| blob | a1d55649016c7e35145867c6034e910b84bc3726 |
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>
61 </ul></li>
63 <ul>
67 method</a></li>
70 </ul></li>
72 class</a>
73 <ul>
76 </ul></li>
77 </ul>
79 <ul>
81 </ul></li>
83 <ul>
85 method</a></li>
86 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
91 </ul></li>
93 <ul>
96 </ul></li>
99 <ul>
101 </ul></li>
103 <ul>
106 </ul></li>
108 <ul>
111 </ul></li>
113 <ul>
115 </ul></li>
116 </ol>
121 </div>
123 <!-- *********************************************************************** -->
126 </div>
127 <!-- *********************************************************************** -->
131 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
132 passes are where most of the interesting parts of the compiler exist. Passes
133 perform the transformations and optimizations that make up the compiler, they
134 build the analysis results that are used by these transformations, and they are,
135 above all, a structuring technique for compiler code.</p>
139 class, which implement functionality by overriding virtual methods inherited
147 more information about what your pass does, and how it can be combined with
148 other passes. One of the main features of the LLVM Pass Framework is that it
149 schedules passes to run in an efficient way based on the constraints that your
150 pass meets (which are indicated by which class they derive from).</p>
152 <p>We start by showing you how to construct a pass, everything from setting up
153 the code, to compiling, loading, and executing it. After the basics are down,
154 more advanced features are discussed.</p>
156 </div>
158 <!-- *********************************************************************** -->
161 </div>
162 <!-- *********************************************************************** -->
167 is designed to simply print out the name of non-external functions that exist in
168 the program being compiled. It does not modify the program at all, it just
169 inspects it. The source code and files for this pass are available in the LLVM
172 </div>
174 <!-- ======================================================================= -->
177 </div>
181 <p>First, you need to create a new directory somewhere in the LLVM source
182 base. For this example, we'll assume that you made
184 (Makefile) that will compile the source code for the new pass. To do this,
186 <hr/>
189 # Makefile for hello pass
191 # Path to top level of LLVM hierarchy
192 LEVEL = ../../..
194 # Name of the library to build
195 LIBRARYNAME = Hello
197 # Make the shared library become a loadable module so the tools can
198 # dlopen/dlsym on the resulting library.
199 LOADABLE_MODULE = 1
201 # Include the makefile implementation stuff
202 include $(LEVEL)/Makefile.common
203 </pre></div>
206 directory are to be compiled and linked together into a
209 If your operating system uses a suffix other than .so (such as windows or
210 Mac OS/X), the appropriate extension will be used.</p>
212 <p>Now that we have the build scripts set up, we just need to write the code for
213 the pass itself.</p>
215 </div>
217 <!-- ======================================================================= -->
220 </div>
224 <p>Now that we have a way to compile our new pass, we just have to write it.
225 Start out with:</p>
230 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
231 </pre></div>
235 we are operating on <tt><a
237 and we will be doing some printing.</p>
242 </pre></div>
243 <p>... which is required because the functions from the include files
244 live in the llvm namespace.
245 </p>
251 </pre></div>
253 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
254 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
255 things declared inside of the anonymous namespace only visible to the current
256 file. If you're not familiar with them, consult a decent C++ book for more
257 information.</p>
263 </pre></div><p>
267 The different builtin pass subclasses are described in detail <a
270 time.</p>
273 static char ID;
274 Hello() : FunctionPass(&ID) {}
275 </pre></div><p>
277 <p> This declares pass identifier used by LLVM to identify pass. This allows LLVM to
278 avoid using expensive C++ runtime information.</p>
284 }
286 </pre></div>
289 which overloads an abstract virtual method inherited from <a
291 to do our thing, so we just print out our message with the name of each
292 function.</p>
295 char Hello::ID = 0;
296 </pre></div>
298 <p> We initialize pass ID here. LLVM uses ID's address to identify pass so
299 initialization value is not important.</p>
303 false /* Only looks at CFG */,
304 false /* Analysis Pass */);
306 </pre></div>
309 giving it a command line
311 Last two arguments describe its behavior.
312 If a pass walks CFG without modifying it then third argument is set to true.
313 If a pass is an analysis pass, for example dominator tree pass, then true
314 is supplied as fourth argument. </p>
321 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
328 static char ID;
329 Hello() : FunctionPass(&ID) {}
334 }
335 };
337 char Hello::ID = 0;
339 }
341 </pre></div>
344 command in the local directory and you should get a new
345 "<tt>Debug+Asserts/lib/Hello.so</tt> file. Note that everything in this file is
346 contained in an anonymous namespace: this reflects the fact that passes are self
347 contained units that do not need external interfaces (although they can have
348 them) to be useful.</p>
350 </div>
352 <!-- ======================================================================= -->
355 </div>
359 <p>Now that you have a brand new shiny shared object file, we can use the
360 <tt>opt</tt> command to run an LLVM program through your pass. Because you
361 registered your pass with the <tt>INITIALIZE_PASS</tt> macro, you will be able to
362 use the <tt>opt</tt> tool to access it, once loaded.</p>
364 <p>To test it, follow the example at the end of the <a
366 LLVM. We can now run the bitcode file (<tt>hello.bc</tt>) for the program
367 through our transformation like this (or course, any bitcode file will
368 work):</p>
371 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello < hello.bc > /dev/null
372 Hello: __main
373 Hello: puts
374 Hello: main
375 </pre></div>
377 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
378 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
380 pass</a>). Because the hello pass does not modify the program in any
381 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
382 <tt>/dev/null</tt>).</p>
384 <p>To see what happened to the other string you registered, try running
385 <tt>opt</tt> with the <tt>-help</tt> option:</p>
388 $ opt -load ../../../Debug+Asserts/lib/Hello.so -help
389 OVERVIEW: llvm .bc -> .bc modular optimizer
391 USAGE: opt [options] <input bitcode>
393 OPTIONS:
394 Optimizations available:
395 ...
