Gitiles
Code Review Sign In
gerrit-public.fairphone.software / fp2-dev / platform / external / llvm / fe6575c2bf1126c9dee3cc73ad321a375634fc42 / . / docs / tutorial / LangImpl7.html
blob: ca0d2f06e034fb9880e4ce620cf4329e14aad9fe [file] [log] [blame]
Chris Lattner00c992d2007-11-03 08:55:29 +0000[diff] [blame]1<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
3
4<html>
5<head>
6 <title>Kaleidoscope: Extending the Language: Mutable Variables / SSA
7 construction</title>
8 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
9 <meta name="author" content="Chris Lattner">
10 <link rel="stylesheet" href="../llvm.css" type="text/css">
11</head>
12
13<body>
14
15<div class="doc_title">Kaleidoscope: Extending the Language: Mutable Variables</div>
16
17<div class="doc_author">
18 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
19</div>
20
21<!-- *********************************************************************** -->
22<div class="doc_section"><a name="intro">Part 7 Introduction</a></div>
23<!-- *********************************************************************** -->
24
25<div class="doc_text">
26
27<p>Welcome to Part 7 of the "<a href="index.html">Implementing a language with
28LLVM</a>" tutorial. In parts 1 through 6, we've built a very respectable,
29albeit simple, <a
30href="http://en.wikipedia.org/wiki/Functional_programming">functional
31programming language</a>. In our journey, we learned some parsing techniques,
32how to build and represent an AST, how to build LLVM IR, and how to optimize
33the resultant code and JIT compile it.</p>
34
35<p>While Kaleidoscope is interesting as a functional language, this makes it
36"too easy" to generate LLVM IR for it. In particular, a functional language
37makes it very easy to build LLVM IR directly in <a
38href="http://en.wikipedia.org/wiki/Static_single_assignment_form">SSA form</a>.
39Since LLVM requires that the input code be in SSA form, this is a very nice
40property and it is often unclear to newcomers how to generate code for an
41imperative language with mutable variables.</p>
42
43<p>The short (and happy) summary of this chapter is that there is no need for
44your front-end to build SSA form: LLVM provides highly tuned and well tested
45support for this, though the way it works is a bit unexpected for some.</p>
46
47</div>
48
49<!-- *********************************************************************** -->
50<div class="doc_section"><a name="why">Why is this a hard problem?</a></div>
51<!-- *********************************************************************** -->
52
53<div class="doc_text">
54
55<p>
56To understand why mutable variables cause complexities in SSA construction,
57consider this extremely simple C example:
58</p>
59
60<div class="doc_code">
61<pre>
62int G, H;
63int test(_Bool Condition) {
64 int X;
65 if (Condition)
66 X = G;
67 else
68 X = H;
69 return X;
70}
71</pre>
72</div>
73
74<p>In this case, we have the variable "X", whose value depends on the path
75executed in the program. Because there are two different possible values for X
76before the return instruction, a PHI node is inserted to merge the two values.
77The LLVM IR that we want for this example looks like this:</p>
78
79<div class="doc_code">
80<pre>
81@G = weak global i32 0 ; type of @G is i32*
82@H = weak global i32 0 ; type of @H is i32*
83
84define i32 @test(i1 %Condition) {
85entry:
86 br i1 %Condition, label %cond_true, label %cond_false
87
88cond_true:
89 %X.0 = load i32* @G
90 br label %cond_next
91
92cond_false:
93 %X.1 = load i32* @H
94 br label %cond_next
95
96cond_next:
97 %X.2 = phi i32 [ %X.1, %cond_false ], [ %X.0, %cond_true ]
98 ret i32 %X.2
99}
100</pre>
101</div>
102
103<p>In this example, the loads from the G and H global variables are explicit in
104the LLVM IR, and they live in the then/else branches of the if statement
105(cond_true/cond_false). In order to merge the incoming values, the X.2 phi node
106in the cond_next block selects the right value to use based on where control
107flow is coming from: if control flow comes from the cond_false block, X.2 gets
108the value of X.1. Alternatively, if control flow comes from cond_tree, it gets
109the value of X.0. The intent of this chapter is not to explain the details of
110SSA form. For more information, see one of the many <a
111href="http://en.wikipedia.org/wiki/Static_single_assignment_form">online
112references</a>.</p>
113
114<p>The question for this article is "who places phi nodes when lowering
115assignments to mutable variables?". The issue here is that LLVM
116<em>requires</em> that its IR be in SSA form: there is no "non-ssa" mode for it.
