1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
6 <title>The Revenge Of The Often Misunderstood GEP Instruction</title>
7 <link rel="stylesheet" href="llvm.css" type="text/css">
8 <style type="text/css">
9 TABLE { text-align: left; border: 1px solid black; border-collapse: collapse; margin: 0 0 0 0; }
14 <div class="doc_title">
15 The Revenge Of The Often Misunderstood GEP Instruction
18 <!-- *********************************************************************** -->
19 <div class="doc_section"><a name="intro"><b>Introduction</b></a></div>
20 <!-- *********************************************************************** -->
21 <div class="doc_text">
22 <p>GEP was mysterious and wily at first, but it turned out that the basic
23 workings were fairly comprehensible. However the dragon was merely subdued;
24 now it's back, and it has more fundamental complexity to confront. This
25 document seeks to uncover misunderstandings of the GEP operator that tend
26 to persist past initial confusion about the funky "extra 0" thing. Here we
27 show that the GEP instruction is really not quite as simple as it seems,
28 even after the initial confusion is overcome.</p>
31 <!-- *********************************************************************** -->
32 <div class="doc_subsection">
33 <a name="lead0"><b>How is GEP different from ptrtoint, arithmetic,
36 <div class="doc_text">
37 <p>It's very similar; there are only subtle differences.</p>
39 <p>With ptrtoint, you have to pick an integer type. One approach is to pick i64;
40 this is safe on everything LLVM supports (LLVM internally assumes pointers
41 are never wider than 64 bits in many places), and the optimizer will actually
42 narrow the i64 arithmetic down to the actual pointer size on targets which
43 don't support 64-bit arithmetic in most cases. However, there are some cases
44 where it doesn't do this. With GEP you can avoid this problem.
46 <p>Also, GEP carries additional pointer aliasing rules. It's invalid to take a
47 GEP from one object, address into a different separately allocated
48 object, and dereference it. IR producers (front-ends) must follow this rule,
49 and consumers (optimizers, specifically alias analysis) benefit from being
50 able to rely on it.</p>
52 <p>And, GEP is more concise in common cases.</p>
54 <p>However, for the underlying integer computation implied, there
59 <!-- *********************************************************************** -->
60 <div class="doc_subsection">
61 <a name="lead0"><b>I'm writing a backend for a target which needs custom
62 lowering for GEP. How do I do this?</b></a>
64 <div class="doc_text">
65 <p>You don't. The integer computation implied by a GEP is target-independent.
66 Typically what you'll need to do is make your backend pattern-match
67 expressions trees involving ADD, MUL, etc., which are what GEP is lowered
68 into. This has the advantage of letting your code work correctly in more
71 <p>GEP does use target-dependent parameters for the size and layout of data
72 types, which targets can customize.</p>
74 <p>If you require support for addressing units which are not 8 bits, you'll
75 need to fix a lot of code in the backend, with GEP lowering being only a
76 small piece of the overall picture.</p>
80 <!-- *********************************************************************** -->
81 <div class="doc_subsection">
82 <a name="lead0"><b>Why do struct member indices always use i32?</b></a>
84 <div class="doc_text">
85 <p>The specific type i32 is probably just a historical artifact, however it's
86 wide enough for all practical purposes, so there's been no need to change it.
87 It doesn't necessarily imply i32 address arithmetic; it's just an identifier
88 which identifies a field in a struct. Requiring that all struct indices be
89 the same reduces the range of possibilities for cases where two GEPs are
90 effectively the same but have distinct operand types.</p>
94 <!-- *********************************************************************** -->
95 <div class="doc_subsection">
96 <a name="lead0"><b>How does VLA addressing work with GEPs?</b></a>
98 <div class="doc_text">
99 <p>GEPs don't natively support VLAs. LLVM's type system is entirely static,
100 and GEP address computations are guided by an LLVM type.</p>
102 <p>VLA indices can be implemented as linearized indices. For example, an
103 expression like X[a][b][c], must be effectively lowered into a form
104 like X[a*m+b*n+c], so that it appears to the GEP as a single-dimensional
107 <p>This means if you want to write an analysis which understands array
108 indices and you want to support VLAs, your code will have to be
109 prepared to reverse-engineer the linearization. One way to solve this
110 problem is to use the ScalarEvolution library, which always presents
111 VLA and non-VLA indexing in the same manner.</p>
115 <!-- *********************************************************************** -->
116 <div class="doc_subsection">
117 <a name="lead0"><b>What happens if an array index is out of bounds?</b></a>
119 <div class="doc_text">
120 <p>There are two senses in which an array index can be out of bounds.</p>
122 <p>First, there's the array type which comes from the (static) type of
123 the first operand to the GEP. Indices greater than the number of elements
124 in the corresponding static array type are valid. There is no problem with
125 out of bounds indices in this sense. Indexing into an array only depends
126 on the size of the array element, not the number of elements.</p>
128 <p>A common example of how this is used is arrays where the size is not known.