396 -funcresolve - Resolve Functions
397 -gcse - Global Common Subexpression Elimination
398 -globaldce - Dead Global Elimination
399 <b>-hello - Hello World Pass</b>
400 -indvars - Canonicalize Induction Variables
401 -inline - Function Integration/Inlining
402 -instcombine - Combine redundant instructions
403 ...
404 </pre></div>
406 <p>The pass name get added as the information string for your pass, giving some
407 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
408 would go ahead and make it do the cool transformations you want. Once you get
409 it all working and tested, it may become useful to find out how fast your pass
411 line option (<tt>--time-passes</tt>) that allows you to get information about
412 the execution time of your pass along with the other passes you queue up. For
413 example:</p>
416 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
417 Hello: __main
418 Hello: puts
419 Hello: main
420 ===============================================================================
421 ... Pass execution timing report ...
422 ===============================================================================
423 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
425 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
426 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bitcode Writer
427 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
428 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
429 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
430 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
431 </pre></div>
433 <p>As you can see, our implementation above is pretty fast :). The additional
434 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
435 that the LLVM emitted by your pass is still valid and well formed LLVM, which
436 hasn't been broken somehow.</p>
438 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
439 about some more details of how they work and how to use them.</p>
441 </div>
443 <!-- *********************************************************************** -->
446 </div>
447 <!-- *********************************************************************** -->
451 <p>One of the first things that you should do when designing a new pass is to
452 decide what class you should subclass for your pass. The <a
455 did not discuss why or when this should occur. Here we talk about the classes
456 available, from the most general to the most specific.</p>
458 <p>When choosing a superclass for your Pass, you should choose the <b>most
459 specific</b> class possible, while still being able to meet the requirements
460 listed. This gives the LLVM Pass Infrastructure information necessary to
461 optimize how passes are run, so that the resultant compiler isn't unnecessarily
462 slow.</p>
464 </div>
466 <!-- ======================================================================= -->
469 </div>
475 class. This pass type is used for passes that do not have to be run, do not
476 change state, and never need to be updated. This is not a normal type of
477 transformation or analysis, but can provide information about the current
478 compiler configuration.</p>
480 <p>Although this pass class is very infrequently used, it is important for
481 providing information about the current target machine being compiled for, and
482 other static information that can affect the various transformations.</p>
484 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
487 </div>
489 <!-- ======================================================================= -->
492 </div>
498 class is the most general of all superclasses that you can use. Deriving from
499 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
500 referring to function bodies in no predictable order, or adding and removing
501 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
502 subclasses, no optimization can be done for their execution.</p>
504 <p>A module pass can use function level passes (e.g. dominators) using
505 the getAnalysis interface
506 <tt>getAnalysis<DominatorTree>(llvm::Function *)</tt> to provide the
507 function to retrieve analysis result for, if the function pass does not require
508 any module or immutable passes. Note that this can only be done for functions for which the
509 analysis ran, e.g. in the case of dominators you should only ask for the
510 DominatorTree for function definitions, not declarations.</p>
512 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
513 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
514 following signature:</p>
516 </div>
518 <!-- _______________________________________________________________________ -->
521 </div>
526 <b>virtual bool</b> runOnModule(Module &M) = 0;
527 </pre></div>
529 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
530 It should return true if the module was modified by the transformation and
531 false otherwise.</p>
533 </div>
535 <!-- ======================================================================= -->
538 </div>
544 is used by passes that need to traverse the program bottom-up on the call graph
545 (callees before callers). Deriving from CallGraphSCCPass provides some
546 mechanics for building and traversing the CallGraph, but also allows the system
547 to optimize execution of CallGraphSCCPass's. If your pass meets the
548 requirements outlined below, and doesn't meet the requirements of a <tt><a
551 <tt>CallGraphSCCPass</tt>.</p>
553 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
555 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
557 <ol>
559 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
560 the current SCC.</li>
562 <li>... <em>not allowed</em> to inspect any Function's other than those in the
563 current SCC and the direct callees of the SCC.</li>
565 <li>... <em>required</em> to preserve the current CallGraph object, updating it
566 to reflect any changes made to the program.</li>
568 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
569 though they may change the contents of an SCC.</li>
571 <li>... <em>allowed</em> to add or remove global variables from the current
572 Module.</li>
574 <li>... <em>allowed</em> to maintain state across invocations of
576 </ol>
578 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
579 because it has to handle SCCs with more than one node in it. All of the virtual
580 methods described below should return true if they modified the program, or
581 false if they didn't.</p>
583 </div>
585 <!-- _______________________________________________________________________ -->
588 method</a>
589 </div>
594 <b>virtual bool</b> doInitialization(CallGraph &CG);
595 </pre></div>
597 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
598 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
599 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
600 is designed to do simple initialization type of stuff that does not depend on
601 the SCCs being processed. The <tt>doInitialization</tt> method call is not
602 scheduled to overlap with any other pass executions (thus it should be very
603 fast).</p>
605 </div>
607 <!-- _______________________________________________________________________ -->
610 </div>
615 <b>virtual bool</b> runOnSCC(CallGraphSCC &SCC) = 0;
616 </pre></div>
618 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
619 should return true if the module was modified by the transformation, false
620 otherwise.</p>
622 </div>
624 <!-- _______________________________________________________________________ -->
627 &)</tt> method</a>
628 </div>
633 <b>virtual bool</b> doFinalization(CallGraph &CG);
634 </pre></div>
636 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
637 called when the pass framework has finished calling <a
639 program being compiled.</p>
641 </div>
643 <!-- ======================================================================= -->
646 </div>
650 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
652 subclasses do have a predictable, local behavior that can be expected by the
653 system. All <tt>FunctionPass</tt> execute on each function in the program
654 independent of all of the other functions in the program.