117However, SSA construction requires non-trivial algorithms and data structures,
118so it is inconvenient and wasteful for every front-end to have to reproduce this
119logic.</p>
120
121</div>
122
123<!-- *********************************************************************** -->
124<div class="doc_section"><a name="memory">Memory in LLVM</a></div>
125<!-- *********************************************************************** -->
126
127<div class="doc_text">
128
129<p>The 'trick' here is that while LLVM does require all register values to be
130in SSA form, it does not require (or permit) memory objects to be in SSA form.
131In the example above, note that the loads from G and H are direct accesses to
132G and H: they are not renamed or versioned. This differs from some other
133compiler systems, which does try to version memory objects. In LLVM, instead of
134encoding dataflow analysis of memory into the LLVM IR, it is handled with <a
135href="../WritingAnLLVMPass.html">Analysis Passes</a> which are computed on
136demand.</p>
137
138<p>
139With this in mind, the high-level idea is that we want to make a stack variable
140(which lives in memory, because it is on the stack) for each mutable object in
141a function. To take advantage of this trick, we need to talk about how LLVM
142represents stack variables.
143</p>
144
145<p>In LLVM, all memory accesses are explicit with load/store instructions, and
146it is carefully designed to not have (or need) an "address-of" operator. Notice
147how the type of the @G/@H global variables is actually "i32*" even though the
148variable is defined as "i32". What this means is that @G defines <em>space</em>
149for an i32 in the global data area, but its <em>name</em> actually refers to the
150address for that space. Stack variables work the same way, but instead of being
151declared with global variable definitions, they are declared with the
152<a href="../LangRef.html#i_alloca">LLVM alloca instruction</a>:</p>
153
154<div class="doc_code">
155<pre>
156define i32 @test(i1 %Condition) {
157entry:
158 %X = alloca i32 ; type of %X is i32*.
159 ...
160 %tmp = load i32* %X ; load the stack value %X from the stack.
161 %tmp2 = add i32 %tmp, 1 ; increment it
162 store i32 %tmp2, i32* %X ; store it back
163 ...
164</pre>
165</div>
166
167<p>This code shows an example of how you can declare and manipulate a stack
168variable in the LLVM IR. Stack memory allocated with the alloca instruction is
169fully general: you can pass the address of the stack slot to functions, you can
170store it in other variables, etc. In our example above, we could rewrite the
171example to use the alloca technique to avoid using a PHI node:</p>
172
173<div class="doc_code">
174<pre>
175@G = weak global i32 0 ; type of @G is i32*
176@H = weak global i32 0 ; type of @H is i32*
177
178define i32 @test(i1 %Condition) {
179entry:
180 %X = alloca i32 ; type of %X is i32*.
181 br i1 %Condition, label %cond_true, label %cond_false
182
183cond_true:
184 %X.0 = load i32* @G
185 store i32 %X.0, i32* %X ; Update X
186 br label %cond_next
187
188cond_false:
189 %X.1 = load i32* @H
190 store i32 %X.1, i32* %X ; Update X
191 br label %cond_next
192
193cond_next:
194 %X.2 = load i32* %X ; Read X
195 ret i32 %X.2
196}
197</pre>
198</div>
199
200<p>With this, we have discovered a way to handle arbitrary mutable variables
201without the need to create Phi nodes at all:</p>
202
203<ol>
204<li>Each mutable variable becomes a stack allocation.</li>
205<li>Each read of the variable becomes a load from the stack.</li>
206<li>Each update of the variable becomes a store to the stack.</li>
207<li>Taking the address of a variable just uses the stack address directly.</li>
208</ol>
209
210<p>While this solution has solved our immediate problem, it introduced another
211one: we have now apparently introduced a lot of stack traffic for very simple
212and common operations, a major performance problem. Fortunately for us, the
213LLVM optimizer has a highly-tuned optimization pass named "mem2reg" that handles
214this case, promoting allocas like this into SSA registers, inserting Phi nodes
215as appropriate. If you run this example through the pass, for example, you'll
216get:</p>
217
218<div class="doc_code">
219<pre>
220$ <b>llvm-as &lt; example.ll | opt -mem2reg | llvm-dis</b>
221@G = weak global i32 0
222@H = weak global i32 0
223
224define i32 @test(i1 %Condition) {
225entry:
226 br i1 %Condition, label %cond_true, label %cond_false
227
228cond_true:
229 %X.0 = load i32* @G
230 br label %cond_next
231
232cond_false:
233 %X.1 = load i32* @H
234 br label %cond_next
235
236cond_next:
237 %X.01 = phi i32 [ %X.1, %cond_false ], [ %X.0, %cond_true ]
238 ret i32 %X.01
239}
240</pre>
Chris Lattnere7198312007-11-03 22:22:30 +0000[diff] [blame]241</div>
