| blob | 5799c99402c0c89ce1ccd6ed1490816fa96228fc |
1 ======================================================
2 Kaleidoscope: Conclusion and other useful LLVM tidbits
3 ======================================================
5 .. contents::
6 :local:
8 Tutorial Conclusion
9 ===================
11 Welcome to the final chapter of the "`Implementing a language with
12 LLVM <index.html>`_" tutorial. In the course of this tutorial, we have
13 grown our little Kaleidoscope language from being a useless toy, to
14 being a semi-interesting (but probably still useless) toy. :)
16 It is interesting to see how far we've come, and how little code it has
17 taken. We built the entire lexer, parser, AST, code generator, an
18 interactive run-loop (with a JIT!), and emitted debug information in
19 standalone executables - all in under 1000 lines of (non-comment/non-blank)
20 code.
22 Our little language supports a couple of interesting features: it
23 supports user defined binary and unary operators, it uses JIT
24 compilation for immediate evaluation, and it supports a few control flow
25 constructs with SSA construction.
27 Part of the idea of this tutorial was to show you how easy and fun it
28 can be to define, build, and play with languages. Building a compiler
29 need not be a scary or mystical process! Now that you've seen some of
30 the basics, I strongly encourage you to take the code and hack on it.
31 For example, try adding:
33 - **global variables** - While global variables have questional value
34 in modern software engineering, they are often useful when putting
35 together quick little hacks like the Kaleidoscope compiler itself.
36 Fortunately, our current setup makes it very easy to add global
37 variables: just have value lookup check to see if an unresolved
38 variable is in the global variable symbol table before rejecting it.
39 To create a new global variable, make an instance of the LLVM
40 ``GlobalVariable`` class.
41 - **typed variables** - Kaleidoscope currently only supports variables
42 of type double. This gives the language a very nice elegance, because
43 only supporting one type means that you never have to specify types.
44 Different languages have different ways of handling this. The easiest
45 way is to require the user to specify types for every variable
46 definition, and record the type of the variable in the symbol table
47 along with its Value\*.
48 - **arrays, structs, vectors, etc** - Once you add types, you can start
49 extending the type system in all sorts of interesting ways. Simple
50 arrays are very easy and are quite useful for many different
51 applications. Adding them is mostly an exercise in learning how the
52 LLVM `getelementptr <../LangRef.html#getelementptr-instruction>`_ instruction
53 works: it is so nifty/unconventional, it `has its own
54 FAQ <../GetElementPtr.html>`_!
55 - **standard runtime** - Our current language allows the user to access
56 arbitrary external functions, and we use it for things like "printd"
57 and "putchard". As you extend the language to add higher-level
58 constructs, often these constructs make the most sense if they are
59 lowered to calls into a language-supplied runtime. For example, if
60 you add hash tables to the language, it would probably make sense to
61 add the routines to a runtime, instead of inlining them all the way.
62 - **memory management** - Currently we can only access the stack in
63 Kaleidoscope. It would also be useful to be able to allocate heap
64 memory, either with calls to the standard libc malloc/free interface
65 or with a garbage collector. If you would like to use garbage
66 collection, note that LLVM fully supports `Accurate Garbage
67 Collection <../GarbageCollection.html>`_ including algorithms that
68 move objects and need to scan/update the stack.
69 - **exception handling support** - LLVM supports generation of `zero
70 cost exceptions <../ExceptionHandling.html>`_ which interoperate with
71 code compiled in other languages. You could also generate code by
72 implicitly making every function return an error value and checking
73 it. You could also make explicit use of setjmp/longjmp. There are
74 many different ways to go here.
75 - **object orientation, generics, database access, complex numbers,
76 geometric programming, ...** - Really, there is no end of crazy
77 features that you can add to the language.
78 - **unusual domains** - We've been talking about applying LLVM to a
79 domain that many people are interested in: building a compiler for a
80 specific language. However, there are many other domains that can use
81 compiler technology that are not typically considered. For example,
82 LLVM has been used to implement OpenGL graphics acceleration,
83 translate C++ code to ActionScript, and many other cute and clever
84 things. Maybe you will be the first to JIT compile a regular
85 expression interpreter into native code with LLVM?
87 Have fun - try doing something crazy and unusual. Building a language
88 like everyone else always has, is much less fun than trying something a
89 little crazy or off the wall and seeing how it turns out. If you get
90 stuck or want to talk about it, feel free to email the `llvm-dev mailing
91 list <http://lists.llvm.org/mailman/listinfo/llvm-dev>`_: it has lots
92 of people who are interested in languages and are often willing to help
93 out.
95 Before we end this tutorial, I want to talk about some "tips and tricks"
96 for generating LLVM IR. These are some of the more subtle things that
97 may not be obvious, but are very useful if you want to take advantage of
98 LLVM's capabilities.
100 Properties of the LLVM IR
101 =========================
103 We have a couple of common questions about code in the LLVM IR form -
104 let's just get these out of the way right now, shall we?
106 Target Independence
107 -------------------
109 Kaleidoscope is an example of a "portable language": any program written
110 in Kaleidoscope will work the same way on any target that it runs on.
111 Many other languages have this property, e.g. lisp, java, haskell,
112 javascript, python, etc (note that while these languages are portable,
113 not all their libraries are).
115 One nice aspect of LLVM is that it is often capable of preserving target
116 independence in the IR: you can take the LLVM IR for a
117 Kaleidoscope-compiled program and run it on any target that LLVM
118 supports, even emitting C code and compiling that on targets that LLVM
119 doesn't support natively. You can trivially tell that the Kaleidoscope
120 compiler generates target-independent code because it never queries for
121 any target-specific information when generating code.