655 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
656 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
658 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
660 <ol>
661 <li>Modify a Function other than the one currently being processed.</li>
662 <li>Add or remove Function's from the current Module.</li>
663 <li>Add or remove global variables from the current Module.</li>
664 <li>Maintain state across invocations of
666 </ol>
668 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
670 may overload three virtual methods to do their work. All of these methods
671 should return true if they modified the program, or false if they didn't.</p>
673 </div>
675 <!-- _______________________________________________________________________ -->
678 method</a>
679 </div>
684 <b>virtual bool</b> doInitialization(Module &M);
685 </pre></div>
687 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
688 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
689 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
690 is designed to do simple initialization type of stuff that does not depend on
691 the functions being processed. The <tt>doInitialization</tt> method call is not
692 scheduled to overlap with any other pass executions (thus it should be very
693 fast).</p>
695 <p>A good example of how this method should be used is the <a
697 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
698 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
699 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
700 free functions that it needs, adding prototypes to the module if necessary.</p>
702 </div>
704 <!-- _______________________________________________________________________ -->
707 </div>
712 <b>virtual bool</b> runOnFunction(Function &F) = 0;
713 </pre></div><p>
715 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
716 the transformation or analysis work of your pass. As usual, a true value should
717 be returned if the function is modified.</p>
719 </div>
721 <!-- _______________________________________________________________________ -->
724 &)</tt> method</a>
725 </div>
730 <b>virtual bool</b> doFinalization(Module &M);
731 </pre></div>
733 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
734 called when the pass framework has finished calling <a
736 program being compiled.</p>
738 </div>
740 <!-- ======================================================================= -->
743 </div>
747 <p> All <tt>LoopPass</tt> execute on each loop in the function independent of
748 all of the other loops in the function. <tt>LoopPass</tt> processes loops in
749 loop nest order such that outer most loop is processed last. </p>
751 <p> <tt>LoopPass</tt> subclasses are allowed to update loop nest using
752 <tt>LPPassManager</tt> interface. Implementing a loop pass is usually
753 straightforward. <tt>Looppass</tt>'s may overload three virtual methods to
754 do their work. All these methods should return true if they modified the
755 program, or false if they didn't. </p>
756 </div>
758 <!-- _______________________________________________________________________ -->
761 LPPassManager &)</tt>
762 method</a>
763 </div>
768 <b>virtual bool</b> doInitialization(Loop *, LPPassManager &LPM);
769 </pre></div>
771 <p>The <tt>doInitialization</tt> method is designed to do simple initialization
772 type of stuff that does not depend on the functions being processed. The
773 <tt>doInitialization</tt> method call is not scheduled to overlap with any
774 other pass executions (thus it should be very fast). LPPassManager
775 interface should be used to access Function or Module level analysis
776 information.</p>
778 </div>
781 <!-- _______________________________________________________________________ -->
784 </div>
789 <b>virtual bool</b> runOnLoop(Loop *, LPPassManager &LPM) = 0;
790 </pre></div><p>
792 <p>The <tt>runOnLoop</tt> method must be implemented by your subclass to do
793 the transformation or analysis work of your pass. As usual, a true value should
794 be returned if the function is modified. <tt>LPPassManager</tt> interface
795 should be used to update loop nest.</p>
797 </div>
799 <!-- _______________________________________________________________________ -->
802 </div>
807 <b>virtual bool</b> doFinalization();
808 </pre></div>
810 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
811 called when the pass framework has finished calling <a
813 program being compiled. </p>
815 </div>
817 <!-- ======================================================================= -->
820 </div>
825 but executes on each single entry single exit region in the function.
826 <tt>RegionPass</tt> processes regions in nested order such that the outer most
827 region is processed last. </p>
829 <p> <tt>RegionPass</tt> subclasses are allowed to update the region tree by using
830 the <tt>RGPassManager</tt> interface. You may overload three virtual methods of
831 <tt>RegionPass</tt> to implementing your own region pass is usually. All these
832 methods should return true if they modified the program, or false if they didn not.