Chris Lattner00c992d2007-11-03 08:55:29 +0000[diff] [blame]242
Chris Lattnere7198312007-11-03 22:22:30 +0000[diff] [blame]243<p>The mem2reg pass implements the standard "iterated dominator frontier"
244algorithm for constructing SSA form and has a number of optimizations that speed
245up very common degenerate cases. mem2reg really is the answer for dealing with
246mutable variables, and we highly recommend that you depend on it. Note that
247mem2reg only works on variables in certain circumstances:</p>
Chris Lattner00c992d2007-11-03 08:55:29 +0000[diff] [blame]248
Chris Lattnere7198312007-11-03 22:22:30 +0000[diff] [blame]249<ol>
250<li>mem2reg is alloca-driven: it looks for allocas and if it can handle them, it
251promotes them. It does not apply to global variables or heap allocations.</li>
Chris Lattner00c992d2007-11-03 08:55:29 +0000[diff] [blame]252
Chris Lattnere7198312007-11-03 22:22:30 +0000[diff] [blame]253<li>mem2reg only looks for alloca instructions in the entry block of the
254function. Being in the entry block guarantees that the alloca is only executed
255once, which makes analysis simpler.</li>
Chris Lattner00c992d2007-11-03 08:55:29 +0000[diff] [blame]256
Chris Lattnere7198312007-11-03 22:22:30 +0000[diff] [blame]257<li>mem2reg only promotes allocas whose uses are direct loads and stores. If
258the address of the stack object is passed to a function, or if any funny pointer
259arithmetic is involved, the alloca will not be promoted.</li>
260
261<li>mem2reg only works on allocas of scalar values, and only if the array size
262of the allocation is 1 (or missing in the .ll file). mem2reg is not capable of
263promoting structs or arrays to registers. Note that the "scalarrepl" pass is
264more powerful and can promote structs, "unions", and arrays in many cases.</li>
265
266</ol>
267
268<p>
269All of these properties are easy to satisfy for most imperative languages, and
270we'll illustrated this below with Kaleidoscope. The final question you may be
271asking is: should I bother with this nonsense for my front-end? Wouldn't it be
272better if I just did SSA construction directly, avoiding use of the mem2reg
273optimization pass? In short, we strongly recommend that use you this technique
274for building SSA form, unless there is an extremely good reason not to. Using
275this technique is:</p>
276
277<ul>
278<li>Proven and well tested: llvm-gcc and clang both use this technique for local
279mutable variables. As such, the most common clients of LLVM are using this to
280handle a bulk of their variables. You can be sure that bugs are found fast and
281fixed early.</li>
282
283<li>Extremely Fast: mem2reg has a number of special cases that make it fast in
284common cases as well as fully general. For example, it has fast-paths for
285variables that are only used in a single block, variables that only have one
286assignment point, good heuristics to avoid insertion of unneeded phi nodes, etc.
287</li>
288
289<li>Needed for debug info generation: <a href="../SourceLevelDebugging.html">
290Debug information in LLVM</a> relies on having the address of the variable
291exposed to attach debug info to it. This technique dovetails very naturally
292with this style of debug info.</li>
293</ul>
294
295<p>If nothing else, this makes it much easier to get your front-end up and
296running, and is very simple to implement. Lets extend Kaleidoscope with mutable
297variables now!
Chris Lattner00c992d2007-11-03 08:55:29 +0000[diff] [blame]298</p>
299</div>
300
301
302<!-- *********************************************************************** -->
303<div class="doc_section"><a name="code">Full Code Listing</a></div>
304<!-- *********************************************************************** -->
305
306<div class="doc_text">
307
308<p>
309Here is the complete code listing for our running example, enhanced with the
310if/then/else and for expressions.. To build this example, use:
311</p>
312
313<div class="doc_code">
314<pre>
315 # Compile
316 g++ -g toy.cpp `llvm-config --cppflags --ldflags --libs core jit native` -O3 -o toy
317 # Run
318 ./toy
319</pre>
320</div>
321
322<p>Here is the code:</p>
323
324<div class="doc_code">
325<pre>
326</pre>
327</div>
328
329</div>
330
331<!-- *********************************************************************** -->
332<hr>
333<address>
334 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
335 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
336 <a href="http://validator.w3.org/check/referer"><img
337 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
338
339 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
340 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
341 Last modified: $Date: 2007-10-17 11:05:13 -0700 (Wed, 17 Oct 2007) $
342</address>
343</body>
344</html>