 <p>Then you need to declare the registry. Example: if your pass registry is
1663 <tt>RegisterMyPasses</tt> then define;</p>
1666 MachinePassRegistry RegisterMyPasses::Registry;
1667 </pre></div>
1669 <p>And finally, declare the command line option for your passes. Example:</p>
1672 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1673 RegisterPassParser<RegisterMyPasses> >
1675 cl::init(&createDefaultMyPass),
1677 </pre></div>
1680 creator.</p>
1682 </div>
1684 <!-- *********************************************************************** -->
1687 </div>
1688 <!-- *********************************************************************** -->
1692 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1693 should be. First of all, you can't set a breakpoint in a shared object that has
1694 not been loaded yet, and second of all there are problems with inlined functions
1695 in shared objects. Here are some suggestions to debugging your pass with
1696 GDB.</p>
1698 <p>For sake of discussion, I'm going to assume that you are debugging a
1699 transformation invoked by <tt>opt</tt>, although nothing described here depends
1700 on that.</p>
1702 </div>
1704 <!-- _______________________________________________________________________ -->
1707 </div>
1711 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1714 $ <b>gdb opt</b>
1715 GNU gdb 5.0
1716 Copyright 2000 Free Software Foundation, Inc.
1717 GDB is free software, covered by the GNU General Public License, and you are
1718 welcome to change it and/or distribute copies of it under certain conditions.
1722 (gdb)
1723 </pre></div>
1725 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1726 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1727 (the shared object isn't loaded until runtime), we must execute the process, and
1728 have it stop before it invokes our pass, but after it has loaded the shared
1729 object. The most foolproof way of doing this is to set a breakpoint in
1730 <tt>PassManager::run</tt> and then run the process with the arguments you
1731 want:</p>
1734 (gdb) <b>break llvm::PassManager::run</b>
1735 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1736 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]</b>
1737 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
1738 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1739 70 bool PassManager::run(Module &M) { return PM->run(M); }
1740 (gdb)
1741 </pre></div>
1743 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1744 now free to set breakpoints in your pass so that you can trace through execution
1745 or do other standard debugging stuff.</p>
1747 </div>
1749 <!-- _______________________________________________________________________ -->
1752 </div>
1756 <p>Once you have the basics down, there are a couple of problems that GDB has,
1757 some with solutions, some without.</p>
1759 <ul>
1760 <li>Inline functions have bogus stack information. In general, GDB does a
1761 pretty good job getting stack traces and stepping through inline functions.
1762 When a pass is dynamically loaded however, it somehow completely loses this
1763 capability. The only solution I know of is to de-inline a function (move it
1764 from the body of a class to a .cpp file).</li>
1766 <li>Restarting the program breaks breakpoints. After following the information
1767 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1768 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1769 and you start getting errors about breakpoints being unsettable. The only way I
1771 already set in your pass, run the program, and re-set the breakpoints once
1772 execution stops in <tt>PassManager::run</tt>.</li>
1774 </ul>
1776 <p>Hopefully these tips will help with common case debugging situations. If
1777 you'd like to contribute some tips of your own, just contact <a
1780 </div>
1782 <!-- *********************************************************************** -->
1785 </div>
1786 <!-- *********************************************************************** -->
1790 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1791 some nifty stuff, there are things we'd like to add in the future. Here is
1792 where we are going:</p>
1794 </div>
1796 <!-- _______________________________________________________________________ -->
1799 </div>
1803 <p>Multiple CPU machines are becoming more common and compilation can never be
1804 fast enough: obviously we should allow for a multithreaded compiler. Because of
1805 the semantics defined for passes above (specifically they cannot maintain state
1806 across invocations of their <tt>run*</tt> methods), a nice clean way to
1807 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1808 to create multiple instances of each pass object, and allow the separate
1809 instances to be hacking on different parts of the program at the same time.</p>
1811 <p>This implementation would prevent each of the passes from having to implement
1812 multithreaded constructs, requiring only the LLVM core to have locking in a few
1813 places (for global resources). Although this is a simple extension, we simply
1814 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1815 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1817 </div>
1819 <!-- *********************************************************************** -->
1820 <hr>
1821 <address>
1829 Last modified: $Date$
1830 </address>
1832 </body>
1833 </html>