833 </p>
834 </div>
836 <!-- _______________________________________________________________________ -->
839 RGPassManager &)</tt>
840 method</a>
841 </div>
846 <b>virtual bool</b> doInitialization(Region *, RGPassManager &RGM);
847 </pre></div>
849 <p>The <tt>doInitialization</tt> method is designed to do simple initialization
850 type of stuff that does not depend on the functions being processed. The
851 <tt>doInitialization</tt> method call is not scheduled to overlap with any
852 other pass executions (thus it should be very fast). RPPassManager
853 interface should be used to access Function or Module level analysis
854 information.</p>
856 </div>
859 <!-- _______________________________________________________________________ -->
862 </div>
867 <b>virtual bool</b> runOnRegion(Region *, RGPassManager &RGM) = 0;
868 </pre></div><p>
870 <p>The <tt>runOnRegion</tt> method must be implemented by your subclass to do
871 the transformation or analysis work of your pass. As usual, a true value should
872 be returned if the region is modified. <tt>RGPassManager</tt> interface
873 should be used to update region tree.</p>
875 </div>
877 <!-- _______________________________________________________________________ -->
880 </div>
885 <b>virtual bool</b> doFinalization();
886 </pre></div>
888 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
889 called when the pass framework has finished calling <a
891 program being compiled. </p>
893 </div>
897 <!-- ======================================================================= -->
900 </div>
904 <p><tt>BasicBlockPass</tt>'s are just like <a
906 their scope of inspection and modification to a single basic block at a time.
907 As such, they are <b>not</b> allowed to do any of the following:</p>
909 <ol>
910 <li>Modify or inspect any basic blocks outside of the current one</li>
911 <li>Maintain state across invocations of
913 <li>Modify the control flow graph (by altering terminator instructions)</li>
914 <li>Any of the things forbidden for
916 </ol>
919 optimizations. They may override the same <a
922 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
924 </div>
926 <!-- _______________________________________________________________________ -->
929 &)</tt> method</a>
930 </div>
935 <b>virtual bool</b> doInitialization(Function &F);
936 </pre></div>
938 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
939 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
940 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
941 to do simple initialization that does not depend on the
942 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
943 scheduled to overlap with any other pass executions (thus it should be very
944 fast).</p>
946 </div>
948 <!-- _______________________________________________________________________ -->
951 </div>
956 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
957 </pre></div>
959 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
960 function is not allowed to inspect or modify basic blocks other than the
961 parameter, and are not allowed to modify the CFG. A true value must be returned
962 if the basic block is modified.</p>
964 </div>
966 <!-- _______________________________________________________________________ -->
969 method</a>
970 </div>
975 <b>virtual bool</b> doFinalization(Function &F);
976 </pre></div>
978 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
979 called when the pass framework has finished calling <a
981 program being compiled. This can be used to perform per-function
982 finalization.</p>
984 </div>
986 <!-- ======================================================================= -->
989 </div>
993 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
994 executes on the machine-dependent representation of each LLVM function in the
995 program.</p>
997 <p>Code generator passes are registered and initialized specially by
998 <tt>TargetMachine::addPassesToEmitFile</tt> and similar routines, so they
999 cannot generally be run from the <tt>opt</tt> or <tt>bugpoint</tt>
1000 commands.</p>
1002 <p>A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
1003 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
1004 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
1005 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
1007 <ol>
1008 <li>Modify or create any LLVM IR Instructions, BasicBlocks, Arguments,
1009 Functions, GlobalVariables, GlobalAliases, or Modules.</li>
1010 <li>Modify a MachineFunction other than the one currently being processed.</li>
1011 <li>Maintain state across invocations of <a
1013 data)</li>
1014 </ol>
1016 </div>
1018 <!-- _______________________________________________________________________ -->
1021 &MF)</tt> method</a>
1022 </div>
1027 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
1028 </pre></div>
1030 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
1031 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
1032 work of your <tt>MachineFunctionPass</tt>.</p>
1034 <p>The <tt>runOnMachineFunction</tt> method is called on every
1035 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
1036 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
1037 representation of the function. If you want to get at the LLVM <tt>Function</tt>
1038 for the <tt>MachineFunction</tt> you're working on, use
1039 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
1040 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
1041 <tt>MachineFunctionPass</tt>.</p>
1043 </div>
1045 <!-- *********************************************************************** -->
1048 </div>
1049 <!-- *********************************************************************** -->
1054 pass registration works, and discussed some of the reasons that it is used and
1055 what it does. Here we discuss how and why passes are registered.</p>
1057 <p>As we saw above, passes are registered with the <b><tt>INITIALIZE_PASS</tt></b>
1058 macro. The first parameter is the name of the pass that is to be used on
1059 the command line to specify that the pass should be added to a program (for
1060 example, with <tt>opt</tt> or <tt>bugpoint</tt>). The second argument is the
1061 name of the pass, which is to be used for the <tt>-help</tt> output of
1062 programs, as
1063 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
1065 <p>If you want your pass to be easily dumpable, you should
1066 implement the virtual <tt>print</tt> method:</p>
1068 </div>
1070 <!-- _______________________________________________________________________ -->
1073 </div>
1078 <b>virtual void</b> print(std::ostream &O, <b>const</b> Module *M) <b>const</b>;
1079 </pre></div>
1082 a human readable version of the analysis results. This is useful for debugging
1083 an analysis itself, as well as for other people to figure out how an analysis
1084 works. Use the <tt>opt -analyze</tt> argument to invoke this method.</p>
1086 <p>The <tt>llvm::OStream</tt> parameter specifies the stream to write the results on,
1087 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
1088 program that has been analyzed. Note however that this pointer may be null in
1089 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
1090 debugger), so it should only be used to enhance debug output, it should not be
1091 depended on.</p>
1093 </div>
1095 <!-- *********************************************************************** -->
1098 </div>
1099 <!-- *********************************************************************** -->
1103 <p>One of the main responsibilities of the <tt>PassManager</tt> is to make sure
1104 that passes interact with each other correctly. Because <tt>PassManager</tt>
1106 know how the passes interact with each other and what dependencies exist between
1107 the various passes. To track this, each pass can declare the set of passes that
1108 are required to be executed before the current pass, and the passes which are
1109 invalidated by the current pass.</p>
1111 <p>Typically this functionality is used to require that analysis results are
1112 computed before your pass is run. Running arbitrary transformation passes can
1113 invalidate the computed analysis results, which is what the invalidation set
1114 specifies. If a pass does not implement the <tt><a
1116 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
1118 </div>
1120 <!-- _______________________________________________________________________ -->
1123 </div>
1128 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
1129 </pre></div>
1131 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
1132 invalidated sets may be specified for your transformation. The implementation
1133 should fill in the <tt><a
1135 object with information about which passes are required and not invalidated. To
1136 do this, a pass may call any of the following methods on the AnalysisUsage
1137 object:</p>
1138 </div>
1140 <!-- _______________________________________________________________________ -->
1142 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
1143 </div>
1146 <p>
1147 If your pass requires a previous pass to be executed (an analysis for example),
1148 it can use one of these methods to arrange for it to be run before your pass.