123 The fact that LLVM provides a compact, target-independent,
124 representation for code gets a lot of people excited. Unfortunately,
125 these people are usually thinking about C or a language from the C
126 family when they are asking questions about language portability. I say
127 "unfortunately", because there is really no way to make (fully general)
128 C code portable, other than shipping the source code around (and of
129 course, C source code is not actually portable in general either - ever
130 port a really old application from 32- to 64-bits?).
132 The problem with C (again, in its full generality) is that it is heavily
133 laden with target specific assumptions. As one simple example, the
134 preprocessor often destructively removes target-independence from the
135 code when it processes the input text:
137 .. code-block:: c
139 #ifdef __i386__
140 int X = 1;
141 #else
142 int X = 42;
143 #endif
145 While it is possible to engineer more and more complex solutions to
146 problems like this, it cannot be solved in full generality in a way that
147 is better than shipping the actual source code.
149 That said, there are interesting subsets of C that can be made portable.
150 If you are willing to fix primitive types to a fixed size (say int =
151 32-bits, and long = 64-bits), don't care about ABI compatibility with
152 existing binaries, and are willing to give up some other minor features,
153 you can have portable code. This can make sense for specialized domains
154 such as an in-kernel language.
156 Safety Guarantees
157 -----------------
159 Many of the languages above are also "safe" languages: it is impossible
160 for a program written in Java to corrupt its address space and crash the
161 process (assuming the JVM has no bugs). Safety is an interesting
162 property that requires a combination of language design, runtime
163 support, and often operating system support.
165 It is certainly possible to implement a safe language in LLVM, but LLVM
166 IR does not itself guarantee safety. The LLVM IR allows unsafe pointer
167 casts, use after free bugs, buffer over-runs, and a variety of other
168 problems. Safety needs to be implemented as a layer on top of LLVM and,
169 conveniently, several groups have investigated this. Ask on the `llvm-dev
170 mailing list <http://lists.llvm.org/mailman/listinfo/llvm-dev>`_ if
171 you are interested in more details.
173 Language-Specific Optimizations
174 -------------------------------
176 One thing about LLVM that turns off many people is that it does not
177 solve all the world's problems in one system (sorry 'world hunger',
178 someone else will have to solve you some other day). One specific
179 complaint is that people perceive LLVM as being incapable of performing
180 high-level language-specific optimization: LLVM "loses too much
181 information".
183 Unfortunately, this is really not the place to give you a full and
184 unified version of "Chris Lattner's theory of compiler design". Instead,
185 I'll make a few observations:
187 First, you're right that LLVM does lose information. For example, as of
188 this writing, there is no way to distinguish in the LLVM IR whether an
189 SSA-value came from a C "int" or a C "long" on an ILP32 machine (other
190 than debug info). Both get compiled down to an 'i32' value and the
191 information about what it came from is lost. The more general issue
192 here, is that the LLVM type system uses "structural equivalence" instead
193 of "name equivalence". Another place this surprises people is if you
194 have two types in a high-level language that have the same structure
195 (e.g. two different structs that have a single int field): these types
196 will compile down into a single LLVM type and it will be impossible to
197 tell what it came from.
199 Second, while LLVM does lose information, LLVM is not a fixed target: we
200 continue to enhance and improve it in many different ways. In addition
201 to adding new features (LLVM did not always support exceptions or debug
202 info), we also extend the IR to capture important information for
203 optimization (e.g. whether an argument is sign or zero extended,
204 information about pointers aliasing, etc). Many of the enhancements are
205 user-driven: people want LLVM to include some specific feature, so they
206 go ahead and extend it.
208 Third, it is *possible and easy* to add language-specific optimizations,
209 and you have a number of choices in how to do it. As one trivial
210 example, it is easy to add language-specific optimization passes that
211 "know" things about code compiled for a language. In the case of the C
212 family, there is an optimization pass that "knows" about the standard C
213 library functions. If you call "exit(0)" in main(), it knows that it is
214 safe to optimize that into "return 0;" because C specifies what the
215 'exit' function does.
217 In addition to simple library knowledge, it is possible to embed a
218 variety of other language-specific information into the LLVM IR. If you
219 have a specific need and run into a wall, please bring the topic up on
220 the llvm-dev list. At the very worst, you can always treat LLVM as if it
221 were a "dumb code generator" and implement the high-level optimizations
222 you desire in your front-end, on the language-specific AST.
224 Tips and Tricks
225 ===============
227 There is a variety of useful tips and tricks that you come to know after
228 working on/with LLVM that aren't obvious at first glance. Instead of
229 letting everyone rediscover them, this section talks about some of these
230 issues.
232 Implementing portable offsetof/sizeof
233 -------------------------------------
235 One interesting thing that comes up, if you are trying to keep the code
236 generated by your compiler "target independent", is that you often need
237 to know the size of some LLVM type or the offset of some field in an
238 llvm structure. For example, you might need to pass the size of a type
239 into a function that allocates memory.
241 Unfortunately, this can vary widely across targets: for example the
242 width of a pointer is trivially target-specific. However, there is a
243 `clever way to use the getelementptr
244 instruction <http://nondot.org/sabre/LLVMNotes/SizeOf-OffsetOf-VariableSizedStructs.txt>`_
245 that allows you to compute this in a portable way.
247 Garbage Collected Stack Frames
248 ------------------------------
250 Some languages want to explicitly manage their stack frames, often so
251 that they are garbage collected or to allow easy implementation of
252 closures. There are often better ways to implement these features than
253 explicit stack frames, but `LLVM does support
254 them, <http://nondot.org/sabre/LLVMNotes/ExplicitlyManagedStackFrames.txt>`_
255 if you want. It requires your front-end to convert the code into
256 `Continuation Passing
257 Style <http://en.wikipedia.org/wiki/Continuation-passing_style>`_ and
258 the use of tail calls (which LLVM also supports).