129 It's common to use array types with zero length to represent these. The
130 fact that the static type says there are zero elements is irrelevant; it's
131 perfectly valid to compute arbitrary element indices, as the computation
132 only depends on the size of the array element, not the number of
133 elements. Note that zero-sized arrays are not a special case here.</p>
135 <p>This sense is unconnected with <tt>inbounds</tt> keyword. The
136 <tt>inbounds</tt> keyword is designed to describe low-level pointer
137 arithmetic overflow conditions, rather than high-level array
140 <p>Analysis passes which wish to understand array indexing should not
141 assume that the static array type bounds are respected.</p>
143 <p>The second sense of being out of bounds is computing an address that's
144 beyond the actual underlying allocated object.</p>
146 <p>With the <tt>inbounds</tt> keyword, the result value of the GEP is
147 undefined if the address is outside the actual underlying allocated
148 object and not the address one-past-the-end.</p>
150 <p>Without the <tt>inbounds</tt> keyword, there are no restrictions
151 on computing out-of-bounds addresses. Obviously, performing a load or
152 a store requires an address of allocated and sufficiently aligned
153 memory. But the GEP itself is only concerned with computing addresses.</p>
157 <!-- *********************************************************************** -->
158 <div class="doc_subsection">
159 <a name="lead0"><b>Can array indices be negative?</b></a>
161 <div class="doc_text">
162 <p>Yes. This is basically a special case of array indices being out
167 <!-- *********************************************************************** -->
168 <div class="doc_subsection">
169 <a name="lead0"><b>Can I compare two values computed with GEPs?</b></a>
171 <div class="doc_text">
172 <p>Yes. If both addresses are within the same allocated object, or
173 one-past-the-end, you'll get the comparison result you expect. If either
174 is outside of it, integer arithmetic wrapping may occur, so the
175 comparison may not be meaningful.</p>
179 <!-- *********************************************************************** -->
180 <div class="doc_subsection">
181 <a name="lead0"><b>Can I do GEP with a different pointer type than the type of
182 the underlying object?</b></a>
184 <div class="doc_text">
185 <p>Yes. There are no restrictions on bitcasting a pointer value to an arbitrary
186 pointer type. The types in a GEP serve only to define the parameters for the
187 underlying integer computation. They need not correspond with the actual
188 type of the underlying object.</p>
190 <p>Furthermore, loads and stores don't have to use the same types as the type
191 of the underlying object. Types in this context serve only to specify
192 memory size and alignment. Beyond that there are merely a hint to the
193 optimizer indicating how the value will likely be used.</p>
197 <!-- *********************************************************************** -->
198 <div class="doc_subsection">
199 <a name="lead0"><b>Can I cast an object's address to integer and add it
202 <div class="doc_text">
203 <p>You can compute an address that way, but if you use GEP to do the add,
204 you can't use that pointer to actually access the object, unless the
205 object is managed outside of LLVM.</p>
207 <p>The underlying integer computation is sufficiently defined; null has a
208 defined value -- zero -- and you can add whatever value you want to it.</p>
210 <p>However, it's invalid to access (load from or store to) an LLVM-aware
211 object with such a pointer. This includes GlobalVariables, Allocas, and
212 objects pointed to by noalias pointers.</p>
214 <p>If you really need this functionality, you can do the arithmetic with
215 explicit integer instructions, and use inttoptr to convert the result to
216 an address. Most of GEP's special aliasing rules do not apply to pointers
217 computed from ptrtoint, arithmetic, and inttoptr sequences.</p>
221 <!-- *********************************************************************** -->
222 <div class="doc_subsection">
223 <a name="lead0"><b>Can I compute the distance between two objects, and add
224 that value to one address to compute the other address?</b></a>
226 <div class="doc_text">
227 <p>As with arithmetic on null, You can use GEP to compute an address that
228 way, but you can't use that pointer to actually access the object if you
229 do, unless the object is managed outside of LLVM.</p>
231 <p>Also as above, ptrtoint and inttoptr provide an alternative way to do this
232 which do not have this restriction.</p>