1149 LLVM has many different types of analyses and passes that can be required,
1150 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
1151 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
1152 be no critical edges in the CFG when your pass has been run.
1153 </p>
1155 <p>
1156 Some analyses chain to other analyses to do their job. For example, an <a
1159 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
1160 used instead of the <tt>addRequired</tt> method. This informs the PassManager
1161 that the transitively required pass should be alive as long as the requiring
1162 pass is.
1163 </p>
1164 </div>
1166 <!-- _______________________________________________________________________ -->
1169 </div>
1172 <p>
1173 One of the jobs of the PassManager is to optimize how and when analyses are run.
1174 In particular, it attempts to avoid recomputing data unless it needs to. For
1175 this reason, passes are allowed to declare that they preserve (i.e., they don't
1176 invalidate) an existing analysis if it's available. For example, a simple
1177 constant folding pass would not modify the CFG, so it can't possibly affect the
1178 results of dominator analysis. By default, all passes are assumed to invalidate
1179 all others.
1180 </p>
1182 <p>
1183 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
1184 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
1185 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
1186 not modify the LLVM program at all (which is true for analyses), and the
1187 <tt>setPreservesCFG</tt> method can be used by transformations that change
1188 instructions in the program but do not modify the CFG or terminator instructions
1189 (note that this property is implicitly set for <a
1191 </p>
1193 <p>
1194 <tt>addPreserved</tt> is particularly useful for transformations like
1195 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
1196 and dominator related analyses if they exist, so it can preserve them, despite
1197 the fact that it hacks on the CFG.
1198 </p>
1199 </div>
1201 <!-- _______________________________________________________________________ -->
1204 </div>
1209 <i>// This is an example implementation from an analysis, which does not modify
1210 // the program at all, yet has a prerequisite.</i>
1211 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1212 AU.setPreservesAll();
1213 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1214 }
1215 </pre></div>
1217 <p>and:</p>
1220 <i>// This example modifies the program, but does not modify the CFG</i>
1221 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1222 AU.setPreservesCFG();
1223 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1224 }
1225 </pre></div>
1227 </div>
1229 <!-- _______________________________________________________________________ -->
1232 <tt>getAnalysisIfAvailable<></tt> methods</a>
1233 </div>
1237 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1238 your class, providing you with access to the passes that you declared that you
1240 method. It takes a single template argument that specifies which pass class you
1241 want, and returns a reference to that pass. For example:</p>
1244 bool LICM::runOnFunction(Function &F) {
1245 LoopInfo &LI = getAnalysis<LoopInfo>();
1246 ...
1247 }
1248 </pre></div>
1250 <p>This method call returns a reference to the pass desired. You may get a
1251 runtime assertion failure if you attempt to get an analysis that you did not
1252 declare as required in your <a
1254 method can be called by your <tt>run*</tt> method implementation, or by any
1255 other local method invoked by your <tt>run*</tt> method.
1257 A module level pass can use function level analysis info using this interface.
1258 For example:</p>
1261 bool ModuleLevelPass::runOnModule(Module &M) {
1262 ...
1263 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
1264 ...
1265 }
1266 </pre></div>
1268 <p>In above example, runOnFunction for DominatorTree is called by pass manager
1269 before returning a reference to the desired pass.</p>
1271 <p>
1272 If your pass is capable of updating analyses if they exist (e.g.,
1273 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1274 <tt>getAnalysisIfAvailable</tt> method, which returns a pointer to the analysis
1275 if it is active. For example:</p>
1278 ...
1279 if (DominatorSet *DS = getAnalysisIfAvailable<DominatorSet>()) {
1280 <i>// A DominatorSet is active. This code will update it.</i>
1281 }
1282 ...
1283 </pre></div>
1285 </div>
1287 <!-- *********************************************************************** -->
1290 </div>
1291 <!-- *********************************************************************** -->
1295 <p>Now that we understand the basics of how passes are defined, how they are
1296 used, and how they are required from other passes, it's time to get a little bit
1297 fancier. All of the pass relationships that we have seen so far are very
1298 simple: one pass depends on one other specific pass to be run before it can run.
1299 For many applications, this is great, for others, more flexibility is
1300 required.</p>
1302 <p>In particular, some analyses are defined such that there is a single simple
1303 interface to the analysis results, but multiple ways of calculating them.
1304 Consider alias analysis for example. The most trivial alias analysis returns
1306 flow-sensitive, context-sensitive interprocedural analysis that can take a
1307 significant amount of time to execute (and obviously, there is a lot of room
1308 between these two extremes for other implementations). To cleanly support
1309 situations like this, the LLVM Pass Infrastructure supports the notion of
1310 Analysis Groups.</p>
1312 </div>
1314 <!-- _______________________________________________________________________ -->
1317 </div>
1321 <p>An Analysis Group is a single simple interface that may be implemented by
1322 multiple different passes. Analysis Groups can be given human readable names
1323 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1324 class. An analysis group may have one or more implementations, one of which is
1327 <p>Analysis groups are used by client passes just like other passes are: the
1328 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1330 scans the available passes to see if any implementations of the analysis group
1331 are available. If none is available, the default implementation is created for
1333 between passes</a> still apply.</p>
1336 passes, all analysis group implementations must be registered, and must use the
1338 implementation pool. Also, a default implementation of the interface
1339 <b>must</b> be registered with <A
1342 <p>As a concrete example of an Analysis Group in action, consider the <a
1344 analysis group. The default implementation of the alias analysis interface (the
1345 <tt><a
1347 pass) just does a few simple checks that don't require significant analysis to
1348 compute (such as: two different globals can never alias each other, etc).
1349 Passes that use the <tt><a
1351 interface (for example the <tt><a
1353 not care which implementation of alias analysis is actually provided, they just
1354 use the designated interface.</p>
1356 <p>From the user's perspective, commands work just like normal. Issuing the
1357 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1358 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1359 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1360 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1361 hypothetical example) instead.</p>
1363 </div>
1365 <!-- _______________________________________________________________________ -->
1368 </div>
1372 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1373 group itself, while the <tt>INITIALIZE_AG_PASS</tt> is used to add pass
1374 implementations to the analysis group. First,
1375 an analysis group should be registered, with a human readable name
1376 provided for it.
1377 Unlike registration of passes, there is no command line argument to be specified
1381 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1382 </pre></div>
1384 <p>Once the analysis is registered, passes can declare that they are valid
1385 implementations of the interface by using the following code:</p>
1388 <b>namespace</b> {
1389 //<i> Declare that we implement the AliasAnalysis interface</i>
1390 INITIALIZE_AG_PASS(FancyAA, <a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, "<i>somefancyaa</i>",
1392 false, // <i>Is CFG Only?</i>
1393 true, // <i>Is Analysis?</i>
1394 false, // <i>Is default Analysis Group implementation?</i>
1395 );
1396 }
1397 </pre></div>
1399 <p>This just shows a class <tt>FancyAA</tt> that
1400 uses the <tt>INITIALIZE_AG_PASS</tt> macro both to register and
1401 to "join" the <tt><a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1402 analysis group. Every implementation of an analysis group should join using
1403 this macro.</p>
1406 <b>namespace</b> {
1407 //<i> Declare that we implement the AliasAnalysis interface</i>
1408 INITIALIZE_AG_PASS(BasicAA, <a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, "<i>basicaa</i>",
1410 false, // <i>Is CFG Only?</i>
1411 true, // <i>Is Analysis?</i>
1412 true, // <i>Is default Analysis Group implementation?</i>
1413 );
1414 }
1415 </pre></div>
1417 <p>Here we show how the default implementation is specified (using the final
1418 argument to the <tt>INITIALIZE_AG_PASS</tt> template). There must be exactly
1419 one default implementation available at all times for an Analysis Group to be
1420 used. Only default implementation can derive from <tt>ImmutablePass</tt>.
1421 Here we declare that the
1422 <tt><a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1423 pass is the default implementation for the interface.</p>
1425 </div>
1427 <!-- *********************************************************************** -->
1430 </div>
1431 <!-- *********************************************************************** -->
1434 <p>The <a
1436 class is designed to be an easy way to expose various success
1437 metrics from passes. These statistics are printed at the end of a
1438 run, when the -stats command line option is enabled on the command
1439 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1441 </div>
1444 <!-- *********************************************************************** -->
1447 </div>
1448 <!-- *********************************************************************** -->
1452 <p>The <a
1454 <a
1457 are set up correctly, and then schedules passes to run efficiently. All of the
1458 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1459 passes.</p>
1461 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1462 time of a series of passes:</p>
1464 <ol>
1465 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1466 recomputing analysis results as much as possible. This means keeping track of
1467 which analyses are available already, which analyses get invalidated, and which
1468 analyses are needed to be run for a pass. An important part of work is that the
1469 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1471 results as soon as they are no longer needed.</li>
1473 <li><b>Pipeline the execution of passes on the program</b> - The
1474 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1475 of a series of passes by pipelining the passes together. This means that, given
1478 the first function, then all of the <a
1480 etc... until the entire program has been run through the passes.