236 <!-- *********************************************************************** -->
237 <div class="doc_subsection">
238 <a name="lead0"><b>Can I do type-based alias analysis on LLVM IR?</b></a>
240 <div class="doc_text">
241 <p>You can't do type-based alias analysis using LLVM's built-in type system,
242 because LLVM has no restrictions on mixing types in addressing, loads or
245 <p>It would be possible to add special annotations to the IR, probably using
246 metadata, to describe a different type system (such as the C type system),
247 and do type-based aliasing on top of that. This is a much bigger
248 undertaking though.</p>
252 <!-- *********************************************************************** -->
254 <div class="doc_subsection">
255 <a name="lead0"><b>What's an uglygep?</b></a>
257 <div class="doc_text">
258 <p>Some LLVM optimizers operate on GEPs by internally lowering them into
259 more primitive integer expressions, which allows them to be combined
260 with other integer expressions and/or split into multiple separate
261 integer expressions. If they've made non-trivial changes, translating
262 back into LLVM IR can involve reverse-engineering the structure of
263 the addressing in order to fit it into the static type of the original
264 first operand. It isn't always possibly to fully reconstruct this
265 structure; sometimes the underlying addressing doesn't correspond with
266 the static type at all. In such cases the optimizer instead will emit
267 a GEP with the base pointer casted to a simple address-unit pointer,
268 using the name "uglygep". This isn't pretty, but it's just as
269 valid, and it's sufficient to preserve the pointer aliasing guarantees
270 that GEP provides.</p>
274 <!-- *********************************************************************** -->
276 <div class="doc_subsection">
277 <a name="lead0"><b>Can GEP index into vector elements?</b></a>
279 <div class="doc_text">
280 <p>Sort of. This hasn't always been forcefully disallowed, though it's
281 not recommended. It leads to awkward special cases in the optimizers.
282 In the future, it may be outright disallowed.</p>
284 <p>Instead, you should cast your pointer types and use arrays instead of
285 vectors for addressing. Arrays have the same in-memory representation
286 as vectors, so the addressing is interchangeable.</p>
290 <!-- *********************************************************************** -->
292 <div class="doc_subsection">
293 <a name="lead0"><b>Can GEP index into unions?</b></a>
295 <div class="doc_text">
300 <!-- *********************************************************************** -->
302 <div class="doc_subsection">
303 <a name="lead0"><b>What happens if a GEP computation overflows?</b></a>
305 <div class="doc_text">
306 <p>If the GEP has the <tt>inbounds</tt> keyword, the result value is
309 <p>Otherwise, the result value is the result from evaluating the implied
310 two's complement integer computation. However, since there's no
311 guarantee of where an object will be allocated in the address space,
312 such values have limited meaning.</p>
316 <!-- *********************************************************************** -->
318 <div class="doc_subsection">
319 <a name="lead0"><b>What effect do address spaces have on GEPs?</b></a>
321 <div class="doc_text">
322 <p>None, except that the address space qualifier on the first operand pointer
323 type always matches the address space qualifier on the result type.</p>
327 <!-- *********************************************************************** -->
329 <div class="doc_subsection">
330 <a name="lead0"><b>Why is GEP designed this way?</b></a>
332 <div class="doc_text">
333 <p>The design of GEP has the following goals, in rough unofficial
334 order of priority:</p>
336 <li>Support C, C-like languages, and languages which can be
337 conceptually lowered into C (this covers a lot).</li>
338 <li>Support optimizations such as those that are common in
340 <li>Provide a consistent method for computing addresses so that
341 address computations don't need to be a part of load and
342 store instructions in the IR.</li>
343 <li>Support non-C-like languages, to the extent that it doesn't
344 interfere with other goals.</li>
345 <li>Minimize target-specific information in the IR.</li>
349 <!-- *********************************************************************** -->
353 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
354 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
355 <a href="http://validator.w3.org/check/referer"><img
356 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
357 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br/>
358 Last modified: $Date$