1482 <p>This improves the cache behavior of the compiler, because it is only touching
1483 the LLVM program representation for a single function at a time, instead of
1484 traversing the entire program. It reduces the memory consumption of compiler,
1485 because, for example, only one <a
1487 needs to be calculated at a time. This also makes it possible to implement
1488 some <a
1491 </ol>
1493 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1494 much information it has about the behaviors of the passes it is scheduling. For
1497 Not implementing when it should be implemented will have the effect of not
1498 allowing any analysis results to live across the execution of your pass.</p>
1500 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1501 options that is useful for debugging pass execution, seeing how things work, and
1502 diagnosing when you should be preserving more analyses than you currently are
1503 (To get information about all of the variants of the <tt>--debug-pass</tt>
1504 option, just type '<tt>opt -help-hidden</tt>').</p>
1506 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1508 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1511 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1512 Module Pass Manager
1513 Function Pass Manager
1514 Dominator Set Construction
1515 Immediate Dominators Construction
1516 Global Common Subexpression Elimination
1517 -- Immediate Dominators Construction
1518 -- Global Common Subexpression Elimination
1519 Natural Loop Construction
1520 Loop Invariant Code Motion
1521 -- Natural Loop Construction
1522 -- Loop Invariant Code Motion
1523 Module Verifier
1524 -- Dominator Set Construction
1525 -- Module Verifier
1526 Bitcode Writer
1527 --Bitcode Writer
1528 </pre></div>
1530 <p>This output shows us when passes are constructed and when the analysis
1531 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1532 GCSE uses dominator and immediate dominator information to do its job. The LICM
1533 pass uses natural loop information, which uses dominator sets, but not immediate
1534 dominators. Because immediate dominators are no longer useful after the GCSE
1535 pass, it is immediately destroyed. The dominator sets are then reused to
1536 compute natural loop information, which is then used by the LICM pass.</p>
1538 <p>After the LICM pass, the module verifier runs (which is automatically added
1539 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1540 resultant LLVM code is well formed. After it finishes, the dominator set
1541 information is destroyed, after being computed once, and shared by three
1542 passes.</p>
1545 World</a> pass in between the two passes:</p>
1548 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1549 Module Pass Manager
1550 Function Pass Manager
1551 Dominator Set Construction
1552 Immediate Dominators Construction
1553 Global Common Subexpression Elimination
1554 <b>-- Dominator Set Construction</b>
1555 -- Immediate Dominators Construction
1556 -- Global Common Subexpression Elimination
1557 <b> Hello World Pass
1558 -- Hello World Pass
1559 Dominator Set Construction</b>
1560 Natural Loop Construction
1561 Loop Invariant Code Motion
1562 -- Natural Loop Construction
1563 -- Loop Invariant Code Motion
1564 Module Verifier
1565 -- Dominator Set Construction
1566 -- Module Verifier
1567 Bitcode Writer
1568 --Bitcode Writer
1569 Hello: __main
1570 Hello: puts
1571 Hello: main
1572 </pre></div>
1575 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1576 we need to add the following <a
1580 <i>// We don't modify the program, so we preserve all analyses</i>
1581 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1582 AU.setPreservesAll();
1583 }
1584 </pre></div>
1586 <p>Now when we run our pass, we get this output:</p>
1589 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1590 Pass Arguments: -gcse -hello -licm
1591 Module Pass Manager
1592 Function Pass Manager
1593 Dominator Set Construction
1594 Immediate Dominators Construction
1595 Global Common Subexpression Elimination
1596 -- Immediate Dominators Construction
1597 -- Global Common Subexpression Elimination
1598 Hello World Pass
1599 -- Hello World Pass
1600 Natural Loop Construction
1601 Loop Invariant Code Motion
1602 -- Loop Invariant Code Motion
1603 -- Natural Loop Construction
1604 Module Verifier
1605 -- Dominator Set Construction
1606 -- Module Verifier
1607 Bitcode Writer
1608 --Bitcode Writer
1609 Hello: __main
1610 Hello: puts
1611 Hello: main
1612 </pre></div>
1614 <p>Which shows that we don't accidentally invalidate dominator information
1615 anymore, and therefore do not have to compute it twice.</p>
1617 </div>
1619 <!-- _______________________________________________________________________ -->
1622 </div>
1627 <b>virtual void</b> releaseMemory();
1628 </pre></div>
1630 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1631 results, and how long to keep them around for. Because the lifetime of the pass
1632 object itself is effectively the entire duration of the compilation process, we
1633 need some way to free analysis results when they are no longer useful. The
1634 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1636 <p>If you are writing an analysis or any other pass that retains a significant
1639 <tt>releaseMemory</tt> to, well, release the memory allocated to maintain this
1640 internal state. This method is called after the <tt>run*</tt> method for the
1641 class, before the next call of <tt>run*</tt> in your pass.</p>
1643 </div>
1645 <!-- *********************************************************************** -->
1648 </div>
1649 <!-- *********************************************************************** -->
1653 <p><i>Size matters</i> when constructing production quality tools using llvm,
1654 both for the purposes of distribution, and for regulating the resident code size
1655 when running on the target system. Therefore, it becomes desirable to
1656 selectively use some passes, while omitting others and maintain the flexibility
1657 to change configurations later on. You want to be able to do all this, and,
1658 provide feedback to the user. This is where pass registration comes into
1659 play.</p>
1661 <p>The fundamental mechanisms for pass registration are the
1662 <tt>MachinePassRegistry</tt> class and subclasses of
1663 <tt>MachinePassRegistryNode</tt>.</p>
1665 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1666 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1667 communicates additions and deletions to the command line interface.</p>
1669 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1670 information provided about a particular pass. This information includes the
1671 command line name, the command help string and the address of the function used
1672 to create an instance of the pass. A global static constructor of one of these
1673 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1674 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1675 in the tool will be registered at start up. A dynamically loaded pass will
1676 register on load and unregister at unload.</p>
1678 </div>
1680 <!-- _______________________________________________________________________ -->
1683 </div>
1687 <p>There are predefined registries to track instruction scheduling
1688 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1689 machine passes. Here we will describe how to <i>register</i> a register
1690 allocator machine pass.</p>
1692 <p>Implement your register allocator machine pass. In your register allocator
1693 .cpp file add the following include;</p>
1697 </pre></div>
1699 <p>Also in your register allocator .cpp file, define a creator function in the
1700 form; </p>
1703 FunctionPass *createMyRegisterAllocator() {
1704 return new MyRegisterAllocator();
1705 }
1706 </pre></div>
1708 <p>Note that the signature of this function should match the type of
1709 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1715 createMyRegisterAllocator);
1716 </pre></div>
1718 <p>Note the two spaces prior to the help string produces a tidy result on the
1719 -help query.</p>
1722 $ llc -help
1723 ...
1724 -regalloc - Register allocator to use (default=linearscan)
1725 =linearscan - linear scan register allocator
1726 =local - local register allocator
1727 =simple - simple register allocator
1728 =myregalloc - my register allocator help string
1729 ...
1730 </pre></div>
1732 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1733 an option. Registering instruction schedulers is similar except use the
1734 <tt>RegisterScheduler</tt> class. Note that the
1735 <tt>RegisterScheduler::FunctionPassCtor</tt> is significantly different from
1736 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1738 <p>To force the load/linking of your register allocator into the llc/lli tools,
1740 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1742 </div>
1745 <!-- _______________________________________________________________________ -->
1748 </div>
1752 <p>The easiest way to get started is to clone one of the existing registries; we
1753 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1754 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1756 <p>Then you need to declare the registry. Example: if your pass registry is
1757 <tt>RegisterMyPasses</tt> then define;</p>
1760 MachinePassRegistry RegisterMyPasses::Registry;
1761 </pre></div>
1763 <p>And finally, declare the command line option for your passes. Example:</p>
1766 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1767 RegisterPassParser<RegisterMyPasses> >
1769 cl::init(&createDefaultMyPass),
1771 </pre></div>
1774 creator.</p>
1776 </div>
1778 <!-- *********************************************************************** -->
1781 </div>
1782 <!-- *********************************************************************** -->
1786 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1787 should be. First of all, you can't set a breakpoint in a shared object that has
1788 not been loaded yet, and second of all there are problems with inlined functions
1789 in shared objects. Here are some suggestions to debugging your pass with
1790 GDB.</p>
1792 <p>For sake of discussion, I'm going to assume that you are debugging a
1793 transformation invoked by <tt>opt</tt>, although nothing described here depends
1794 on that.</p>
1796 </div>
1798 <!-- _______________________________________________________________________ -->
1801 </div>
1805 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1808 $ <b>gdb opt</b>
1809 GNU gdb 5.0
1810 Copyright 2000 Free Software Foundation, Inc.
1811 GDB is free software, covered by the GNU General Public License, and you are
1812 welcome to change it and/or distribute copies of it under certain conditions.
1816 (gdb)
1817 </pre></div>
1819 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1820 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1821 (the shared object isn't loaded until runtime), we must execute the process, and
1822 have it stop before it invokes our pass, but after it has loaded the shared
1823 object. The most foolproof way of doing this is to set a breakpoint in
1824 <tt>PassManager::run</tt> and then run the process with the arguments you
1825 want:</p>
1828 (gdb) <b>break llvm::PassManager::run</b>
1829 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1830 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]</b>
1831 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1832 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1833 70 bool PassManager::run(Module &M) { return PM->run(M); }
1834 (gdb)
1835 </pre></div>
1837 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1838 now free to set breakpoints in your pass so that you can trace through execution
1839 or do other standard debugging stuff.</p>
1841 </div>
1843 <!-- _______________________________________________________________________ -->
1846 </div>
1850 <p>Once you have the basics down, there are a couple of problems that GDB has,
1851 some with solutions, some without.</p>
1853 <ul>
1854 <li>Inline functions have bogus stack information. In general, GDB does a
1855 pretty good job getting stack traces and stepping through inline functions.
1856 When a pass is dynamically loaded however, it somehow completely loses this
1857 capability. The only solution I know of is to de-inline a function (move it
1858 from the body of a class to a .cpp file).</li>
1860 <li>Restarting the program breaks breakpoints. After following the information
1861 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1862 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1863 and you start getting errors about breakpoints being unsettable. The only way I
1865 already set in your pass, run the program, and re-set the breakpoints once
1866 execution stops in <tt>PassManager::run</tt>.</li>
1868 </ul>
1870 <p>Hopefully these tips will help with common case debugging situations. If
1871 you'd like to contribute some tips of your own, just contact <a
1874 </div>
1876 <!-- *********************************************************************** -->
1879 </div>
1880 <!-- *********************************************************************** -->
1884 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1885 some nifty stuff, there are things we'd like to add in the future. Here is
1886 where we are going:</p>
1888 </div>
1890 <!-- _______________________________________________________________________ -->
1893 </div>
1897 <p>Multiple CPU machines are becoming more common and compilation can never be
1898 fast enough: obviously we should allow for a multithreaded compiler. Because of
1899 the semantics defined for passes above (specifically they cannot maintain state
1900 across invocations of their <tt>run*</tt> methods), a nice clean way to
1901 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1902 to create multiple instances of each pass object, and allow the separate
1903 instances to be hacking on different parts of the program at the same time.</p>
1905 <p>This implementation would prevent each of the passes from having to implement
1906 multithreaded constructs, requiring only the LLVM core to have locking in a few
1907 places (for global resources). Although this is a simple extension, we simply
1908 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1909 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1911 </div>
1913 <!-- *********************************************************************** -->
1914 <hr>
1915 <address>
1923 Last modified: $Date$
1924 </address>
1926 </body>
1927 </html>